November 17, 2025 • 59 mins
Non-technical execs are being told “AI agents will change everything” but no one tells them how to keep those agents from breaking things.
In this episode of AI for the C-Suite, Chad Harvey talks with Oren Michels, co-founder & CEO of Barn Door AI (and founder of Mashery), about:
-
Why agents are like enthusiastic interns with no fear of consequences
-
What an AI control plane actually does
-
How MCP is becoming essential infrastructure
-
Real, early ROI use cases and where most companies are still stuck in “science experiment” mode
If you’re a middle-market leader trying to move beyond chatbots and copilots into real automation with guardrails then this conversation is a must listen.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:01):
Greetings, innovators and leaders.
This is AI for the C-suite, your compass for navigating the exponential age.
I'm your host, Chad Harvey, and we're here to bring you cutting edge insights on AItailored specifically for middle market organizations.
Buckle up as we embark on a journey through the transformative world of artificialintelligence.
(00:22):
Today I am pleased to welcome Orin Michaels, co-founder and CEO of barndoor.ai to AI forthe C-suite.
Orin brings a unique perspective to enterprise AI, having successfully navigated the earlydays of another transformative technology that changed how businesses operate.
Back in 2006, when most people had never heard of APIs, founded Mashery to help companiesmanage these critical business connections.
(00:50):
He built that company over seven years and sold it to Intel in 2013 just as mobile appswere exploding and every business needed an API strategy.
Today, Orin sees a similar pattern emerging with AI agents in the enterprise.
Last November, he co-founded Barn Door AI after recognizing that companies face the same
(01:11):
fundamental challenge they did with APIs.
do you safely connect powerful new tech to your critical business systems?
Barn door raised a 13.5 million seed round and has grown to a team of 22 building whatOrin calls the control plane for agentic AI in the enterprise.
What makes Orin's perspective particularly valuable is his focus on a problem many execsare just beginning to grapple with.
(01:35):
AI agents aren't just tools that provide information, they take actions.
And unlike humans,
employees, agents, have no conscience, no fear of consequences, and no inherentunderstanding of business context.
Today, we'll explore how middle market companies can harness the productivity benefits ofAI agents while maintaining the security and governance necessary to protect their
(01:58):
businesses.
Oren, welcome to AI for the C-suite.
Thanks Chad, that was lovely intro.
Thank you!
We put a lot of time crafting intros so I appreciate the props.
to help you craft that intro?
just a little bit yes just a little bit I've got a secret sauce here that's a sprinkle ofme a sprinkle you and a sprinkle something else so hopefully it comes off alright so let's
(02:23):
dive right in uh I mentioned in the intro that you are describing barn door as the controlplane for agentic AI and I would like to have you explain what a control plane is for
anybody that's not familiar with that terminology and why is it that enterprises uh
need one specifically for AI agents.
(02:45):
Well, when we hire humans, we give them permissions of what they're allowed to do and whatthey're not allowed to do.
We say, can use Salesforce and you can do these things in it.
You can use Google Docs and you can do those things in it and you can see these ones, butyou can't see those ones.
um agents and any AI for that matter that's going to access the same systems that we ashumans use need to have that level of governance as well.
(03:13):
And it already exists for the humans, but that's not quite good enough.
As you mentioned, the agents have a lot more of a blast radius.
They can do a lot more, a lot faster than I can.
They're not afraid of being fired.
And they're like enthusiastic interns.
They really want to just take their instructions and go do it, no matter what happens.
And so just saying that you have an identity system that says that Orin's allowed to haveaccess to these things is not enough if I start to use things that act.
(03:43):
on my behalf but independently without necessarily my governance and my conscience.
So when you think about each of those things needs to have that governance.
And there's going to be a lot of those things, right?
Right now we've got more AIs than you and I can name in the time we have here.
Yes.
And I just got it off a call with a very large financial institution where the CIO theresays he fully expects each of his employees to have hundreds or perhaps even thousands of
(04:12):
agents that they use in their job and he has 20,000 employees.
And so if you want to be able to start managing that, you can either try to go agent byagent, platform by platform, each of which presents its own interface or.
admin tools to try to make that happen, or you can say no, we're just going to putsomething in between that is a gateway, it's a controlling device, and it says here are
(04:40):
all the different things, and we understand what they're, what they're.
capabilities are and what their risks are and that sort of thing.
Here are all the humans and their roles and we understand what they do.
On the other side, here are all the tools people use.
Here's Salesforce and Snowflake and Notion and Google Docs and all those things.
And something called a control plane will manage the traffic that goes between them.
(05:02):
It will give the administrators the ability to see what these agents are trying to do.
So if you say to the Salesforce agent or an agent using Salesforce, know, go update myopportunities and it...
goes to try to delete them instead, well, you'll see that, right?
And you'll have rules in place to keep that from happening.
And so it manages this across many, many, many agents on one side and many, many, tools onthe other side with the humans sort of on the third side.
(05:30):
So how is this different from some of the security and access control systems thatcompanies already have in place that give them visibility into network traffic and uh some
level of control?
Yeah, well network traffic is important.
uh That is one layer and you think of the network layer models, right?
And so network traffic says, and it's important, for instance, suggest that you tell theMCP servers you're running or ours if you're using ours, that they can only accept network
(06:00):
traffic that comes through us.
So that's a particular kind of rule.
Right, a network traffic rule might say even, and there's authentication things andidentity that say, Oren is allowed to use this set of records in Salesforce and he's
allowed to update them.
But none at the packet level, you're not going to be able to say, Oren is a low levelsales employee, he's allowed to update opportunities with his agent, but only those that
(06:28):
are worth less than $2,000.
you
He's allowed to, as I do actually, I'm CEO with my agents.
When I finish a call, uh use Cloud to update Salesforce about it.
And I make it very clear that Cloud is allowed to create a new call record.
But it's not allowed to modify the ones that are there.
And it's not allowed to modify the underlying record or the account because that's notwhat I want it to be doing.
(06:53):
And although I'm allowed to do those things, I don't want to have to worry about it.
Mm-hmm.
I make sure that when I'm using Cloud for this task, it's only allowed to do the thingsthat it should be allowed to do given the context of what it's doing.
And that is not, you you talk about security systems, security systems are all designedaround, there's these nefarious people outside who are trying to penetrate and get in and
(07:16):
steal your stuff.
And we have to deal with that.
That's an important part of what we do.
It's what I sort of call security with a capital S.
That is, know, have a CISO and they do that.
But those folks are not used to saying here's a perfectly authorized human using a toolthey're allowed to use to access a system they're allowed to access, but.
That tool is doing stupid stuff because it wasn't necessarily, it didn't have the contextor it got a bad prompt or it just interpreted it wrong or it had a bad day or whatever
(07:47):
reason.
And so you want to be able to be sure that this kind of management is also in place andthat's not the role of traditional security.
Okay.
So help me understand, uh because a lot of our listeners are non-technical execs, as manyare, right?
So I'm a CEO.
(08:09):
Well, and I even just heard you in your prior explanation describe yourself as a low-levelsales employee.
you know, right.
So let's say that I'm leading an organization.
um What
is it that your technology is going to do to flag these bad use cases because again thecall is coming from inside the house here essentially so how are they going to determine
(08:35):
what is appropriate and what is not appropriate and how are they going to flag that
So that is set by a combination of people, right?
So again, going back to your security person, your security person, your CISO, or theirtitle is, they understand these outside actors and there's going to be a series of
policies they dictate.
But they don't know the ins and outs of Salesforce.
They don't know the ins and outs of Google Docs because that's not their world.
(08:56):
And so those policies will be set perhaps by a Salesforce admin or an HR team or whateverand have to be able to manage by them.
So when you say how does the system know it,
Well, the system knows it because we tell it.
mean, we or our customers tell it.
But what's more important is to say, how do you go about doing that?
And the answer is you do it the same way you would when you have that new intern show up.
(09:19):
And you say, OK, I'm going to give you a task, and I'm going to see how you attempt to dothat task.
And it's going to be a task that you may or may not have all the ability to go do rightnow.
But I want to see how you do that task and what your
Process is to do it so I can decide if I can trust you to really do it for real right andthat's sort of how we give people more and more authority and with these it's the same
(09:41):
sort of thing so you have you know you have Claude and you say Okay, got Claude, you knowgo clean up my opportunities And then you can look and see what how it interprets that and
what the specific calls are it is to salesforce
At that point in time, you may have shut all that stuff out.
You have given it only read access and it can't actually go do it.
It's gonna error out, which is fine because then you can say, okay, it's error out, buthow did it try?
(10:06):
When I gave it that instruction, did it try to do the things that were consistent withthat task?
And if the answer is yes, you start turning those on and you say, okay, now try it againand it succeeds and you begin to get to a place where this set of instructions going to
that.
AI tool using these set of enterprise tools are likely to not cause trouble and you'rewilling to bring that on.
(10:31):
You also will have, as we talked about, hundreds perhaps of these different agentsworking, depending on the context and what that agent's doing, that exact same agent may
have different permissions assigned to it.
So you may say, on the occasion I'm doing this particular task, I've noticed it's capableof...
interpreting it correctly that I'm allowed to have it right into opportunities.
(10:52):
But when it does that thing, sometimes it tries to do things it shouldn't, so I'm cuttingit off at that.
I find the intersection and the overlap between permissions and process for agentsabsolutely fascinating because as we are promised the era of true agentic AI, we're
promised this idea that agents are going to learn and they're going to be able to takeadditionally increased actions.
(11:17):
I just wonder how do we continue to evolve and adapt both the permissions and the processthrough the type of infrastructure that you're talking about to ensure that we are getting
the results that we desire.
Does that make sense?
If you could restate it, I'm not sure exactly what you're.
(11:38):
Sure.
So we talked to you, you give a very good uh answer about how this works, right, from anon-technical perspective.
We talked about policy, but we also talked about permissions and what is allowed andwhat's not allowed.
And it sounds as if there's also an overlap there in terms of the documented process thatcertain agents are going to follow.
(11:58):
And I wonder, as we unleash more uh self-evolving agents, how do we continue to ensurethat we understand the process
that they're following as well as adapt and update the permissions based on that changingenvironment.
Does that make sense?
Yeah.
I think that certainly in the current time horizon, we don't understand it.
(12:20):
And that's part of the problem, right?
We also don't understand humans, by the way.
And so I know, it's this thing, right?
And so when I have an employee, as much as I would love to drill into their brain andunderstand why they think that way, I can't go do that.
And I can't really do that with
AI agents either because we really don't understand how the technology works.
(12:40):
so instead, what you do with humans and what you do with agents is you observe how theyinteract with the world around them.
And is that consistent with the task at hand and are they doing things or attempting to dothings that don't make sense?
If the low level sales employee is trying to access records in the payroll team, that'sprobably an indication that person is doing stuff they shouldn't be doing.
(13:05):
And you'll want to shut that down and be aware of it.
I think that as the agents get more complex and as they interact with each other, havemulti-step agents and agent to agent and all these sorts of things.
you will still be in a situation where ultimately the best way to ensure that the rightthing is happening is to know what's going in, what's coming out, what's going between the
(13:31):
different agents, be aware of those lines of communication and set your rules there.
Because trying to set them within the AIs themselves, at least at this point, is notsomething that I would necessarily rely on being able to do.
Fair enough.
When do you typically find that you're getting the calls to get involved withorganizations that are either deploying, experimenting, or thinking about agents?
(13:58):
And I know the ideal solution here is, we want to be in upfront before you've even startedthinking about this.
But when do you typically start to get these calls?
Yeah, so and you know, we're talking about agents, but it's any AI that interacts with it,right?
So, you know, is Claude agentic?
I don't know, I mean, I type, I ask to do something, sometimes it acts on my behalf,sometimes it just gets me information.
(14:22):
you know, it's any AI that needs to talk to underlying systems, generally today using thisprotocol called MCP, which we can get into if you want to.
But.
Generally speaking, people contact us when the science experiment didn't really yield muchin the way of ROI because everybody was kind of afraid to actually let it do something.
(14:48):
Mm-hmm.
you had the side, you hooked it up and you said, I can list all my opportunities.
Well, I can do that in Salesforce too, that's not helpful, right?
What's helpful is for me to be able to bring in information from multiple different placesand I can, at the beginning of the week, can pull the schedules of all my salespeople,
can...
(15:10):
bump those up against the opportunities we have open, I can pull in all the Slack chatsthat we're talking about them and I can get an overview of what's likely to happen that
week and how I can be of assistance and when I should pop in and say hi to these people.
And so, you know, it's when they come to us when they generally have some known use case,there's something they wanna do that they think is gonna make them money or make them more
(15:35):
productive or whatever it is businesses wanna do, whatever their goals are.
and give them more time at home, whatever it could be.
And they're ready to really make this work for reals.
But it's scary and complicated and they see an array of all these different MCP serversout there, they can't really trust them, they don't know what they are, they're not.
(16:02):
They're sort of worried about what will happen and um and they want to be aware and becontrolling.
That's one place.
The other place we get called in is that we're able to deploy not we our standarddeployment is in the cloud or really loves the cloud but we can also deploy in private
clouds and on-prem.
So like we have major financial institutions that want to implement these technologies MCPand other technologies inside in their fire inside their firewall and
(16:31):
you know, the do-it-yourself versions out there are just not enterprise-grade for that,right?
So we're able to do that, and that allows us to serve a particularly security-consciousclass of customer.
you mentioned mcp and i do want to dive into that before because that is uh...
(16:54):
absolutely a very hot topic right now before we do that though i want to jump back in thethe way back machine here a little bit and uh...
i want to talk about your build of mashery and your observation about APIs right i knowthat you have said that the AI agentic challenge looks like the same movie that you saw
back in 06 and you've got wonderful movie posters behind you so
(17:16):
are Broadway posters.
These are shows.
These are Broadway.
okay.
I was thinking about the good night, good luck there that I saw with Clooney.
So, where are you now?
boy.
All right.
We need to explore that in a separate episode, I think.
uh So, uh what are the specific parallels that you're seeing right now?
And what were some of the lessons that you learned from the API era that might be relevanttoday?
(17:39):
Well, know, so in the API era and I started Mashery in 2006, okay, so that was beforethere was an iPhone and before there was really AWS to any real extent, we were one of
their first customers.
I had been running business development for the world's second best blog search engine.
The world didn't really need the first best one either, but there was some reallyinteresting technology we had that a large company wanted to integrate in their
(18:06):
capabilities.
And so they came to me and I was a biz dev guy.
and they said, we would like to access your API.
And I said, great.
So I went to our engineers, I said, we got this big deal, they'll pay us a bunch of money,all we have to do is give them our API.
And they're like, well, we don't have one and we can't.
And I'm like, well, why not?
Well, they gave me all these reasons that boiled down to they didn't want to build theinfrastructure around doing it.
(18:26):
And I thought, well, this is silly because I can't dev any biz if I can't make this thinghappen.
And I started realizing that.
as the world was sort of moving to these web services concept, this was gonna becomereally, really important.
And so I left my job there and I went to one of the investors who had been an investor inthis company, we having dinner and I was explaining this problem and...
(18:50):
He said, well, you should start a company to do that.
I said, great, I will.
So I did, and that became Mashery.
we began building, and what it was that was, back then, the engineers had the keys to thekingdom, and if you wanted someone to have access to your API, you found the dude who knew
where the appliance was and had the password to it and knew how to use Terminal and couldmake something, create something that maybe wouldn't get disrupted, but if someone tried
(19:16):
to use it too much, then you were all screwed.
And that just didn't seem
seemed like a way to run a business.
so now, what we did at that point was we turned it into a business tool that companiescould use.
Conveniently, Apple brought the iPhone out in the app store.
Everybody needed a mobile strategy.
Mobile apps, talked to APIs, and here we were.
(19:38):
And so now, what we're seeing is we have this AI, and the initial version of what we allthink of AI and businesses at least is, you know,
these chat interfaces where you, the very, very best and brightest uses and mosteconomically viable uses we've come up with for these chat interfaces are coding copilots.
(20:02):
Coding copilots are saving people a lot of money and a lot of time.
Why?
They mimic how coders work.
If a coder has code that's not working, they go to a super smart PhD, their CTO, and say,here's all my code, fix it for me.
And this,
They do, right?
that is something chat is incredibly good at doing.
So once again, the engineers built the tool that they can love.
(20:24):
And the rest of us are like, wait, I could use something like that if only.
And the if only is, if only it could talk to the stuff I use.
If only it could access the systems I can access.
And then we have this agentic thing coming along that wants to write to things, which isnot really what the role of your typical chat interface was, right?
(20:45):
And so now you're saying, okay, there's this thing and it needs access to the realbusiness stuff to go do things and be productive.
And just as that large company needed access to my stuff to make me money.
And so in that sense, it seems to be, there's a lot of parallels and in fact, you know,there's the technological challenges.
(21:08):
We are a proxy, we are a gateway between this and that.
That's essentially what API management is, right?
And so the way I viewed it um in terms of the spec of what we're building has an awful lotto do with it.
It's extended because it's not merely, there are API management companies and they have AIproducts, but again, they're all, they're based specifically on a key and an access
(21:32):
control and don't have the same level of visibility and management because that's not.
part of API management, so at the level beyond, the fact that you have access to this callis not sufficient to say that you should be allowed to do it.
I appreciate you walking me through that because I think the way that you answered thatactually does lend itself to us going in a little bit deeper on MCPs now since you talked
(21:57):
about all that connective tissue.
So for the folks out there that MCP uh is just uh they've never heard of it before.
Let's talk about model context protocol, MCP.
uh What is MCP?
Why does it matter?
why does it matter?
So what I mentioned, I started this company to help people connect to, connect their AIsto the systems they use.
(22:22):
And conveniently and nicely, the folks at Anthropic last November put out this protocolthat purports to make that easy and rational.
And what it essentially does, you have,
We'll spend a sec on APIs here.
APIs are essentially how computers talk to computers.
(22:43):
So if I want to interact with a particular application, I can do it through the userinterface or a computer over here can sort of bypass that and go straight to it with a
conduit that they each understand called an API.
And it has some security on it.
But an AI isn't a normal computer.
(23:03):
It's not just a here is my query.
to a, the AI act, or the computer accessing an API has been programmed by programmers whoknow the rules of that API and they say, well, if we want it to happen, you send a query
in this form and you'll get an answer in that form.
And that, and it's a known thing in programmers program that.
(23:24):
AIs don't work like that.
AIs like us, they're sort of ambiguous.
You're given something like, how am gonna go about doing this?
What tools do I have at my disposal?
So MCP is essentially, it is a menu of the various options and things and tools andcapabilities that some system somewhere is giving this AI as sort of the universe of toys
(23:52):
they can play with to answer that thing and then also provides it sort of the form that itwould need to play with that in terms of how it should make the request and what.
information it's going to get back.
And you might have multiples of these.
I could have my AI talking to Salesforce and then interacting with Notion and then sendingan email on my behalf and putting it into a PDF and dropping it at Google Drive.
(24:19):
So each of those things can be, has a select, a group of menu items that you're able toplay with.
And.
On a very rudimentary level, often, there's an ability to turn specific ones of these onand off in a fairly uh broad way.
(24:41):
And then, of course, we provide a much more sophisticated and nuanced way of managingthat.
But.
But fundamentally, it's saying to the AI, in a form the AI can understand, here are theexternal resources, tools, abilities, data at your disposal that you can use to either
answer questions or take the actions that your human is telling you you're supposed to do.
(25:08):
you
So I think that begs the question then because we know that the AI tools that are outthere, the large language models, the agents upon which they're built, they're statistical
and they're probabilistic in nature, right?
They're not true thinking machines.
So, right, okay.
So I'm an AI agent.
Well, I'm not an AI agent.
Let's say I'm an AI agent.
(25:29):
And uh I'm calling up the MCP server here and I'm looking at all the different thingsthere.
How am I going to make the determination probabilistically, statistically,
which one of those menu options I need to select to do my task.
How does that look?
Yeah, it's an interesting question and it also, know, so I would say that the depths ofhow AI models function and think is probably beyond the scope of today's podcast.
(26:01):
But ultimately, there are things you're going to want to do to influence the AI to choosethe right tools, the tools you'd like it to use.
So for instance,
or to use them in a certain way.
So for instance, uh I can interact with Cloud and Cloud like most AIs count these thingscalled tokens which are essentially a unit of measure that shows how much data is going in
(26:27):
and out of them that they have to deal with.
It's a bad definition but we'll go with it.
So if the AI says I want to pull up the
my contact, Chad Harvey in Salesforce, one way it could do it is it could assume thatthere's a Chad Harvey contacting and say I want the Chad Harvey contact.
(26:50):
Or the easier for the AI way is.
give me all the contacts and I'm look for contacts that sound like something like ChadHarvey, I'm gonna get the contact ID for it and I'm gonna go back and request it
specifically by the ID.
Well, that second way just burned a bunch of tokens because it looked at the entire thing,right?
This is how it usually runs if it's allowed to.
(27:11):
And so you now have other, there's a layer of sophistication that's being built into this.
It's not as important an issue today because
many of these AIs aren't charging you for tokens yet, but they will be, where our moresophisticated customers are like, we need your help managing that as well.
(27:32):
And so you're not just using the MCP to give it the full menu.
You're managing the waiter in the restaurant.
They're giving you the full menu, but they're also saying, today's special is this one.
And if you can use that one, you probably want to, and it's a better deal kind of thing.
Yeah.
Okay.
(27:53):
I think that that is a perfect focal point to go back to something you addressed earlierwhen you categorized agents as enthusiastic interns, right?
um And I think if I'm an executive and I'm thinking about deploying agents, that helps meunderstand uh the fact that the agents don't necessarily understand the consequences of
(28:14):
their actions.
um So when you're working with organizations and uh you come across this type of subjectmatter, how,
How else do you help execs think about deploying agents and what have you learned that isgood counsel for leaders when this topic comes up?
I think it starts with understand which agents are good at what.
(28:40):
Using the right tool for the job certainly helps, but it's also to understand what you'reexpecting them to do and how you're expecting them to do it so that you can instruct them
in a rational way and experiment with how you instruct them ah in order to
(29:01):
try to get to the most efficient and effective way to get the job done.
But it's, you you're right.
They're not, it's not linear.
It's not an easy thing to do as managing humans is an easy thing to do.
And so, you know, I think it's also that, other thing I advise executives to do is don'tput all your eggs in one basket.
(29:24):
Don't say, you know, I just saw, you know,
and that's gonna be the cool AI, I'm gonna use that one, right?
It is a cool AI, it's not the only cool AI, right?
And so, I was with a group of insurance executives yesterday, and one of those like, we'reimplementing Rider, and it's great, except when it's not, right?
(29:44):
And so, we encourage people to have a low bar for trying things.
And also to realize it's all changing so fast.
So we see this with the different models.
What ChatGBT is really, really good at today, they bring out a new model and suddenly,Gemini's better at that.
(30:04):
And so understanding when to use which of the arrows in your quiver uh is a big part ofthis and not insisting on sticking with ones just because it's the one you happen to be
used to use.
Or in one of the things that's coming across very clearly in our conversation today for meat least is that you are very comfortable speaking truth to power and for our audio only
(30:31):
listeners there's a poster we referenced earlier on the wall behind you good night goodluck which is all about speaking truth to power uh
yes it is uh...
and so i'm i'm curious how well does uh...
your message land with c e o's that don't want to hear some of what you're saying rightnow because you've got a what i would call a a spiky point of view uh...
(30:54):
occasionally about certain things which is probably contrary to a vision that a lot ofleaders have in their mind about the way this stuff is going to work
Well, I think it's the vision, I think the folks that would be more upset with the visionthat we have are the ones who want to be the only AI anyone needs, right?
(31:16):
And so if your software company acts, know, pick a SaaS company, they all have AI agents,and they're like, use my AI agent, it'll do everything you need.
Because that's what they want, but they won't, right?
And so the folks who I think are mostly pushed aside by this are the ones who have tobecome convinced that their AI has to play well with other people's AIs and that they're
(31:40):
not building the one and only, and no one is at this point, right?
Certainly not.
know, an agent built into an existing piece of software.
That, that, you know, no human only uses one piece of software to do their job.
So I don't know how you could expect any agent to only use one piece of software to dotheir job.
And so, you know, that, that little conversation is, is a little bit.
(32:04):
Dice here and you know the platform folks, you we run our our Company on platform X andtherefore will use their AI management.
Well, well, okay except when you want you know, you when your competitor is using thisother AI over here that isn't compatible with that and you know, that's problem, right so
You know, think a lot of companies are trying to to portray themselves as the only AI AIyou'll ever need
(32:31):
And I don't think that any company will ever be that, but certainly no company is thattoday.
When I go to our customers and talk to them, I'm not really speaking truth to power.
I'm basically saying, look, you have a desire to make this happen.
You recognize this is super important for your company, for your job, for your competitivepositioning, and you are unhappy with the productivity you're getting from AI so far.
(32:54):
And I know very few CEOs.
who will say, am really happy with our AI productivity so far.
Actually, I know one, but he's an unusual person running an unusual company, and he'sthrilled, and we should all be like him.
But the rest of them um are not very happy at all.
And so,
You know the other folks who we probably irritate a bit are the tinkerers who want to, I'mgonna download this thing and do this thing and I'm gonna sort of run everything through
(33:23):
me so I'm important.
And those folks don't necessarily like to relinquish power but we lived through thatbefore.
The other thing we lived through during the API time was of course the cloud revolution.
And you'll recall that when Mashery started in 2006,
My previous company, I had to raise $4 million before I wrote a line of code.
(33:47):
A million went to Sun for two servers.
A million went to Cisco for two routers.
I had to get a cage.
I had to buy a bunch of Oracle software.
All of that stuff's free now, right?
And so during that transition, I remember speaking at a conference at Interop, theconference for the big data center buyers, right?
And I was on the cloud panel because we were one of the first companies that deployed inthe cloud.
(34:10):
And the basic point of the cloud panel, so nostalgic there was such a thing, was we got upthere on all these
dinosaurs in the room were saying all the reasons why are we were wrong and the cloudwould never survive and I'm like luck we're in Vegas and interrupt and You're talking to
me your CEO is at a much better conference in Hawaii and she's learning that you can havedata centers with no capex who's gonna win that argument like just you guys are out of
(34:38):
your minds if you think this isn't gonna work and And further I'm sure you guys are allsuper smart, but I know Amazon is better at running data centers than you are
And so, this evolution is going to happen.
The smart CEOs get it.
And they just want to see it happen.
They want to see it happen now.
(34:58):
I want to see it happen now.
I'm not happy with the level of AI in my company.
I've got 22 people, all of whom know about using AI.
So I want more.
you
I want more.
That could be, I think, the theme of the era right now.
uh I want to go back and tug a little bit more on a topic that we started exploringearlier in terms of the visibility of what your platform enables.
(35:23):
And uh I'm interested in what patterns have you seen in the data or what patterns maybehave your customers seen that surprised them?
uh Things happening with their agents that they didn't expect or something else?
I don't know what.
Has there been anything surprising?
Well, I think the big surprise that everybody has to have with AI in general, and it'sjust amplified when it's agenting and it's taking actions, is when it does things that are
(35:53):
completely counter to what you tell it to do.
And you say to it, why did you do this?
This is not what I said.
And of course, it's always very apologetic.
Oh, I'm so sorry.
I didn't mean to do that.
And that
You know, that the concept of how you ask for what you want really, really matters.
(36:15):
And that I would say that the, and by the way, it's also not consistent.
You could have the same.
you know, I can say, you know, go do this thing in Salesforce that you did, that it didyesterday and today it says, I don't have access to Salesforce.
I'm like, well, you had it 10 minutes ago.
my goodness.
Yes, I do.
You know, and, and, uh, you know, so I think a lot of it is still, you know, it is thisincredibly powerful technology that often just seems kind of stupid and it, and, that you
(36:46):
have to understand that, you know, I'll go back to the cloud thing when we were, we wereon the first.
companies using EC2 that we bet the farm on it.
We were like, we're all in.
People are like, know, how can you manage something where you're not the ones controllingthe actual computers?
And I said, you know, the traditional data center management is you have these computersrunning and then you have things you do when they go down and you fix them.
(37:15):
And we decided to turn that on its head and say, we need to manage in such a way as weassume they'll often be down.
and understand how to always be up, even if they weren't, and figure out what that lookedlike.
And we built a bunch of tools around that that are now standard stuff.
But the mentality is, how do you leverage that uncertainty?
(37:37):
And how do you exist in it and still get the kind of return that you want?
And it's a lot of experimentation and a lot of accepting that there will be the occasionalbad outcome.
Mm-hmm.
there are with humans, humans do stupid stuff too, and that if you can mitigate the risksand manage the downside, if you can put guardrails up that make it so that when it doesn't
(38:05):
do the thing you thought it was gonna do, that's not the end of the world.
And then you figure out how to make it do what you want.
You raised a number of important points there, one of which I think is risk appetite andthe larger the organization, the less willing they are to take risks in general.
you seeing that mentality or that mindset shift at all given the rapid rise of AI and theneed to start to take risks and take more of a bottom-up approach in many instances?
(38:35):
Yeah, I did it, I was at a conference, a big CIO, major company CIO conference recently,and they had one of these things on stage where they asked this question, it responded in
an app in real time, and it was, with AI, do you want to be fast and risky or slow andcareful?
And these CIOs in the room were like 60, 40 fast and take risks.
(39:00):
Okay, which, know, okay, great, good.
um But it also depends on the kind of company, right?
So these financial service companies we're dealing with where they're insisting on usdeploying in private cloud and they're very, very, very, very careful, but also recognize
that they are competing with people who aren't.
(39:23):
And ah they have to manage that risk.
um And then there are companies that are really trying to say,
How do we mitigate this?
One company I talked to, a big e-commerce, big retailer uh in person, oh a bricks andmortar and e-commerce retailer, and they have a particular category of product that has a
lot of defects.
(39:43):
uh The defects arise when people bring the product home and install it.
And so they created an agent that has the ability when the customer service folks reach aparticular thing with someone who calls them or who chats with them or who comes in
through the agent, that if it's this category of product, it's under $200 and they'resaying certain things that make it likely that the product has failed, that this agent
(40:09):
will automatically order and drop ship a replacement.
But they've put guardrails around it where they say, I know
I'm taking some risks.
I'm gonna give away some widgets I should know.
Someone's gonna figure out how to take advantage of this.
But on the other hand, I'm saving a bunch of money from people going through a bunch ofmachinations before they do the inevitable and I just ordered the thing to replace this
(40:33):
thing, right?
And if the risk is greater than...
If the dollar cost of the bad outcomes is lower than the money saved by the good outcomesand you get customers who are happy because they get instant service, um it's a win.
(40:53):
And I think a lot of technology and automation uh falls into that.
You're never going to have no bad outcomes, but you have to manage the risk and manage thedownside.
And if you can, you're going to succeed and take advantage of the folks who are too riskaverse.
I think that's a very fair assessment.
(41:14):
And talking about money and risk.
our audience is predominantly middle market companies that don't have the resources of aFortune 500, but they've got the same challenges, quite frankly, especially.
Yeah.
What do you suggest as an approach for these organizations in terms of implementing agentgovernance?
(41:35):
Some of the things that we've been talking about today, what what are exploratory stepsthat they should be taking right now?
ah If say I'm not, you know, ah I
resource like a fortune 500 company.
and we have plans for teams that start at 500 bucks a month and are designed for SMBs,right?
So we're there, right there with you.
(41:58):
uh And the good thing about programs like that, whether it's ours or software in general,is you want to have the level of security and management afforded
to the enterprises, you just don't need bunch of the bells and whistles they do.
And so oh you look for that.
(42:21):
But I think the other thing is that the best thing you can do is talk to folks who aredoing this, that are your peers.
And we go, the thing that all of my team is laser focused on, I remind them about this allthe time, is this is so early, agentic ROI is like not much of a thing right now.
(42:42):
No.
And so we as technologists, we as humans are pattern matchers.
And so the more things we can go out and say, here is this little use case that someone'susing to make, to actually have positive ROI on an.
really getting something out of this agent, the more you can surface those things and hearabout how they were done, you then can pattern-match them.
(43:04):
Well, I don't have that problem, but that sounds a lot like this thing I do have, and ifthey could solve it with that way, this might give me some insight as to how to go do it.
And so I would encourage people to talk to folks even outside their businesses and otherkinds of companies, know, whenever you're at some event where you're talking to your
peers, just bring this up and say, you know, how are you?
(43:28):
making AI happen.
This is actually one of the great icebreakers that um some of the dinners I attend thatthis one organization uses.
And they go around the room and they say, other than coding copilots and customer servicechat bots, what is the most productive way you're using AI in your company?
And I've heard some really interesting things there.
And it's through those that the rest of the people in room go, huh, I hadn't thought oftrying to use it for that.
(43:52):
That's a really good, because at this point it's all new.
Right.
It is and I think you are spot on in your analysis and your remarks about this being earlydays and we really don't understand uh yet where this is all going yet.
You clearly have had the insight that we're going to need some different types ofinfrastructure to make all this work.
(44:15):
And I'm thinking back to a conversation I had uh much, much earlier in 2025 with PapiManon from outshift by Cisco and they are thinking very much about
what they're calling the Internet of Agents and what is going to be required, right, interms of infrastructure and whether our current Internet infrastructure can even, and
networking infrastructure is capable of withstanding that.
(44:36):
So, all that word salad out there.
What else are you seeing in the landscape that, uh much like you're reevaluating thisparticular component uh for AI, what else are you seeing out there that you think we need
to really be taking a hard look at and saying, hmm, I'm not sure this is going to serve uswell going forward?
know, AI specific, I would say, I actually don't believe that the chat interface survivesas the dominant way for enterprise, least, to deal with AI.
(45:09):
Because again, it's not how we ordinarily do our jobs.
We don't.
As you tend to chat, there's a whole reason why, you know, could say user interface folksexist because computers are hard to deal with and so it simplifies that so we humans can
do it.
Okay, fine.
But over time, we've found more efficient ways to give instructions and to get informationback.
(45:35):
And I think the chat interface, again, it's a combination of it's a great way to get yourcode fixed, which is why engineers like it.
Yep.
And it's sort of an extension of how we interact with search engines and other thingswhere we sort of are now able to ask questions and get answers or suggestions.
(45:56):
But asking questions and getting answers or suggestions is either none or a very smallpart of most people's jobs.
uh It is more of what they do for leisure and recreation.
That's great.
So chat interfaces are great there.
But I think that we will do well to
figure out what the user interface and the interaction modality is of people interactingwith agents and agents interacting with underlying services uh in a way that makes sense
(46:34):
given the strengths and weaknesses of this technology.
But my guess is it's not gonna be the chat interface.
I think that's a fairly good prediction and I'm interested to see where all this goesbecause I don't feel like we've really had what I would call an iPhone moment yet uh with
AI in terms, yeah.
And you know, for me the iPhone moment was the first time I slipped my thumb to unlock it.
(46:58):
I'm like, okay, this is just totally different.
But you know, it's also, every technology is like that.
So I'm a theater buff, I love going to theater.
The initial movies, the first movies where they put a camera in the middle of theater andthey showed what was on stage, which is fine, but.
you know, it solved the problem.
(47:19):
If you couldn't be there, you could at least see it, great.
But that's not movies.
Charlie Chaplin came along and said, wait, what if I put a camera here and another oneover here and a third one there and then I edit that together and that became movies,
right?
And so I think that, you know, when we had the same thing with, you know, CD-ROMs wereinitially, you know, they were kind of janky and then they got better and internet...
(47:45):
applications were initially like CD-ROMs and then they got better and we added more andyou know, all of these things evolve when really, really creative people see the world as
it is, they see this new technology, they see the problems that exist and they go, huh, ifI have this and this and this, I can do that thing over there that no one's thought of but
(48:05):
is going to make a huge difference and it's those use cases where a lot of those werewaiting for.
I'm so glad um that you gave that example about early cinema, right?
Setting up the camera in the theater.
One of the examples I often give in workshops or talks is the early electrification offactories.
ah You know, you'll have a steam powered factory, but then we're just stringing lights init, right?
(48:29):
And it's the same thing with what you're speaking of.
We don't have those use cases.
We haven't figured it out yet.
And it's impossibly exciting.
And for folks that aren't thinking about it the way that you are, any of our listeners, Ireally encourage
to lean into that idea because we are at the very nascent stage of this technology fromwhere I'm sitting and we don't quite know what that killer app is going to be, what that
(48:53):
iPhone moment is going to be, what it's going to look like but that's why we got greatpeople like you inventing really cool technology to help us along the way.
It's fun to be part of it, but I can't wait to see what people do with it.
That's the really fun part.
Me too.
Speaking of inventing technology, ah believe Barn Door is coming up on its firstanniversary, is that correct?
(49:14):
are, technically we were incorporated a year ago, two days ago.
And yeah, we had to incorporate before we were able to accept money from investors.
And yeah, they have to wire it somewhere and to have a bank account you have to have acompany.
But yeah, so we're basically depending on when you count the first anniversary, we'rebasically there and you know a lot has happened.
(49:41):
Not just in our little company but in the world, in AI and you know it is moving at apretty steady clip.
You know we talked about MCP.
MCP is less than a year old and has undergone so many transformations so far that
uh you know, part of our job is to manage that and deal with that on behalf of ourcustomers so they don't have to.
(50:09):
um And, you know, I have two people whose full-time job is just basically paying attentionto what's happening at MCP because it's just, there's all kinds of wacky things going on.
Well, and I would imagine uh in addition to the typical pressure cooker of a startup inthe intensity of the first year, you're layering on the rapid development and iteration of
(50:32):
all the AI technology.
uh Plus, you're out there in the field working with customers, early customers.
What have you learned that you didn't expect when you started?
I mean, I know you founded a company before you quote been down this road before, but thisis a uniquely different road you're traveling now.
It is, think the biggest surprise so far has been the level to which really, reallysophisticated and large companies are jumping in very early.
(50:59):
you know, the API thing we went through was sort of a proxy for the mobile revolution.
But even that, you the concept of, you know, every company needs a mobile app and a mobilestrategy took a lot longer to happen.
you know, when we...
we would have to educate people.
Large hotel chain acts and they want to have the ability to book a hotel room on an app.
(51:24):
And I was like, well, what's that gonna talk to?
They're like, I don't know.
I said, well, it's this thing called an API.
This is why you need us, right?
And you now have mandates in banks and financial services companies to make this stuffhappen in a very secure way from day one.
And our...
(51:45):
Several of our first uh enterprise customers are folks who have very, very deep security,uh have to be private cloud or on premise.
We have to deal with deployment issues that weren't something that we had to deal with foryears back before.
(52:09):
that's just been really interesting to see the level of speed and enthusiasm.
that this is getting implemented.
cool.
We're recording this episode in mid-October.
It's going to air probably late November.
I know that you've got some stuff cooking.
ah Is there anything that you can share with us or talk about?
(52:33):
Yes.
have released uh the ability to, so MCP is a funny thing.
There's lots and lots of different servers out there.
Some of them are in and of themselves unsecure.
Many of them don't really have much capability.
They sort of limit the things you can do to things that could not possibly cause problems,but also could not possibly be productive.
(52:53):
um And so,
We have had a lot of demand from folks saying, help me just get connected to my varioussystems just as an individual, as an individual user.
And that is being launched prior to the end of November when this will, so if you come andsee us by the time this is launched, because uh the last bits of it are in the current
(53:15):
sprint.
So I'm not at all concerned that it will get out there.
As well as we're working on, and if it hasn't been launched, uh
yet, which I hope it is.
It will be launched shortly after this airs.
The ability for even enterprises to come and self-serve, just set up the system on theirown, take it for a run, do some stuff with it, and see what they can play with so that
(53:40):
they don't, know, engineers don't like talking to salespeople, go figure.
So we're going to make it possible for our platform to be.
test driven by anybody who wants to play with it.
very exciting very exciting alright let me ask you one more hardball question here you'reyou're working with a lot of leaders you're working with a lot of CEOs if you were a
(54:06):
middle market CEO today not a barn door which you are but if you were a different middlemarket CEO today what is the one thing that you do differently than you're seeing most
companies do with their AI strategies is there a zig when everybody is zagging that youwould encourage folks to think about making
I won't, know, obviously some folks are doing it really well, right?
So I wouldn't say that everyone is screwing it up, but I will say that um there is uh agap between this concept of we all know AI is important and actually execution and making
(54:43):
it happen.
And I think that a lot of companies are putting, so there's two things.
One is that a lot of companies have,
bought into this myth that Joe over in finance, you give him access to Claude and suddenlyhe's gonna figure out the way to automate his job away.
(55:06):
And we've had low code, no code things for a long time.
We've had no shortage of tools that purport to make it possible for non-technical peopleto make computers do technical jobs.
But fundamentally, the folks in those jobs aren't usually the ones who think the waycomputers do and understand how to apply these tools to solve those problems.
(55:35):
And I think that a lot of mid-market and large company CEOs feel like, buying thisplatform, I'm putting it out there, people should figure out how to use it.
And it still is the case that there's a group of people
that are particularly good at understanding and dissecting business challenges, looking atthe available tools and implementing those tools.
(56:00):
And sometimes they're consultants and sometimes they're people in your organization,right?
But there are folks who are, whose task it is to sort of go through, we used to run thesehackathons and we kind of do a similar thing now with AI.
It's like, get the people in the room, talk about it and have someone, we have people onour team who do this, someone who is really good at using AI to solve problems.
(56:21):
understand, try to understand your problems and teach you how to solve them rather thanassuming it's just gonna sort of happen on its own because it doesn't generally happen on
its own and and you know the other version of this is you have you know the sort of thesenior executive who's a year from retirement and you're like you run the innovation team
(56:42):
and make AI happen and that's not necessarily the best casting for that you know you youreally you know I look at
some of our younger team members make use of AI.
And I'm really impressed, right?
This is a generation that is pretty native on it.
(57:03):
I encourage them to teach the rest of us what we should be doing.
I can't think of a better parting thought than that.
uh Learning from the younger generation.
So, Orin, an hour goes by so quickly on this show.
I really enjoyed speaking with you today and I thought that this was a fantasticconversation.
(57:26):
Is there anything else that you want to leave us with beyond that mic drop of learningfrom the youth?
uh
I would plug my latest Broadway show, but I won't do that here.
okay, I'll plug my latest Broadway show.
So we are opening Queen of Versailles on Broadway with starring Kristin Chenoweth, who isin the regional cast of Wicked, written by Stephen Schwartz, who wrote Wicked.
(57:52):
It's his first new musical in 25 years.
Wow.
And it will be open, actually performances started this week, so by the time this airs,will have been open on Broadway, and you should definitely come see it.
If you see one show on Broadway, you wanna see Kristin Chenoweth perform this music, it'sincredible.
Fantastic.
uh I can almost guarantee that my daughter will be on the next train to New York.
(58:15):
uh She loves the live Broadway theater up there.
So that sounds great.
Thank you.
That's good.
uh People that want to get in touch with you, maybe they want to help you back your nextshow.
Maybe they just want to learn more about Barn Door.
I'm Orin at barndoror.ai.
Easiest way to reach me.
Okay, that's fair enough.
So I appreciate you sharing your insights with us today.
(58:37):
This was just wonderful.
Thanks so much, Chad.
Appreciate it.
Absolutely.
All right.
Our listeners, once again, we've come to the end of yet another episode of AI for theC-suite.
So thank you for tuning in.
We are committed to helping middle market leaders thrive during the exponential age.
And if you found value in today's discussion, be sure to subscribe to our podcast, followus on all the socials and visit our site, AI for the C-suite dot com.
(59:03):
Please join us next time as we continue to unlock the potential of AI for yourorganization and until then keep your algorithms running, your leadership evolving, and
your AI in check.
Thanks everybody.
Greetings, innovators and leaders.
This is AI for the C-suite, your compass for navigating the exponential age.
I'm your host, Chad Harvey, and we're here to bring you cutting edge insights on AItailored specifically for middle market organizations.
Buckle up as we embark on a journey through the transformative world of artificialintelligence.
(00:22):
Today I am pleased to welcome Orin Michaels, co-founder and CEO of barndoor.ai to AI forthe C-suite.
Orin brings a unique perspective to enterprise AI, having successfully navigated the earlydays of another transformative technology that changed how businesses operate.
Back in 2006, when most people had never heard of APIs, founded Mashery to help companiesmanage these critical business connections.
(00:50):
He built that company over seven years and sold it to Intel in 2013 just as mobile appswere exploding and every business needed an API strategy.
Today, Orin sees a similar pattern emerging with AI agents in the enterprise.
Last November, he co-founded Barn Door AI after recognizing that companies face the same
(01:11):
fundamental challenge they did with APIs.
do you safely connect powerful new tech to your critical business systems?
Barn door raised a 13.5 million seed round and has grown to a team of 22 building whatOrin calls the control plane for agentic AI in the enterprise.
What makes Orin's perspective particularly valuable is his focus on a problem many execsare just beginning to grapple with.
(01:35):
AI agents aren't just tools that provide information, they take actions.
And unlike humans,
employees, agents, have no conscience, no fear of consequences, and no inherentunderstanding of business context.
Today, we'll explore how middle market companies can harness the productivity benefits ofAI agents while maintaining the security and governance necessary to protect their
(01:58):
businesses.
Oren, welcome to AI for the C-suite.
Thanks Chad, that was lovely intro.
Thank you!
We put a lot of time crafting intros so I appreciate the props.
to help you craft that intro?
just a little bit yes just a little bit I've got a secret sauce here that's a sprinkle ofme a sprinkle you and a sprinkle something else so hopefully it comes off alright so let's
(02:23):
dive right in uh I mentioned in the intro that you are describing barn door as the controlplane for agentic AI and I would like to have you explain what a control plane is for
anybody that's not familiar with that terminology and why is it that enterprises uh
need one specifically for AI agents.
(02:45):
Well, when we hire humans, we give them permissions of what they're allowed to do and whatthey're not allowed to do.
We say, can use Salesforce and you can do these things in it.
You can use Google Docs and you can do those things in it and you can see these ones, butyou can't see those ones.
um agents and any AI for that matter that's going to access the same systems that we ashumans use need to have that level of governance as well.
(03:13):
And it already exists for the humans, but that's not quite good enough.
As you mentioned, the agents have a lot more of a blast radius.
They can do a lot more, a lot faster than I can.
They're not afraid of being fired.
And they're like enthusiastic interns.
They really want to just take their instructions and go do it, no matter what happens.
And so just saying that you have an identity system that says that Orin's allowed to haveaccess to these things is not enough if I start to use things that act.
(03:43):
on my behalf but independently without necessarily my governance and my conscience.
So when you think about each of those things needs to have that governance.
And there's going to be a lot of those things, right?
Right now we've got more AIs than you and I can name in the time we have here.
Yes.
And I just got it off a call with a very large financial institution where the CIO theresays he fully expects each of his employees to have hundreds or perhaps even thousands of
(04:12):
agents that they use in their job and he has 20,000 employees.
And so if you want to be able to start managing that, you can either try to go agent byagent, platform by platform, each of which presents its own interface or.
admin tools to try to make that happen, or you can say no, we're just going to putsomething in between that is a gateway, it's a controlling device, and it says here are
(04:40):
all the different things, and we understand what they're, what they're.
capabilities are and what their risks are and that sort of thing.
Here are all the humans and their roles and we understand what they do.
On the other side, here are all the tools people use.
Here's Salesforce and Snowflake and Notion and Google Docs and all those things.
And something called a control plane will manage the traffic that goes between them.
(05:02):
It will give the administrators the ability to see what these agents are trying to do.
So if you say to the Salesforce agent or an agent using Salesforce, know, go update myopportunities and it...
goes to try to delete them instead, well, you'll see that, right?
And you'll have rules in place to keep that from happening.
And so it manages this across many, many, many agents on one side and many, many, tools onthe other side with the humans sort of on the third side.
(05:30):
So how is this different from some of the security and access control systems thatcompanies already have in place that give them visibility into network traffic and uh some
level of control?
Yeah, well network traffic is important.
uh That is one layer and you think of the network layer models, right?
And so network traffic says, and it's important, for instance, suggest that you tell theMCP servers you're running or ours if you're using ours, that they can only accept network
(06:00):
traffic that comes through us.
So that's a particular kind of rule.
Right, a network traffic rule might say even, and there's authentication things andidentity that say, Oren is allowed to use this set of records in Salesforce and he's
allowed to update them.
But none at the packet level, you're not going to be able to say, Oren is a low levelsales employee, he's allowed to update opportunities with his agent, but only those that
(06:28):
are worth less than $2,000.
you
He's allowed to, as I do actually, I'm CEO with my agents.
When I finish a call, uh use Cloud to update Salesforce about it.
And I make it very clear that Cloud is allowed to create a new call record.
But it's not allowed to modify the ones that are there.
And it's not allowed to modify the underlying record or the account because that's notwhat I want it to be doing.
(06:53):
And although I'm allowed to do those things, I don't want to have to worry about it.
Mm-hmm.
I make sure that when I'm using Cloud for this task, it's only allowed to do the thingsthat it should be allowed to do given the context of what it's doing.
And that is not, you you talk about security systems, security systems are all designedaround, there's these nefarious people outside who are trying to penetrate and get in and
(07:16):
steal your stuff.
And we have to deal with that.
That's an important part of what we do.
It's what I sort of call security with a capital S.
That is, know, have a CISO and they do that.
But those folks are not used to saying here's a perfectly authorized human using a toolthey're allowed to use to access a system they're allowed to access, but.
That tool is doing stupid stuff because it wasn't necessarily, it didn't have the contextor it got a bad prompt or it just interpreted it wrong or it had a bad day or whatever
(07:47):
reason.
And so you want to be able to be sure that this kind of management is also in place andthat's not the role of traditional security.
Okay.
So help me understand, uh because a lot of our listeners are non-technical execs, as manyare, right?
So I'm a CEO.
(08:09):
Well, and I even just heard you in your prior explanation describe yourself as a low-levelsales employee.
you know, right.
So let's say that I'm leading an organization.
um What
is it that your technology is going to do to flag these bad use cases because again thecall is coming from inside the house here essentially so how are they going to determine
(08:35):
what is appropriate and what is not appropriate and how are they going to flag that
So that is set by a combination of people, right?
So again, going back to your security person, your security person, your CISO, or theirtitle is, they understand these outside actors and there's going to be a series of
policies they dictate.
But they don't know the ins and outs of Salesforce.
They don't know the ins and outs of Google Docs because that's not their world.
(08:56):
And so those policies will be set perhaps by a Salesforce admin or an HR team or whateverand have to be able to manage by them.
So when you say how does the system know it,
Well, the system knows it because we tell it.
mean, we or our customers tell it.
But what's more important is to say, how do you go about doing that?
And the answer is you do it the same way you would when you have that new intern show up.
(09:19):
And you say, OK, I'm going to give you a task, and I'm going to see how you attempt to dothat task.
And it's going to be a task that you may or may not have all the ability to go do rightnow.
But I want to see how you do that task and what your
Process is to do it so I can decide if I can trust you to really do it for real right andthat's sort of how we give people more and more authority and with these it's the same
(09:41):
sort of thing so you have you know you have Claude and you say Okay, got Claude, you knowgo clean up my opportunities And then you can look and see what how it interprets that and
what the specific calls are it is to salesforce
At that point in time, you may have shut all that stuff out.
You have given it only read access and it can't actually go do it.
It's gonna error out, which is fine because then you can say, okay, it's error out, buthow did it try?
(10:06):
When I gave it that instruction, did it try to do the things that were consistent withthat task?
And if the answer is yes, you start turning those on and you say, okay, now try it againand it succeeds and you begin to get to a place where this set of instructions going to
that.
AI tool using these set of enterprise tools are likely to not cause trouble and you'rewilling to bring that on.
(10:31):
You also will have, as we talked about, hundreds perhaps of these different agentsworking, depending on the context and what that agent's doing, that exact same agent may
have different permissions assigned to it.
So you may say, on the occasion I'm doing this particular task, I've noticed it's capableof...
interpreting it correctly that I'm allowed to have it right into opportunities.
(10:52):
But when it does that thing, sometimes it tries to do things it shouldn't, so I'm cuttingit off at that.
I find the intersection and the overlap between permissions and process for agentsabsolutely fascinating because as we are promised the era of true agentic AI, we're
promised this idea that agents are going to learn and they're going to be able to takeadditionally increased actions.
(11:17):
I just wonder how do we continue to evolve and adapt both the permissions and the processthrough the type of infrastructure that you're talking about to ensure that we are getting
the results that we desire.
Does that make sense?
If you could restate it, I'm not sure exactly what you're.
(11:38):
Sure.
So we talked to you, you give a very good uh answer about how this works, right, from anon-technical perspective.
We talked about policy, but we also talked about permissions and what is allowed andwhat's not allowed.
And it sounds as if there's also an overlap there in terms of the documented process thatcertain agents are going to follow.
(11:58):
And I wonder, as we unleash more uh self-evolving agents, how do we continue to ensurethat we understand the process
that they're following as well as adapt and update the permissions based on that changingenvironment.
Does that make sense?
Yeah.
I think that certainly in the current time horizon, we don't understand it.
(12:20):
And that's part of the problem, right?
We also don't understand humans, by the way.
And so I know, it's this thing, right?
And so when I have an employee, as much as I would love to drill into their brain andunderstand why they think that way, I can't go do that.
And I can't really do that with
AI agents either because we really don't understand how the technology works.
(12:40):
so instead, what you do with humans and what you do with agents is you observe how theyinteract with the world around them.
And is that consistent with the task at hand and are they doing things or attempting to dothings that don't make sense?
If the low level sales employee is trying to access records in the payroll team, that'sprobably an indication that person is doing stuff they shouldn't be doing.
(13:05):
And you'll want to shut that down and be aware of it.
I think that as the agents get more complex and as they interact with each other, havemulti-step agents and agent to agent and all these sorts of things.
you will still be in a situation where ultimately the best way to ensure that the rightthing is happening is to know what's going in, what's coming out, what's going between the
(13:31):
different agents, be aware of those lines of communication and set your rules there.
Because trying to set them within the AIs themselves, at least at this point, is notsomething that I would necessarily rely on being able to do.
Fair enough.
When do you typically find that you're getting the calls to get involved withorganizations that are either deploying, experimenting, or thinking about agents?
(13:58):
And I know the ideal solution here is, we want to be in upfront before you've even startedthinking about this.
But when do you typically start to get these calls?
Yeah, so and you know, we're talking about agents, but it's any AI that interacts with it,right?
So, you know, is Claude agentic?
I don't know, I mean, I type, I ask to do something, sometimes it acts on my behalf,sometimes it just gets me information.
(14:22):
you know, it's any AI that needs to talk to underlying systems, generally today using thisprotocol called MCP, which we can get into if you want to.
But.
Generally speaking, people contact us when the science experiment didn't really yield muchin the way of ROI because everybody was kind of afraid to actually let it do something.
(14:48):
Mm-hmm.
you had the side, you hooked it up and you said, I can list all my opportunities.
Well, I can do that in Salesforce too, that's not helpful, right?
What's helpful is for me to be able to bring in information from multiple different placesand I can, at the beginning of the week, can pull the schedules of all my salespeople,
can...
(15:10):
bump those up against the opportunities we have open, I can pull in all the Slack chatsthat we're talking about them and I can get an overview of what's likely to happen that
week and how I can be of assistance and when I should pop in and say hi to these people.
And so, you know, it's when they come to us when they generally have some known use case,there's something they wanna do that they think is gonna make them money or make them more
(15:35):
productive or whatever it is businesses wanna do, whatever their goals are.
and give them more time at home, whatever it could be.
And they're ready to really make this work for reals.
But it's scary and complicated and they see an array of all these different MCP serversout there, they can't really trust them, they don't know what they are, they're not.
(16:02):
They're sort of worried about what will happen and um and they want to be aware and becontrolling.
That's one place.
The other place we get called in is that we're able to deploy not we our standarddeployment is in the cloud or really loves the cloud but we can also deploy in private
clouds and on-prem.
So like we have major financial institutions that want to implement these technologies MCPand other technologies inside in their fire inside their firewall and
(16:31):
you know, the do-it-yourself versions out there are just not enterprise-grade for that,right?
So we're able to do that, and that allows us to serve a particularly security-consciousclass of customer.
you mentioned mcp and i do want to dive into that before because that is uh...
(16:54):
absolutely a very hot topic right now before we do that though i want to jump back in thethe way back machine here a little bit and uh...
i want to talk about your build of mashery and your observation about APIs right i knowthat you have said that the AI agentic challenge looks like the same movie that you saw
back in 06 and you've got wonderful movie posters behind you so
(17:16):
are Broadway posters.
These are shows.
These are Broadway.
okay.
I was thinking about the good night, good luck there that I saw with Clooney.
So, where are you now?
boy.
All right.
We need to explore that in a separate episode, I think.
uh So, uh what are the specific parallels that you're seeing right now?
And what were some of the lessons that you learned from the API era that might be relevanttoday?
(17:39):
Well, know, so in the API era and I started Mashery in 2006, okay, so that was beforethere was an iPhone and before there was really AWS to any real extent, we were one of
their first customers.
I had been running business development for the world's second best blog search engine.
The world didn't really need the first best one either, but there was some reallyinteresting technology we had that a large company wanted to integrate in their
(18:06):
capabilities.
And so they came to me and I was a biz dev guy.
and they said, we would like to access your API.
And I said, great.
So I went to our engineers, I said, we got this big deal, they'll pay us a bunch of money,all we have to do is give them our API.
And they're like, well, we don't have one and we can't.
And I'm like, well, why not?
Well, they gave me all these reasons that boiled down to they didn't want to build theinfrastructure around doing it.
(18:26):
And I thought, well, this is silly because I can't dev any biz if I can't make this thinghappen.
And I started realizing that.
as the world was sort of moving to these web services concept, this was gonna becomereally, really important.
And so I left my job there and I went to one of the investors who had been an investor inthis company, we having dinner and I was explaining this problem and...
(18:50):
He said, well, you should start a company to do that.
I said, great, I will.
So I did, and that became Mashery.
we began building, and what it was that was, back then, the engineers had the keys to thekingdom, and if you wanted someone to have access to your API, you found the dude who knew
where the appliance was and had the password to it and knew how to use Terminal and couldmake something, create something that maybe wouldn't get disrupted, but if someone tried
(19:16):
to use it too much, then you were all screwed.
And that just didn't seem
seemed like a way to run a business.
so now, what we did at that point was we turned it into a business tool that companiescould use.
Conveniently, Apple brought the iPhone out in the app store.
Everybody needed a mobile strategy.
Mobile apps, talked to APIs, and here we were.
(19:38):
And so now, what we're seeing is we have this AI, and the initial version of what we allthink of AI and businesses at least is, you know,
these chat interfaces where you, the very, very best and brightest uses and mosteconomically viable uses we've come up with for these chat interfaces are coding copilots.
(20:02):
Coding copilots are saving people a lot of money and a lot of time.
Why?
They mimic how coders work.
If a coder has code that's not working, they go to a super smart PhD, their CTO, and say,here's all my code, fix it for me.
And this,
They do, right?
that is something chat is incredibly good at doing.
So once again, the engineers built the tool that they can love.
(20:24):
And the rest of us are like, wait, I could use something like that if only.
And the if only is, if only it could talk to the stuff I use.
If only it could access the systems I can access.
And then we have this agentic thing coming along that wants to write to things, which isnot really what the role of your typical chat interface was, right?
(20:45):
And so now you're saying, okay, there's this thing and it needs access to the realbusiness stuff to go do things and be productive.
And just as that large company needed access to my stuff to make me money.
And so in that sense, it seems to be, there's a lot of parallels and in fact, you know,there's the technological challenges.
(21:08):
We are a proxy, we are a gateway between this and that.
That's essentially what API management is, right?
And so the way I viewed it um in terms of the spec of what we're building has an awful lotto do with it.
It's extended because it's not merely, there are API management companies and they have AIproducts, but again, they're all, they're based specifically on a key and an access
(21:32):
control and don't have the same level of visibility and management because that's not.
part of API management, so at the level beyond, the fact that you have access to this callis not sufficient to say that you should be allowed to do it.
I appreciate you walking me through that because I think the way that you answered thatactually does lend itself to us going in a little bit deeper on MCPs now since you talked
(21:57):
about all that connective tissue.
So for the folks out there that MCP uh is just uh they've never heard of it before.
Let's talk about model context protocol, MCP.
uh What is MCP?
Why does it matter?
why does it matter?
So what I mentioned, I started this company to help people connect to, connect their AIsto the systems they use.
(22:22):
And conveniently and nicely, the folks at Anthropic last November put out this protocolthat purports to make that easy and rational.
And what it essentially does, you have,
We'll spend a sec on APIs here.
APIs are essentially how computers talk to computers.
(22:43):
So if I want to interact with a particular application, I can do it through the userinterface or a computer over here can sort of bypass that and go straight to it with a
conduit that they each understand called an API.
And it has some security on it.
But an AI isn't a normal computer.
(23:03):
It's not just a here is my query.
to a, the AI act, or the computer accessing an API has been programmed by programmers whoknow the rules of that API and they say, well, if we want it to happen, you send a query
in this form and you'll get an answer in that form.
And that, and it's a known thing in programmers program that.
(23:24):
AIs don't work like that.
AIs like us, they're sort of ambiguous.
You're given something like, how am gonna go about doing this?
What tools do I have at my disposal?
So MCP is essentially, it is a menu of the various options and things and tools andcapabilities that some system somewhere is giving this AI as sort of the universe of toys
(23:52):
they can play with to answer that thing and then also provides it sort of the form that itwould need to play with that in terms of how it should make the request and what.
information it's going to get back.
And you might have multiples of these.
I could have my AI talking to Salesforce and then interacting with Notion and then sendingan email on my behalf and putting it into a PDF and dropping it at Google Drive.
(24:19):
So each of those things can be, has a select, a group of menu items that you're able toplay with.
And.
On a very rudimentary level, often, there's an ability to turn specific ones of these onand off in a fairly uh broad way.
(24:41):
And then, of course, we provide a much more sophisticated and nuanced way of managingthat.
But.
But fundamentally, it's saying to the AI, in a form the AI can understand, here are theexternal resources, tools, abilities, data at your disposal that you can use to either
answer questions or take the actions that your human is telling you you're supposed to do.
(25:08):
you
So I think that begs the question then because we know that the AI tools that are outthere, the large language models, the agents upon which they're built, they're statistical
and they're probabilistic in nature, right?
They're not true thinking machines.
So, right, okay.
So I'm an AI agent.
Well, I'm not an AI agent.
Let's say I'm an AI agent.
(25:29):
And uh I'm calling up the MCP server here and I'm looking at all the different thingsthere.
How am I going to make the determination probabilistically, statistically,
which one of those menu options I need to select to do my task.
How does that look?
Yeah, it's an interesting question and it also, know, so I would say that the depths ofhow AI models function and think is probably beyond the scope of today's podcast.
(26:01):
But ultimately, there are things you're going to want to do to influence the AI to choosethe right tools, the tools you'd like it to use.
So for instance,
or to use them in a certain way.
So for instance, uh I can interact with Cloud and Cloud like most AIs count these thingscalled tokens which are essentially a unit of measure that shows how much data is going in
(26:27):
and out of them that they have to deal with.
It's a bad definition but we'll go with it.
So if the AI says I want to pull up the
my contact, Chad Harvey in Salesforce, one way it could do it is it could assume thatthere's a Chad Harvey contacting and say I want the Chad Harvey contact.
(26:50):
Or the easier for the AI way is.
give me all the contacts and I'm look for contacts that sound like something like ChadHarvey, I'm gonna get the contact ID for it and I'm gonna go back and request it
specifically by the ID.
Well, that second way just burned a bunch of tokens because it looked at the entire thing,right?
This is how it usually runs if it's allowed to.
(27:11):
And so you now have other, there's a layer of sophistication that's being built into this.
It's not as important an issue today because
many of these AIs aren't charging you for tokens yet, but they will be, where our moresophisticated customers are like, we need your help managing that as well.
(27:32):
And so you're not just using the MCP to give it the full menu.
You're managing the waiter in the restaurant.
They're giving you the full menu, but they're also saying, today's special is this one.
And if you can use that one, you probably want to, and it's a better deal kind of thing.
Yeah.
Okay.
(27:53):
I think that that is a perfect focal point to go back to something you addressed earlierwhen you categorized agents as enthusiastic interns, right?
um And I think if I'm an executive and I'm thinking about deploying agents, that helps meunderstand uh the fact that the agents don't necessarily understand the consequences of
(28:14):
their actions.
um So when you're working with organizations and uh you come across this type of subjectmatter, how,
How else do you help execs think about deploying agents and what have you learned that isgood counsel for leaders when this topic comes up?
I think it starts with understand which agents are good at what.
(28:40):
Using the right tool for the job certainly helps, but it's also to understand what you'reexpecting them to do and how you're expecting them to do it so that you can instruct them
in a rational way and experiment with how you instruct them ah in order to
(29:01):
try to get to the most efficient and effective way to get the job done.
But it's, you you're right.
They're not, it's not linear.
It's not an easy thing to do as managing humans is an easy thing to do.
And so, you know, I think it's also that, other thing I advise executives to do is don'tput all your eggs in one basket.
(29:24):
Don't say, you know, I just saw, you know,
and that's gonna be the cool AI, I'm gonna use that one, right?
It is a cool AI, it's not the only cool AI, right?
And so, I was with a group of insurance executives yesterday, and one of those like, we'reimplementing Rider, and it's great, except when it's not, right?
(29:44):
And so, we encourage people to have a low bar for trying things.
And also to realize it's all changing so fast.
So we see this with the different models.
What ChatGBT is really, really good at today, they bring out a new model and suddenly,Gemini's better at that.
(30:04):
And so understanding when to use which of the arrows in your quiver uh is a big part ofthis and not insisting on sticking with ones just because it's the one you happen to be
used to use.
Or in one of the things that's coming across very clearly in our conversation today for meat least is that you are very comfortable speaking truth to power and for our audio only
(30:31):
listeners there's a poster we referenced earlier on the wall behind you good night goodluck which is all about speaking truth to power uh
yes it is uh...
and so i'm i'm curious how well does uh...
your message land with c e o's that don't want to hear some of what you're saying rightnow because you've got a what i would call a a spiky point of view uh...
(30:54):
occasionally about certain things which is probably contrary to a vision that a lot ofleaders have in their mind about the way this stuff is going to work
Well, I think it's the vision, I think the folks that would be more upset with the visionthat we have are the ones who want to be the only AI anyone needs, right?
(31:16):
And so if your software company acts, know, pick a SaaS company, they all have AI agents,and they're like, use my AI agent, it'll do everything you need.
Because that's what they want, but they won't, right?
And so the folks who I think are mostly pushed aside by this are the ones who have tobecome convinced that their AI has to play well with other people's AIs and that they're
(31:40):
not building the one and only, and no one is at this point, right?
Certainly not.
know, an agent built into an existing piece of software.
That, that, you know, no human only uses one piece of software to do their job.
So I don't know how you could expect any agent to only use one piece of software to dotheir job.
And so, you know, that, that little conversation is, is a little bit.
(32:04):
Dice here and you know the platform folks, you we run our our Company on platform X andtherefore will use their AI management.
Well, well, okay except when you want you know, you when your competitor is using thisother AI over here that isn't compatible with that and you know, that's problem, right so
You know, think a lot of companies are trying to to portray themselves as the only AI AIyou'll ever need
(32:31):
And I don't think that any company will ever be that, but certainly no company is thattoday.
When I go to our customers and talk to them, I'm not really speaking truth to power.
I'm basically saying, look, you have a desire to make this happen.
You recognize this is super important for your company, for your job, for your competitivepositioning, and you are unhappy with the productivity you're getting from AI so far.
(32:54):
And I know very few CEOs.
who will say, am really happy with our AI productivity so far.
Actually, I know one, but he's an unusual person running an unusual company, and he'sthrilled, and we should all be like him.
But the rest of them um are not very happy at all.
And so,
You know the other folks who we probably irritate a bit are the tinkerers who want to, I'mgonna download this thing and do this thing and I'm gonna sort of run everything through
(33:23):
me so I'm important.
And those folks don't necessarily like to relinquish power but we lived through thatbefore.
The other thing we lived through during the API time was of course the cloud revolution.
And you'll recall that when Mashery started in 2006,
My previous company, I had to raise $4 million before I wrote a line of code.
(33:47):
A million went to Sun for two servers.
A million went to Cisco for two routers.
I had to get a cage.
I had to buy a bunch of Oracle software.
All of that stuff's free now, right?
And so during that transition, I remember speaking at a conference at Interop, theconference for the big data center buyers, right?
And I was on the cloud panel because we were one of the first companies that deployed inthe cloud.
(34:10):
And the basic point of the cloud panel, so nostalgic there was such a thing, was we got upthere on all these
dinosaurs in the room were saying all the reasons why are we were wrong and the cloudwould never survive and I'm like luck we're in Vegas and interrupt and You're talking to
me your CEO is at a much better conference in Hawaii and she's learning that you can havedata centers with no capex who's gonna win that argument like just you guys are out of
(34:38):
your minds if you think this isn't gonna work and And further I'm sure you guys are allsuper smart, but I know Amazon is better at running data centers than you are
And so, this evolution is going to happen.
The smart CEOs get it.
And they just want to see it happen.
They want to see it happen now.
(34:58):
I want to see it happen now.
I'm not happy with the level of AI in my company.
I've got 22 people, all of whom know about using AI.
So I want more.
you
I want more.
That could be, I think, the theme of the era right now.
uh I want to go back and tug a little bit more on a topic that we started exploringearlier in terms of the visibility of what your platform enables.
(35:23):
And uh I'm interested in what patterns have you seen in the data or what patterns maybehave your customers seen that surprised them?
uh Things happening with their agents that they didn't expect or something else?
I don't know what.
Has there been anything surprising?
Well, I think the big surprise that everybody has to have with AI in general, and it'sjust amplified when it's agenting and it's taking actions, is when it does things that are
(35:53):
completely counter to what you tell it to do.
And you say to it, why did you do this?
This is not what I said.
And of course, it's always very apologetic.
Oh, I'm so sorry.
I didn't mean to do that.
And that
You know, that the concept of how you ask for what you want really, really matters.
(36:15):
And that I would say that the, and by the way, it's also not consistent.
You could have the same.
you know, I can say, you know, go do this thing in Salesforce that you did, that it didyesterday and today it says, I don't have access to Salesforce.
I'm like, well, you had it 10 minutes ago.
my goodness.
Yes, I do.
You know, and, and, uh, you know, so I think a lot of it is still, you know, it is thisincredibly powerful technology that often just seems kind of stupid and it, and, that you
(36:46):
have to understand that, you know, I'll go back to the cloud thing when we were, we wereon the first.
companies using EC2 that we bet the farm on it.
We were like, we're all in.
People are like, know, how can you manage something where you're not the ones controllingthe actual computers?
And I said, you know, the traditional data center management is you have these computersrunning and then you have things you do when they go down and you fix them.
(37:15):
And we decided to turn that on its head and say, we need to manage in such a way as weassume they'll often be down.
and understand how to always be up, even if they weren't, and figure out what that lookedlike.
And we built a bunch of tools around that that are now standard stuff.
But the mentality is, how do you leverage that uncertainty?
(37:37):
And how do you exist in it and still get the kind of return that you want?
And it's a lot of experimentation and a lot of accepting that there will be the occasionalbad outcome.
Mm-hmm.
there are with humans, humans do stupid stuff too, and that if you can mitigate the risksand manage the downside, if you can put guardrails up that make it so that when it doesn't
(38:05):
do the thing you thought it was gonna do, that's not the end of the world.
And then you figure out how to make it do what you want.
You raised a number of important points there, one of which I think is risk appetite andthe larger the organization, the less willing they are to take risks in general.
you seeing that mentality or that mindset shift at all given the rapid rise of AI and theneed to start to take risks and take more of a bottom-up approach in many instances?
(38:35):
Yeah, I did it, I was at a conference, a big CIO, major company CIO conference recently,and they had one of these things on stage where they asked this question, it responded in
an app in real time, and it was, with AI, do you want to be fast and risky or slow andcareful?
And these CIOs in the room were like 60, 40 fast and take risks.
(39:00):
Okay, which, know, okay, great, good.
um But it also depends on the kind of company, right?
So these financial service companies we're dealing with where they're insisting on usdeploying in private cloud and they're very, very, very, very careful, but also recognize
that they are competing with people who aren't.
(39:23):
And ah they have to manage that risk.
um And then there are companies that are really trying to say,
How do we mitigate this?
One company I talked to, a big e-commerce, big retailer uh in person, oh a bricks andmortar and e-commerce retailer, and they have a particular category of product that has a
lot of defects.
(39:43):
uh The defects arise when people bring the product home and install it.
And so they created an agent that has the ability when the customer service folks reach aparticular thing with someone who calls them or who chats with them or who comes in
through the agent, that if it's this category of product, it's under $200 and they'resaying certain things that make it likely that the product has failed, that this agent
(40:09):
will automatically order and drop ship a replacement.
But they've put guardrails around it where they say, I know
I'm taking some risks.
I'm gonna give away some widgets I should know.
Someone's gonna figure out how to take advantage of this.
But on the other hand, I'm saving a bunch of money from people going through a bunch ofmachinations before they do the inevitable and I just ordered the thing to replace this
(40:33):
thing, right?
And if the risk is greater than...
If the dollar cost of the bad outcomes is lower than the money saved by the good outcomesand you get customers who are happy because they get instant service, um it's a win.
(40:53):
And I think a lot of technology and automation uh falls into that.
You're never going to have no bad outcomes, but you have to manage the risk and manage thedownside.
And if you can, you're going to succeed and take advantage of the folks who are too riskaverse.
I think that's a very fair assessment.
(41:14):
And talking about money and risk.
our audience is predominantly middle market companies that don't have the resources of aFortune 500, but they've got the same challenges, quite frankly, especially.
Yeah.
What do you suggest as an approach for these organizations in terms of implementing agentgovernance?
(41:35):
Some of the things that we've been talking about today, what what are exploratory stepsthat they should be taking right now?
ah If say I'm not, you know, ah I
resource like a fortune 500 company.
and we have plans for teams that start at 500 bucks a month and are designed for SMBs,right?
So we're there, right there with you.
(41:58):
uh And the good thing about programs like that, whether it's ours or software in general,is you want to have the level of security and management afforded
to the enterprises, you just don't need bunch of the bells and whistles they do.
And so oh you look for that.
(42:21):
But I think the other thing is that the best thing you can do is talk to folks who aredoing this, that are your peers.
And we go, the thing that all of my team is laser focused on, I remind them about this allthe time, is this is so early, agentic ROI is like not much of a thing right now.
(42:42):
No.
And so we as technologists, we as humans are pattern matchers.
And so the more things we can go out and say, here is this little use case that someone'susing to make, to actually have positive ROI on an.
really getting something out of this agent, the more you can surface those things and hearabout how they were done, you then can pattern-match them.
(43:04):
Well, I don't have that problem, but that sounds a lot like this thing I do have, and ifthey could solve it with that way, this might give me some insight as to how to go do it.
And so I would encourage people to talk to folks even outside their businesses and otherkinds of companies, know, whenever you're at some event where you're talking to your
peers, just bring this up and say, you know, how are you?
(43:28):
making AI happen.
This is actually one of the great icebreakers that um some of the dinners I attend thatthis one organization uses.
And they go around the room and they say, other than coding copilots and customer servicechat bots, what is the most productive way you're using AI in your company?
And I've heard some really interesting things there.
And it's through those that the rest of the people in room go, huh, I hadn't thought oftrying to use it for that.
(43:52):
That's a really good, because at this point it's all new.
Right.
It is and I think you are spot on in your analysis and your remarks about this being earlydays and we really don't understand uh yet where this is all going yet.
You clearly have had the insight that we're going to need some different types ofinfrastructure to make all this work.
(44:15):
And I'm thinking back to a conversation I had uh much, much earlier in 2025 with PapiManon from outshift by Cisco and they are thinking very much about
what they're calling the Internet of Agents and what is going to be required, right, interms of infrastructure and whether our current Internet infrastructure can even, and
networking infrastructure is capable of withstanding that.
(44:36):
So, all that word salad out there.
What else are you seeing in the landscape that, uh much like you're reevaluating thisparticular component uh for AI, what else are you seeing out there that you think we need
to really be taking a hard look at and saying, hmm, I'm not sure this is going to serve uswell going forward?
know, AI specific, I would say, I actually don't believe that the chat interface survivesas the dominant way for enterprise, least, to deal with AI.
(45:09):
Because again, it's not how we ordinarily do our jobs.
We don't.
As you tend to chat, there's a whole reason why, you know, could say user interface folksexist because computers are hard to deal with and so it simplifies that so we humans can
do it.
Okay, fine.
But over time, we've found more efficient ways to give instructions and to get informationback.
(45:35):
And I think the chat interface, again, it's a combination of it's a great way to get yourcode fixed, which is why engineers like it.
Yep.
And it's sort of an extension of how we interact with search engines and other thingswhere we sort of are now able to ask questions and get answers or suggestions.
(45:56):
But asking questions and getting answers or suggestions is either none or a very smallpart of most people's jobs.
uh It is more of what they do for leisure and recreation.
That's great.
So chat interfaces are great there.
But I think that we will do well to
figure out what the user interface and the interaction modality is of people interactingwith agents and agents interacting with underlying services uh in a way that makes sense
(46:34):
given the strengths and weaknesses of this technology.
But my guess is it's not gonna be the chat interface.
I think that's a fairly good prediction and I'm interested to see where all this goesbecause I don't feel like we've really had what I would call an iPhone moment yet uh with
AI in terms, yeah.
And you know, for me the iPhone moment was the first time I slipped my thumb to unlock it.
(46:58):
I'm like, okay, this is just totally different.
But you know, it's also, every technology is like that.
So I'm a theater buff, I love going to theater.
The initial movies, the first movies where they put a camera in the middle of theater andthey showed what was on stage, which is fine, but.
you know, it solved the problem.
(47:19):
If you couldn't be there, you could at least see it, great.
But that's not movies.
Charlie Chaplin came along and said, wait, what if I put a camera here and another oneover here and a third one there and then I edit that together and that became movies,
right?
And so I think that, you know, when we had the same thing with, you know, CD-ROMs wereinitially, you know, they were kind of janky and then they got better and internet...
(47:45):
applications were initially like CD-ROMs and then they got better and we added more andyou know, all of these things evolve when really, really creative people see the world as
it is, they see this new technology, they see the problems that exist and they go, huh, ifI have this and this and this, I can do that thing over there that no one's thought of but
(48:05):
is going to make a huge difference and it's those use cases where a lot of those werewaiting for.
I'm so glad um that you gave that example about early cinema, right?
Setting up the camera in the theater.
One of the examples I often give in workshops or talks is the early electrification offactories.
ah You know, you'll have a steam powered factory, but then we're just stringing lights init, right?
(48:29):
And it's the same thing with what you're speaking of.
We don't have those use cases.
We haven't figured it out yet.
And it's impossibly exciting.
And for folks that aren't thinking about it the way that you are, any of our listeners, Ireally encourage
to lean into that idea because we are at the very nascent stage of this technology fromwhere I'm sitting and we don't quite know what that killer app is going to be, what that
(48:53):
iPhone moment is going to be, what it's going to look like but that's why we got greatpeople like you inventing really cool technology to help us along the way.
It's fun to be part of it, but I can't wait to see what people do with it.
That's the really fun part.
Me too.
Speaking of inventing technology, ah believe Barn Door is coming up on its firstanniversary, is that correct?
(49:14):
are, technically we were incorporated a year ago, two days ago.
And yeah, we had to incorporate before we were able to accept money from investors.
And yeah, they have to wire it somewhere and to have a bank account you have to have acompany.
But yeah, so we're basically depending on when you count the first anniversary, we'rebasically there and you know a lot has happened.
(49:41):
Not just in our little company but in the world, in AI and you know it is moving at apretty steady clip.
You know we talked about MCP.
MCP is less than a year old and has undergone so many transformations so far that
uh you know, part of our job is to manage that and deal with that on behalf of ourcustomers so they don't have to.
(50:09):
um And, you know, I have two people whose full-time job is just basically paying attentionto what's happening at MCP because it's just, there's all kinds of wacky things going on.
Well, and I would imagine uh in addition to the typical pressure cooker of a startup inthe intensity of the first year, you're layering on the rapid development and iteration of
(50:32):
all the AI technology.
uh Plus, you're out there in the field working with customers, early customers.
What have you learned that you didn't expect when you started?
I mean, I know you founded a company before you quote been down this road before, but thisis a uniquely different road you're traveling now.
It is, think the biggest surprise so far has been the level to which really, reallysophisticated and large companies are jumping in very early.
(50:59):
you know, the API thing we went through was sort of a proxy for the mobile revolution.
But even that, you the concept of, you know, every company needs a mobile app and a mobilestrategy took a lot longer to happen.
you know, when we...
we would have to educate people.
Large hotel chain acts and they want to have the ability to book a hotel room on an app.
(51:24):
And I was like, well, what's that gonna talk to?
They're like, I don't know.
I said, well, it's this thing called an API.
This is why you need us, right?
And you now have mandates in banks and financial services companies to make this stuffhappen in a very secure way from day one.
And our...
(51:45):
Several of our first uh enterprise customers are folks who have very, very deep security,uh have to be private cloud or on premise.
We have to deal with deployment issues that weren't something that we had to deal with foryears back before.
(52:09):
that's just been really interesting to see the level of speed and enthusiasm.
that this is getting implemented.
cool.
We're recording this episode in mid-October.
It's going to air probably late November.
I know that you've got some stuff cooking.
ah Is there anything that you can share with us or talk about?
(52:33):
Yes.
have released uh the ability to, so MCP is a funny thing.
There's lots and lots of different servers out there.
Some of them are in and of themselves unsecure.
Many of them don't really have much capability.
They sort of limit the things you can do to things that could not possibly cause problems,but also could not possibly be productive.
(52:53):
um And so,
We have had a lot of demand from folks saying, help me just get connected to my varioussystems just as an individual, as an individual user.
And that is being launched prior to the end of November when this will, so if you come andsee us by the time this is launched, because uh the last bits of it are in the current
(53:15):
sprint.
So I'm not at all concerned that it will get out there.
As well as we're working on, and if it hasn't been launched, uh
yet, which I hope it is.
It will be launched shortly after this airs.
The ability for even enterprises to come and self-serve, just set up the system on theirown, take it for a run, do some stuff with it, and see what they can play with so that
(53:40):
they don't, know, engineers don't like talking to salespeople, go figure.
So we're going to make it possible for our platform to be.
test driven by anybody who wants to play with it.
very exciting very exciting alright let me ask you one more hardball question here you'reyou're working with a lot of leaders you're working with a lot of CEOs if you were a
(54:06):
middle market CEO today not a barn door which you are but if you were a different middlemarket CEO today what is the one thing that you do differently than you're seeing most
companies do with their AI strategies is there a zig when everybody is zagging that youwould encourage folks to think about making
I won't, know, obviously some folks are doing it really well, right?
So I wouldn't say that everyone is screwing it up, but I will say that um there is uh agap between this concept of we all know AI is important and actually execution and making
(54:43):
it happen.
And I think that a lot of companies are putting, so there's two things.
One is that a lot of companies have,
bought into this myth that Joe over in finance, you give him access to Claude and suddenlyhe's gonna figure out the way to automate his job away.
(55:06):
And we've had low code, no code things for a long time.
We've had no shortage of tools that purport to make it possible for non-technical peopleto make computers do technical jobs.
But fundamentally, the folks in those jobs aren't usually the ones who think the waycomputers do and understand how to apply these tools to solve those problems.
(55:35):
And I think that a lot of mid-market and large company CEOs feel like, buying thisplatform, I'm putting it out there, people should figure out how to use it.
And it still is the case that there's a group of people
that are particularly good at understanding and dissecting business challenges, looking atthe available tools and implementing those tools.
(56:00):
And sometimes they're consultants and sometimes they're people in your organization,right?
But there are folks who are, whose task it is to sort of go through, we used to run thesehackathons and we kind of do a similar thing now with AI.
It's like, get the people in the room, talk about it and have someone, we have people onour team who do this, someone who is really good at using AI to solve problems.
(56:21):
understand, try to understand your problems and teach you how to solve them rather thanassuming it's just gonna sort of happen on its own because it doesn't generally happen on
its own and and you know the other version of this is you have you know the sort of thesenior executive who's a year from retirement and you're like you run the innovation team
(56:42):
and make AI happen and that's not necessarily the best casting for that you know you youreally you know I look at
some of our younger team members make use of AI.
And I'm really impressed, right?
This is a generation that is pretty native on it.
(57:03):
I encourage them to teach the rest of us what we should be doing.
I can't think of a better parting thought than that.
uh Learning from the younger generation.
So, Orin, an hour goes by so quickly on this show.
I really enjoyed speaking with you today and I thought that this was a fantasticconversation.
(57:26):
Is there anything else that you want to leave us with beyond that mic drop of learningfrom the youth?
uh
I would plug my latest Broadway show, but I won't do that here.
okay, I'll plug my latest Broadway show.
So we are opening Queen of Versailles on Broadway with starring Kristin Chenoweth, who isin the regional cast of Wicked, written by Stephen Schwartz, who wrote Wicked.
(57:52):
It's his first new musical in 25 years.
Wow.
And it will be open, actually performances started this week, so by the time this airs,will have been open on Broadway, and you should definitely come see it.
If you see one show on Broadway, you wanna see Kristin Chenoweth perform this music, it'sincredible.
Fantastic.
uh I can almost guarantee that my daughter will be on the next train to New York.
(58:15):
uh She loves the live Broadway theater up there.
So that sounds great.
Thank you.
That's good.
uh People that want to get in touch with you, maybe they want to help you back your nextshow.
Maybe they just want to learn more about Barn Door.
I'm Orin at barndoror.ai.
Easiest way to reach me.
Okay, that's fair enough.
So I appreciate you sharing your insights with us today.
(58:37):
This was just wonderful.
Thanks so much, Chad.
Appreciate it.
Absolutely.
All right.
Our listeners, once again, we've come to the end of yet another episode of AI for theC-suite.
So thank you for tuning in.
We are committed to helping middle market leaders thrive during the exponential age.
And if you found value in today's discussion, be sure to subscribe to our podcast, followus on all the socials and visit our site, AI for the C-suite dot com.
(59:03):
Please join us next time as we continue to unlock the potential of AI for yourorganization and until then keep your algorithms running, your leadership evolving, and
your AI in check.
Thanks everybody.
Popular Podcasts
Las Culturistas with Matt Rogers and Bowen Yang
Ding dong! Join your culture consultants, Matt Rogers and Bowen Yang, on an unforgettable journey into the beating heart of CULTURE. Alongside sizzling special guests, they GET INTO the hottest pop-culture moments of the day and the formative cultural experiences that turned them into Culturistas. Produced by the Big Money Players Network and iHeartRadio.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.