In this episode, I sat down with Arya Bolurfrushan, founder and CEO of Applied AI, to talk about what most companies are getting wrong about AI.
Instead of asking where AI fits into existing workflows, Arya flips the question entirely — where do humans belong in an AI-driven system?
We get into:
- Why most AI pilots fail
- The real bottleneck in AI adoption
- How companies are unlocking 10x productivity
- Why legacy workflows are built around human limitations
- What leaders need to do right now to stay competitive
This is a practical and honest look at what it actually takes to implement AI inside real organizations.
Full episode: https://youtu.be/BYjN58XyJKQ Learn more about Arya’s work: https://opus.com
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:02):
Right, and we're live now, perfect.
I'm Chad Harvey, and this is AI for the C-suite, the show for senior leaders who know thatAI matters and need to figure out what to do about it.
Each episode, we dig into what it all actually means for middle market organizations, howit's changing decisions, strategy, leadership, and the nature of work.
(00:24):
Let's get into it.
I'm very excited about today's conversation because my guest isn't just building AI tools,
He's rethinking how humans and AI should work together, especially in industries wheremistakes aren't an option.
Arya Balorfushan is the founder and CEO of Applied AI, which he deliberately calls theworld's most boring AI company.
(00:48):
That's not false modesty, it's strategy.
Arya's team focuses on automating the critical, regulated workflows in healthcare,banking, and insurance where you can't afford hallucinations or black box decisions.
Arya brings a unique lens to this work.
He started at Goldman Sachs, earned his MBA at Harvard, and took a company public on theOslo Stock Exchange at 29.
(01:10):
Now he's based in Abu Dhabi, building AI that keeps humans in the loop in a veryintentional way.
We're going to dig into a question I think every leader needs to wrestle with.
When AI can do so much, where do humans actually belong?
Arya, welcome to AI for the C-suite.
Thank you very much, Chad.
It's a pleasure and honor to be here.
(01:32):
Very excited to get into our discussion today.
So I'm going to jump right into it with something that you said to me when we were talkingabout this conversation a few weeks ago.
You said to me that most companies are asking the wrong question about AI.
And instead of asking the question, where do I put AI in my workflow?
Your argument is that they should be asking, where do I put humans in an AI workflow?
(01:57):
And that's a pretty significant inversion.
I want to know what you meant by that.
Yeah, so we've been at this for a little over four years now.
So kind of before AI was cool.
And the reason I bring that up is we've been at the front lines deploying these things inmission critical industries and have seen the realities of implementation of production
(02:18):
and also have made.
you know, all the mistakes you can imagine and kind of iterated very quickly on them.
There's, and we've tried to kind of incorporate those into Opus, which is the thirdgeneration of our software.
And uh one of things we realized is anything but business process re-engineering,re-imagination is a band aid.
(02:47):
Mm-hmm.
And the kind of limiting factor is not really technology, but the limiting factor is thehuman mind.
If you uh give somebody a blank canvas with a low code solution, the human mind just kindof recreates the old process again.
(03:07):
And then we'll think about where to add an AI agent to find a local optimum to increaseproductivity by 20%, 30%.
Where we've seen the real uptick in productivity is when you look for a global optimum,where essentially you say uh these legacy processes that were written in the 90s have been
(03:31):
kind of an amalgamation of a bunch of things has a lot of false constraints in there thatwere designed for the limitations of the human.
uh
And once we reimagine it and ask the question, imagine the whole thing is AI.
And where do we want to intentionally put humans that are needed for supervision, forchecks, for edge cases?
(03:56):
Then you see an increase of productivity by 10x, 15x.
um And it's actually a very hard, hard, hard process to do well.
um
And we think, AI can also help us.
We call it workflow generation, uh to kind of remap us onto the map of possibilities forus to begin from a different starting point.
(04:27):
I love that you focused and went right to the psychological component here, the human mindcomponent.
I think you said something to the effect, if we give someone a blank canvas, they're justgoing to reinvent what they already know.
And I think that that's a real challenge.
So when you walk into an organization, how do get people to flip their mental model?
(04:48):
Yeah, so there's a.
to be very clear-eyed about the realities and deliver kind of the bad news.
That's what we start with.
One of which is this won't necessarily succeed in a centralized fashion.
(05:13):
So it's not IT internally that can do this.
It's not management.
A lot of times when we first started, every time there's an AI initiative and a bigannouncement for a company,
The people on the front lines, the process owners, roll their eyes.
Because they know this is just another management initiative who doesn't know what itthey're doing.
And they have to build a workaround on the solutions imposed on them.
(05:37):
That's why we've seen AI adoption falter.
That's why 95 % of these pilots fail.
So we think it's most powerful to do the reimagination at the edge of enterprise.
Mm-hmm.
way, where you have the business process owner involved, the non-technical user be the onewho's building these workflows and iterating on them.
(06:03):
Once they've built consensus at the edge, only then the guardrails of IT for security,cetera, are kind of instilled.
So uh in a way, we see that being most powerful is not to focus on
their existing process and how they do things, focus on what we call the intention or thejob to be done.
(06:28):
What are you really trying to accomplish here?
If it's, I'm trying to make sure there's no fraud in this claim.
Okay, that's the intention.
How you do it is secondary.
How you did it before is secondary.
So that initial prompt that will go into Opus is, what's the job to be done?
And then we append to it all of the constraints, which include the policies that theenterprise has, law, anything else that you put in.
(06:57):
And then our large work model, which has been trained on 1.6 million workflows, willgenerate the first rendition.
It says, you know what?
Here's how it can be done.
And very clearly, it will not be 100 % correct.
But it'll get you 80 % of the way there.
So once you look at it for the first time, you say, oh my god, all of these things that Ithought I needed, I don't need.
(07:23):
And here's where the only places that we recommend humans to be.
And then you do 30 iterations to that as the process owner.
Oh, you forgot about this edge case, you forgot about that, and you add to that.
And after one day, you have a 90 % accurate in-production workflow that you test.
(07:45):
Let's have some files that we run through it and then what goes out.
So having AI generate the first proposal lays bare to you all of the self-limiting uhbeliefs that were in the functional requirements.
(08:07):
1.6 million workflows that really jumped out oh at me in terms of the training base thatyou've got.
What have you found as you've evolved the model?
you finding you?
You gave an example.
You said we run it through.
We get the workflow and 40 iterations later or so and you know one full work day.
We've got something that works.
(08:27):
Have you found that that time is compressing as the models learn more about how to betteriterate uh ideal workflows based on the client?
I'll give you two answers.
Number one is the marginal workflow you built within your enterprise gets exceedinglybetter.
(08:49):
We call that kind of enterprise memory.
Once your policies are uploaded and it knows how you do things, then the first one thatyou generate is not 80 % away there, it's 85 % away there.
Then it's 90 % away there.
The time to value will go down dramatically.
Mm-hmm.
over time.
uh For a new enterprise or if you're on opus.com as kind of a self-serve user, you willbenefit from the kind of increasing number of workflows in our work knowledge graph,
(09:21):
right?
So when the 1.6 because 1.7, 1.8, the kind of base reality and best practices improve anduh as kind of LLMs also improve because we're LLM agnostic, you'll also see benefits
there.
You also said something that jumped out at me and I want to make sure that I understandand maybe I'm even splitting a hair here.
(09:44):
You talked about the intention of the user and the intention of the existing workflowprocess and taking that intention and dropping it into the AI to iterate the new process.
I often think about
StephenCovey101, starting with the end in mind, what's the result?
And it seems like starting with the intention is a flavor of that, distinctly different.
(10:05):
Agree, disagree.
Yeah, so I mean, we follow the Clayton Christensen kind of approach of what's the job tobe done.
What is it that you're trying to really accomplish?
eh And then what's the output that you need to accomplish that?
So it's almost, and there's a big philosophical thing here where human intent and machineexecution in a way, and then human checks.
(10:35):
Mm-hmm.
a way to think about it where there's...
um
You know, like, I'll give you a very clear example to kind of make this all real.
(10:55):
One of the workflows that we began with was independent medical examinations in the US,where it's a mission-critical kind of high-cost-of-error workflow in the sense that
there's a human life on the line.
And it's usually when...
ah your uh insurance company refuses to pay your insurance claim uh and then you sue.
(11:20):
So the law in almost every state is uh they subpoena all your providers.
You can get your entire medical history in this Frankenstein document that's like 30,000pages.
And then an independent medical examiner must adjudicate whether they should pay you ornot.
And it's usually being litigated.
So it's a pretty high cost of error.
(11:43):
If you can't hallucinate, you can't have all these kind of risks there.
And it's around a $5 billion industry.
And it's a private equity and a BPO darling.
uh
because it's a cash cow as you can imagine.
uh the majority of the time the man hours spent in the industry is organizing pages.
(12:07):
Because you can imagine, because you have to organize then summarize.
And no human can summarize before they organize a 30,000 page PDF.
You can't scroll up and down.
You have to put all the, in overall chronological order, in types, in various different,all duplicates must be removed.
(12:28):
So, on the 48 hours that it takes to have this organized and summarized, BPO firms chargeyou 18 to 36 cents a page.
and there's about 90 % accuracy.
We came in and then initially we were doing it the old way without inverting the questionand we got great results.
(12:52):
We were 30 % faster, 30 % cheaper, it was good.
But then we generated the workflow from the intention of I want to know if this insuranceclaim should be paid or not.
That's the intention.
The output is a sorted medical document with a summarization of all of it and medicalcodes, but the whole point is, is this person sick enough that they should stay in
(13:15):
hospital?
Or can they go back to work?
And then what Opus did was when it generated the workflow, it removed the organized phase.
It just summarized it.
The entire organizing of pages was a human limitation of our eyes.
(13:35):
and mental capacity to scroll up and down.
So it went from 48 hours to 4 minutes and it went from 36 cents a page to 1 2 cents a pageand it went from 90 % accuracy to 95 % accuracy.
Just because humans make mistakes you can't do 30,000 pages without making a mistake.
(13:56):
So who pays that cost?
Ultimately all of us do in our insurance premiums.
Mm-hmm.
So there's this global administration attack.
So that's kind of why intention versus just the output is a more effective means togenerate.
That's a really good example and I appreciate you walking us through there.
(14:18):
One of the things that I've been finding in my conversations with folks that aredeveloping, deploying AI technology, as well as working with organizations to help
accelerate their AI transformation is what I would characterize as inherent pushback byhumans against mistakes made by AI and a greater level of grace afforded to humans that
(14:39):
make mistakes, right?
By other humans.
And
Fascinating.
that you just gave, I would imagine that even in an ideal situation with a human reviewerand a human organizer, there will be multiple mistakes or errors of some sort made in that
final product.
Yet we afford that level of grace because it's a human.
(15:00):
And I'm wondering, have you experienced any pushback if an error has been found after thefact.
when it's been organized and summarized by the AI.
Have you found that there's any inherent pushback out there right now about this, or doesthe speed and the decreased cost offset that?
uh
I think it's a very deep, deep point.
(15:25):
So one of the other realities and the bad news that we kind of share when we do thesethings is one of the blockers of adoption is sabotage.
AI is scarier than other humans doing it because your livelihood is on the line.
Like people look at it as a direct threat.
(15:46):
So they look for errors.
as a way to basically do a gotcha moment and have the pilot fail and say, he was better.
uh So we've also seen pretty sophisticated sabotage in the sense of having the pagesmissing in the document or white pages.
(16:07):
So to make the AI make a mistake and then point at it and be like, that didn't work out.
uh Which, by the way, is not.
I understand it, it's not that malicious, it's just human nature to realize what's coming.
So way that we tackle that is, one, we call it supervised automation, where every singledocument has had the human review.
(16:33):
em the AI does the carbs and the human does the protein, where it's almost like you just.
focused on the edge cases, on the parts that require judgment, on the part that the AI haslow confidence in.
uh And this way, the human fingerprints are still on it.
Someone holds liability.
(16:55):
You have a throat to choke in the sense of this person made the mistake.
uh So when the kind of end user sees that, they don't see the AI's making a mistake.
They see this human, this manager hadn't checked.
the work of the AI.
So that kind of empathy is still there.
(17:15):
But also, we are on the side that humans will stay employed.
And the future is this marriage between AI and humans.
And this way, we keep the human relevant.
So it sounds like you've found less, I'm going to call it organizational inertia andresistance with respect to this type of sabotage than individual resistance.
(17:41):
Is that fair or do you find an equal amount of individual and organizational resistanceand inertia?
It goes back to one of the earlier points on if you don't include the business processowners as part of the solution, then they see it as the solution being imposed on them.
And then you see individual sabotage and kind of organizational.
(18:04):
We call it an antibody reaction.
That comes up.
But when you actually give them agency and you give them a sense that they're building it,
ah It's kind of the Ikea effect, ah where now you feel like it's your workflow.
You built it.
It understands you.
ah And then we see a big decline in sabotage.
(18:29):
Got it.
And it's your fault when the table collapses, right?
Okay.
You didn't build it well enough or you forgot to that screw in.
Got it.
So let me ask one other question on what you just outlined before we switch gears.
I was very interested and I think you had in our uh pre call talked about mid mile versuslast mile checking and I think that's what you were just describing right now.
(18:53):
And what I'm interested in is how do you insert the mid mile versus last mile checking ina continuous oversight model without creating new bottlenecks where the humans become
those bottlenecks?
Yeah, so I realize that this is not as intuitive for lot of listeners, so let me just get,we kind of use this.
(19:19):
But the main metric is increasing human productivity, which is output per person per day.
We saw when the human review was only at the end, last mile, like the end of the workflow,AI does everything, and then the human checks, productivity started going down sometimes.
uh because of a compound error problem, where you weren't sure where the probabilisticcode made a small error.
(19:46):
So you have to go back 36 agents and you spend more time fixing than actually just doingit.
So the fundamental issue with agentic workflows, as you know, uh probabilistic code and uhdeterministic outcome requirements.
having the human reviews actually mid-mile, which means like in the middle of theworkflow, this this dance, AI works, human works, AI works, human works, you kind of reset
(20:19):
the probabilistic drift in the middle of these workflows, and you catch the errors muchearlier.
So then your productivity really goes up.
Now, your question is very,
a very correct chat because you now have to kind of leave the workflow and wait for thehuman.
(20:39):
That's also a very expensive thing because that pipeline was spun up in the cloud and youcan't wait four or five hours.
So we actually have to spin that whole pipeline back down, wait for the human input, spinit back up, run the next thing and spin it back down or else the cost will be kind of
astronomical to have all these pipelines kind of waiting.
(21:02):
uh So we actually within Opus have a system where it goes out and every human has theirreviews for the day.
So you're notified, it knows which human is online, which one's offline, uh almost like anUber model.
uh And then they kind of, and you add a deadline, say four hours, three hours, two hoursfor a response.
(21:23):
uh And then we have a dispatch kind of engine, essentially.
that will call a human at right time.
That was very helpful area and I said we're going to shift gears, but it actually unlockedanother question I want to drill into here.
It what you just outlined there reminded me a lot of the conversations I've had with folksregarding software development and how a lot of the coding is now moving into an automated
(21:51):
realm and coders are needing to become more editors uh of the code, right?
Very different skill set and what you just described is uh human checkpoints that are not.
necessarily in production, they are more in an editor role, for lack of better word.
So here's my question.
Are you finding that the production level skills translate fairly well to an editor role,or are there new skills that those folks are needing to learn as they become that mid-mile
(22:20):
checkpoint?
So I will give you a quick answer and one that we are very passionate about.
(22:41):
The humans are moving from a maker to a checker, to use banking terminology.
We're all becoming managers.
We're all supervisors, in a way.
And we're managing AI agents.
The issue is we can't walk the halls.
(23:03):
So in that, it's a challenge of design.
And the onus is on us to kind of showcase the AI recommendation in a way that the humancan easily interact with it.
So the training is more, how do I figure out what I'm looking at?
How does the AI communicate to me what it's sure, what it's not sure of?
(23:25):
And how do I optimize the UI UX in a way where I can respond very easily and quickly?
I don't have to review a 30,000 page file every time there's a checkpoint.
That defeats the purpose.
So there's an element of...
uh of human-computer interaction design, uh which we think is the singular long-term thingto optimize.
(23:51):
That's one piece.
Number two.
So this is, I think, a very existential point on a species level, where for labor to beable to compete in a post-AI world, it needs to reprice itself or re-unitize a unit of
(24:13):
labor from time to output.
Where the old paradigm is, why would I pay you more if it takes you two hours?
I'll pay you more if takes you two minutes.
And that paradigm has there's a lot of retraining.
(24:34):
yeah, and I think by the way, that already happened in the transportation industry, right?
um So I do think there'll be a new breed of lawyer that will charge you by output and giveyou faster results.
So the consumer will win, but
(24:55):
uh In a world of abundance, the thing that's most scarce, the only thing that may bescarce is time.
uh And why are we selling that, essentially?
uh So I think there will be a world in which you have to incentivize the increasedproductivity and not penalize it.
(25:24):
If I'm getting paid for an hour, why would I do 30 times more cases?
I'll do my three or four.
uh So the productivity gains that uh these kind of agentic workflows unlock must be sharedbetween the capital owners and the humans.
ah Or else, I think, and that's a major retraining of not only the human.
(25:50):
in the loop, also the enterprise and the payment apparatus around that.
Only when that's done properly, uh then I think you will see massive increases inproductivity.
(26:12):
So let's stick with this thread for a few moments because I think.
This goes directly to something you and I had previously talked about.
I actually I wrote this down.
So I'm going to quote something you said that directly pertains to your answer here.
You had said to me during our prior call, if we don't share productivity gains with labor,there will be no place for capital owners to hide.
(26:32):
And I couldn't agree more with that, quite frankly.
You even referenced the idea that it took two world wars for us to digest the last majorproductivity gains from mechanization and industrialization.
Yeah.
think that's a pretty stark warning and I'm bringing this up right now because I thinkyou've outlined where the nature of productivity is going.
You talked about output versus hours worked, right?
(26:54):
And I think you can't have this conversation without a larger conversation here about whatthe impact is on capital owners and business leaders.
And I'm interested in exploring this a little bit.
uh I know it's not a direct takeaway maybe for folks in our audience, but this is wherethe puck is moving to quote Gretzky.
I don't think there's any question.
Yeah, we're going through a major shift and we don't entirely know what it's going to looklike on the other side.
(27:19):
So from your point of view, you're on the front lines, you're developing this technology,you're working with a lot of folks, you're seeing the impact it's having now, you're
projecting forward.
What does sharing those productivity gains look like?
And is this something that self-interest is going to drive or is there another motivatingfactor here?
(27:45):
So I think we had a town hall yesterday and I kind of walked through this with the team.
I think if we don't get this right, nothing else kind of matters.
ah Where, you know, the last two world wars and kind of the rise of fascism...
(28:07):
and communism was literally, if you read those books again, the alienation of labor, thesurplus profit, a critique to capital owners, these are all the reasons why we had these
battles these days.
And um we also fought over the source of the gain of productivity, which is energy andoil.
(28:32):
So this gets very real very quick.
Yeah.
Um
But in that case, with hydrocarbons and electricity, we still had minimum wage.
And there were still uprisings.
In this case, there's a good chance to have no wage.
(28:55):
And in particular, in democracies, you're going to have some major disruption to theprocess.
And I do think it may be of interest to your audience too because on the supply side, thecost of delivery of the goods and services will go down with automation.
(29:23):
So we can increase our gross margins, we can do all the good things to kind of increaseefficiencies.
But if no one's on the other side to buy them, so on the demand side, it doesn't matter.
uh
if prices equally drop, automatically.
Or if there's no safety, where do you sell it to?
So I think it's extremely important that we get this right.
(29:46):
um And there are things that kind of block us that we've to go through, like minimum wagelaws.
And there's a lot of these kind of issues where we're mandating to pay people by time.
ah And then a lot of this work goes offshore.
ah
So there is this kind of on-shoring of knowledge work.
(30:08):
There's a billion knowledge workers in the world earning 12 trillion in salary.
The mass majority of is outside the US.
I think if we, to all folks who can, and who hear this, whether you're regulators or theprivate sector, public sector, figure out a way to share these gains of productivity with
(30:43):
humans.
uh And I think that, in fact, we may regulate ourselves into relevancy in the sense thatit's not a free market.
the government is still run by humans.
uh So there's no AI representation in the regulator.
So we could force this down, uh which we should do sooner rather than later.
(31:09):
Indeed.
That's a very interesting perspective that we can regulate ourselves into relevancy.
I appreciate you taking time to unpack that and walk through that with us because one ofthe things that's very clear from the history of technology and its impact on society is
that when you've got a new general purpose technology and AI, however you want to defineit, is definitely a general purpose technology.
(31:32):
It reshapes everything from the language that we use to the way that economies operate.
And I don't feel that that conversation is being had enough within the business community.
So thank you for taking time to walk through that with me right now.
Let's return to something a little bit more practical.
uh Less next year, next decade.
uh Let's talk about your rise framework.
(31:55):
Redesign, integrate, supervise, evolve.
If I'm a mid-market company, maybe I've done some AI experiments, but I haven't reallytransformed my operations.
How does this apply?
How can I use it?
Where should I start?
Does this work?
Yeah.
Yeah, so em I think the R we've covered in the fact that eventually you have to take thepain of business process re-engineering, whether you like it or not.
(32:24):
The issue is do you do it now or do do it later?
Everything but that is a band-aid.
So you need to reimagine these SOPs and processes that you have.
um
and not just one time, but for all of them.
And what we've launched now is this enterprise digital twin, which is now, imagine walkingthe halls of a hybrid workforce where you can see where you can get to.
(32:49):
And then we map here as is and have a kind of roadmap to get there.
So that's the kind of R.
The I is, uh
to restack your kind of thing is just way too painful.
So we think an orchestration layer can sit on top of your tech stack.
(33:10):
So essentially just integrate it with all the ERP and the various things that youcurrently use.
And uh that's still, by the way, hand-to-hand combat.
So that's where you'll have the most pain.
But I think the idea is you do it once.
uh You do it once and then we call that liberation from the tyranny of point solutions.
(33:32):
eh We're working with a bank here who has 3,400 point solutions and we're just like, canyou imagine upkeeping all the credentials and all databases?
So that's the I phase.
(33:53):
The S is the kind of execution phase, is supervised automation, where now you're justwatching it work and making sure that it's auditable, transparent, it has full human
supervision, so you have that reliability that you can run with it.
The E for us is actually the most important.
um We think enterprises will become kind of self-evolving in the sense of
(34:20):
their gain share, which is called in the old world.
The existing mechanisms around RPA and a few other things are pretty brittle.
I've never seen a workflow that stays the same for a year or two.
Things change all the time.
And then some folks say, we spend more time fixing the RPA than we do doing the work, sowe just do the work now.
(34:44):
So having this kind of self-evolving workflow where
It learns from the edge cases.
It learns from the advancement of them.
And you have this kind of dynamic, living enterprise that you can view and can evolve withyou that's not so brittle and rules-based.
uh And that's how you kind of, we call this potentially the last digital transformationbecause it can evolve itself now.
(35:13):
Once you have all the intentions and all the policies and all the constraints.
uh and humans in the right places, you can self-evolve.
And I think I'm of being ready for that, where you don't have to do this again in fiveyears.
That's fascinating and it begs so many questions I could keep you here for another twohours.
(35:38):
So I'll just pick one question out of the 30 that just popped into my head here.
And that is what you're describing reminds me in many ways of how easy it is now to spinup a website versus 20 years ago, right?
Website in a box and
As you have enterprises in your words that go through the different stages and get to apoint where the software and the process and the automations become self evolving, I
(36:09):
wonder how difficult it is going to be to have, uh let's say, a distribution center in abox, right?
Where you've got
preformatted, tested, battle tested in your language workflows and automations that folkscan either rent or just kind of plop in almost like the preformatted website templates,
(36:30):
right?
That changes the calculus and it removes the advantage that enterprise scale organizationshave against small to upstart, right?
um Yeah.
very interesting.
we call that, there's two parts of the product.
One is the rewire, which is the RISE framework.
We have this other part called wire.
(36:51):
So if you're a startup now, you can like buy back office in a box.
You can buy finance workloads.
You can buy your procurement.
You can buy this, you can buy that.
And actually be AI native from the beginning.
Which I think is quite interesting because you have more degrees of freedom.
Mm-hmm.
in the sense that you can really optimize this and then you can start hiring for the rightroles, right, up front.
(37:13):
And you save on the retraining, you save on a lot of the other kind of legacy issues,which gives new companies a massive advantage that others don't have.
uh So I think it makes the imperative for the rewire just that much more important, whereyou will have structural disadvantage if you don't.
And uh the cost of
(37:36):
iterating goes down dramatically where imagine you can launch 30 companies for the cost ofone in one year and see which one works.
So the cost of testing out in the market, the same way you can have 30 e-commercecompanies for the same price as 15 years ago to build one.
And then you just see which one works.
(37:57):
All the pieces are automated.
Shopify, delivery is all done.
AB testing of company concepts.
Actually even more than AB testing it A through Z, right?
Wow, that is crazy.
uh We say crazy because it's never been possible before, uh but we are seeing all kinds ofthings open up as possibilities now that we previously did not have uh within our
(38:23):
vernacular or even access to.
Yeah, yeah, 100%.
Okay, where are you seeing right now the highest level of demand come from for the type ofwork you talked about the sectors that you work with, but I'm sure you're also fielding
calls from other industries.
um So I'm interested.
Where are you seeing the most interest?
(38:44):
So we started with the eat the frog first approach of what's the hardest thing to do, andthat's the highly regulated industries.
that's kind of our initial focus.
If we get that right, then it's easier to do low cost of our workflows.
So we're doing banking, insurance, health care, pharma, energy, gaming, anything whereit's the work that we depend on.
(39:14):
to go right.
um We are seeing some in-bounds from Series A, a company that just want to do the wireproduct, how do I just build the entire back office of Opus?
uh And for that, we've launched a self-serve version where these will be a full kind ofAPI and dev site where you can kind of use it yourself.
(39:36):
um I think a challenge we have now is
Kind of within enterprise adoption is a top-down is a bottom-up is it side in you knowthat It is playing out the majority of things we're seeing is pretty top-down um But again
(40:00):
one last minor point this this is funny and I shouldn't be saying this is The quality ofthe product is much higher when it's maximally horizontal
Hmm.
uh So the same way in LLM is better at answering a radiology question than a radiologyspecific model because it knows more about uh where it disciplines and outperforms.
(40:37):
So the more we have more industries in our work knowledge graph, the more best practices
Mm-hmm.
processes we learn.
So the quality of the generation is going up dramatically.
However, from a sales perspective, people want to buy extremely vertically.
And not because the quality is better, product, but just because the credibility ishigher.
(41:00):
I want to buy for somebody in my industry, in my region.
So it's an interesting psychology we've seen.
That is.
Well, I also think that we've been conditioned to look for things vertically as well asopposed to horizontally.
yeah, that's true.
You mentioned uh top down versus bottom up.
I'm also, because we haven't really touched on this, but you're working in a lot ofindustries where there's a high level of regulation and the stakes are quite high.
(41:28):
And we haven't really talked other than a little bit on a side trail we took aboutregulators and compliance officers.
So I'm interested, how do you satisfy those two categories of folks who I would imagineare naturally skeptical of AI and what you're bringing to the table?
Yeah, so we actually call it compliance as code, where we think this can actually really,really help those folks.
(41:54):
Because in Opus Now, it actually went into production yesterday.
Every single workflow is rated on how much is compliant with your policies of yourenterprise.
Yeah.
uh So now you can actually have recommendations on to improve it or where you're kind ofleaving policy.
So the visibility of a compliance leader is now much, more.
(42:16):
Plus, when you update a policy, which happens all the time, you know which workflows arekind of, uh you know, inherit that policy.
So in the sense that it can just update all the workflows at once with your approval.
uh Whereas back in the day you had to go and hunt for what work, which processes were inthe business were linked to that policy and kind of catch people who would break the speed
(42:44):
limit.
Whereas now it's like in uh autonomous driving where you you can't exceed that limit.
uh So we see an uptick in compliance with these tools.
ah
But that still takes explanation.
There's still a lot of questions around IP data privacy and these things where, which iswhy we offer this kind of like self-hosted solution just to appease any worries about
(43:17):
things leaving the instance.
Okay, Ari, I want to tap into your perspective.
So shifting gears slightly here a little bit.
You see an awful lot and I love talking to folks that get to see a lot inside the walls ofdifferent organizations.
And I would imagine that you've seen quite a few uh pilot production fails, right?
(43:40):
There's a lot of impressive demos out there that just they don't survive the real world.
So.
What is, or maybe what are some of the biggest mistakes that you have seen companiesmaking when they try to move from a pilot to production with AI?
And of course not with your system, right?
Yeah, yeah, yeah, yeah, I think we've been very lucky with that.
(44:01):
I think we've had over 80 % of our pilots go to production.
I think, so a few things I've realized is, very early on we said this, is why we're theworld's most boring company, because we're focused on like real ROI and real results at
low hanging fruit that's measurable, quantifiable.
So there's a lot of, and by the way, there was disagreement with our tech team, like no,we shouldn't do that, we should just do software, we should be a purist.
(44:29):
And I said no, there needs to be, the application needs to be, have a very clear measureof all return.
And so for example, there's a lot of pilots that do enterprise search.
It's valuable, but it's very hard to quantify the value.
the enterprise machinery of procurement and all these kind of balances, chosen balances ofbudget, have a hard time approving big expenses that aren't linked to direct gains.
(44:58):
uh So that's one, uh a clear kind of ROI.
Then two is just mapping out all the stakeholders where the head of legal will haveconcerns, the head of finance will have concerns, the head of HR will have concerns, the
CEO will have concerns.
So kind of really kind of
addressing all the bad news upfront and say, here's how this is going to go wrong.
(45:24):
And just be very clear about it.
um Lastly, I think we've touched on most of these points already, which is the risk ofsabotage and the antibody reaction, how to address that.
um Having a kind of
(45:45):
sustainable solution with business process re-engineering as opposed to a band-aid.
Getting ownership from the edge of enterprise and not those in the ivory tower only.
yeah, I think these are the current pieces.
We'll obviously make more mistakes in this current round.
(46:07):
And our fourth version is coming out in April, which will incorporate any learnings wehave this year.
Excellent.
What without giving away any secret sauce since you mentioned your fourth version comingout in April 26, what's the biggest change or evolution from version three to version four
ah to the extent that you're able to speak about it?
(46:28):
I think it's going to be quite exciting.
uh
I can't say much about it, I will say, imagine an enterprise digital twin, where theoutput of every workflow is the input of another.
And you can kind of, for the first time, see your enterprise.
(46:50):
Hmm, beautiful.
That sounds pretty cool.
All right.
A couple other questions here as our time over the next 10, 12 minutes here starts to drawto a close.
You're somebody that impresses me uh with the ability to look around corners and we'vealready talked about one or two of the corners that you're looking around.
(47:10):
What are some of the other corners that you're looking around and what are you seeingeither from a practical perspective uh or a longer term almost intellectual academic
perspective?
So something practical and something a little bit more uh high level.
(47:32):
I there's, I don't know how controversial you want to go here, Chad, but let's open it up.
I think...
uh
There's a...
(47:52):
There's going be massive geopolitical changes that I think we're just seeing start.
uh the first country to have, it's almost like the Manhattan Project all over again, butwhoever gets AGI to run a military autonomously, I think will have phenomenal power.
(48:22):
globally.
And that's gonna be a big part of the next couple years.
think the issue that we spoke about, I think it was Elon who called it, we're in thesupersonic tsunami.
(48:43):
ah Where I think there's gonna be serious
unemployment issues to worry about, I think would cascade very quickly into meaning.
A lot of us get our meaning from our work.
um So I think there's big spiritual component to what happens, what we do.
(49:11):
And there are two ways that can go.
we spiral into addiction and entertainment and that world ah or rise into judgment and ahI think building the culture to get to one of those is very, important.
ah And what do we celebrate more um in the culture?
(49:37):
ah I think there's a self-governance question that's going to come to the fore.
Mm-hmm.
the previous models from a first principles perspective will get challenged.
Even like the basis of economics is pricing scarcity.
(50:00):
And in a world in which our productivity goes up, and if production of GDP is just numberof people multiplied by the productivity.
Hmm.
massive increase in production will mean abundance.
And then what is scarce?
How do we do pricing?
What is the need for money?
There's kind of these very, a uh lot of the pillars of what we built our global ahframeworks on have to be questioned.
(50:33):
um
and I've got where I read this but it may have also been Elon but he said that kind of theultimate irony may be that to get to universal high universal basic income you need
maximal capitalism to get there whereas with socialism you'll get to universal low basicincome so the ultimate irony may be that to have you know the
(51:09):
socialist dream materialized the answer was maximal capitalism which is the irony.
God.
No, I think I appreciate you laying out those uh those different ideas because we are at aperiod in time right now where I don't feel enough conversations are happening about where
(51:33):
we're going and what the possibilities are.
There's a lot of people that are very excited about the tech.
There's a lot of folks that are very wrapped up in the changes as.
what we have experienced for the last 40, 50, 70 years in many instances comes apart atthe seams.
There's no doubt in my mind that we're going through a transition period.
This is to use construction terminology and interstitial space.
(51:56):
is the space between.
folks like you are working very diligently to invent the future.
And I'm wondering based on that perspective, do you have any advice for mid-market execs?
that are listening right now about what they can do to help shape and invent the future asopposed to just react to it.
(52:19):
I would say don't wait for the adopt AI immediately.
For example, one of the hospital folks we spoke to said, why don't I just wait a few yearsto see who wins the race and then use that version?
What's the rush?
And I said, but that point is too late.
Make the mistake now.
Even though it may cost you one or two pilots or three things, but build the muscle.
(52:42):
It's an existential imperative.
ah
and then really look to how you can reward your folks working with you to share the upsidein some way, where it is a zero-sum game.
(53:04):
I think that's, as a mid-market manager, I that's what I would focus on the most.
and then see if you can like harness the productivity gains instead of cost reduction intocapacity growth and earn much more.
(53:25):
So there's also an element of, so that's one.
Two.
See you then.
Make sure your kids have the right education and are ready for the world to come.
ah
(53:49):
And that's the whole conversation that we can have afterwards.
But get the next generation ready.
That's good.
And for those interested in what that might look like, uh the episode with Brent Orle fromthe American Enterprise Institute, we talk a little bit about that.
uh He's got some opinions and some thoughts and I think his thinking on that is sound.
So thank you for your perspective, Aria.
(54:12):
Aria, if someone wants to learn more about your organization, about how they can engagewith applied AI, about how they can connect with you, where should they start to look?
ah So opus.com is our website, opus.com, and you can email me directly at a.opus.com andwe're very available.
(54:35):
Very good.
As you mentioned earlier in our conversation, time is a finite resource and it may becomeone of the most valuable commodities out there.
Your time is valuable.
I appreciate you spending some of your time with us today.
I really enjoyed our discussion.
So you bet.
All right.
And for our audience, thank you for once again listening to AI for the C-suite.
(54:57):
If this episode was useful, subscribe wherever you get your podcasts.
Follow us on LinkedIn.
and check out AIforthecsuite.com.
Until next time, keep your algorithms running, your leadership evolving, and your AI incheck.
Take care.
Thank you.
Right, and we're live now, perfect.
I'm Chad Harvey, and this is AI for the C-suite, the show for senior leaders who know thatAI matters and need to figure out what to do about it.
Each episode, we dig into what it all actually means for middle market organizations, howit's changing decisions, strategy, leadership, and the nature of work.
(00:24):
Let's get into it.
I'm very excited about today's conversation because my guest isn't just building AI tools,
He's rethinking how humans and AI should work together, especially in industries wheremistakes aren't an option.
Arya Balorfushan is the founder and CEO of Applied AI, which he deliberately calls theworld's most boring AI company.
(00:48):
That's not false modesty, it's strategy.
Arya's team focuses on automating the critical, regulated workflows in healthcare,banking, and insurance where you can't afford hallucinations or black box decisions.
Arya brings a unique lens to this work.
He started at Goldman Sachs, earned his MBA at Harvard, and took a company public on theOslo Stock Exchange at 29.
(01:10):
Now he's based in Abu Dhabi, building AI that keeps humans in the loop in a veryintentional way.
We're going to dig into a question I think every leader needs to wrestle with.
When AI can do so much, where do humans actually belong?
Arya, welcome to AI for the C-suite.
Thank you very much, Chad.
It's a pleasure and honor to be here.
(01:32):
Very excited to get into our discussion today.
So I'm going to jump right into it with something that you said to me when we were talkingabout this conversation a few weeks ago.
You said to me that most companies are asking the wrong question about AI.
And instead of asking the question, where do I put AI in my workflow?
Your argument is that they should be asking, where do I put humans in an AI workflow?
(01:57):
And that's a pretty significant inversion.
I want to know what you meant by that.
Yeah, so we've been at this for a little over four years now.
So kind of before AI was cool.
And the reason I bring that up is we've been at the front lines deploying these things inmission critical industries and have seen the realities of implementation of production
(02:18):
and also have made.
you know, all the mistakes you can imagine and kind of iterated very quickly on them.
There's, and we've tried to kind of incorporate those into Opus, which is the thirdgeneration of our software.
And uh one of things we realized is anything but business process re-engineering,re-imagination is a band aid.
(02:47):
Mm-hmm.
And the kind of limiting factor is not really technology, but the limiting factor is thehuman mind.
If you uh give somebody a blank canvas with a low code solution, the human mind just kindof recreates the old process again.
(03:07):
And then we'll think about where to add an AI agent to find a local optimum to increaseproductivity by 20%, 30%.
Where we've seen the real uptick in productivity is when you look for a global optimum,where essentially you say uh these legacy processes that were written in the 90s have been
(03:31):
kind of an amalgamation of a bunch of things has a lot of false constraints in there thatwere designed for the limitations of the human.
uh
And once we reimagine it and ask the question, imagine the whole thing is AI.
And where do we want to intentionally put humans that are needed for supervision, forchecks, for edge cases?
(03:56):
Then you see an increase of productivity by 10x, 15x.
um And it's actually a very hard, hard, hard process to do well.
um
And we think, AI can also help us.
We call it workflow generation, uh to kind of remap us onto the map of possibilities forus to begin from a different starting point.
(04:27):
I love that you focused and went right to the psychological component here, the human mindcomponent.
I think you said something to the effect, if we give someone a blank canvas, they're justgoing to reinvent what they already know.
And I think that that's a real challenge.
So when you walk into an organization, how do get people to flip their mental model?
(04:48):
Yeah, so there's a.
to be very clear-eyed about the realities and deliver kind of the bad news.
That's what we start with.
One of which is this won't necessarily succeed in a centralized fashion.
(05:13):
So it's not IT internally that can do this.
It's not management.
A lot of times when we first started, every time there's an AI initiative and a bigannouncement for a company,
The people on the front lines, the process owners, roll their eyes.
Because they know this is just another management initiative who doesn't know what itthey're doing.
And they have to build a workaround on the solutions imposed on them.
(05:37):
That's why we've seen AI adoption falter.
That's why 95 % of these pilots fail.
So we think it's most powerful to do the reimagination at the edge of enterprise.
Mm-hmm.
way, where you have the business process owner involved, the non-technical user be the onewho's building these workflows and iterating on them.
(06:03):
Once they've built consensus at the edge, only then the guardrails of IT for security,cetera, are kind of instilled.
So uh in a way, we see that being most powerful is not to focus on
their existing process and how they do things, focus on what we call the intention or thejob to be done.
(06:28):
What are you really trying to accomplish here?
If it's, I'm trying to make sure there's no fraud in this claim.
Okay, that's the intention.
How you do it is secondary.
How you did it before is secondary.
So that initial prompt that will go into Opus is, what's the job to be done?
And then we append to it all of the constraints, which include the policies that theenterprise has, law, anything else that you put in.
(06:57):
And then our large work model, which has been trained on 1.6 million workflows, willgenerate the first rendition.
It says, you know what?
Here's how it can be done.
And very clearly, it will not be 100 % correct.
But it'll get you 80 % of the way there.
So once you look at it for the first time, you say, oh my god, all of these things that Ithought I needed, I don't need.
(07:23):
And here's where the only places that we recommend humans to be.
And then you do 30 iterations to that as the process owner.
Oh, you forgot about this edge case, you forgot about that, and you add to that.
And after one day, you have a 90 % accurate in-production workflow that you test.
(07:45):
Let's have some files that we run through it and then what goes out.
So having AI generate the first proposal lays bare to you all of the self-limiting uhbeliefs that were in the functional requirements.
(08:07):
1.6 million workflows that really jumped out oh at me in terms of the training base thatyou've got.
What have you found as you've evolved the model?
you finding you?
You gave an example.
You said we run it through.
We get the workflow and 40 iterations later or so and you know one full work day.
We've got something that works.
(08:27):
Have you found that that time is compressing as the models learn more about how to betteriterate uh ideal workflows based on the client?
I'll give you two answers.
Number one is the marginal workflow you built within your enterprise gets exceedinglybetter.
(08:49):
We call that kind of enterprise memory.
Once your policies are uploaded and it knows how you do things, then the first one thatyou generate is not 80 % away there, it's 85 % away there.
Then it's 90 % away there.
The time to value will go down dramatically.
Mm-hmm.
over time.
uh For a new enterprise or if you're on opus.com as kind of a self-serve user, you willbenefit from the kind of increasing number of workflows in our work knowledge graph,
(09:21):
right?
So when the 1.6 because 1.7, 1.8, the kind of base reality and best practices improve anduh as kind of LLMs also improve because we're LLM agnostic, you'll also see benefits
there.
You also said something that jumped out at me and I want to make sure that I understandand maybe I'm even splitting a hair here.
(09:44):
You talked about the intention of the user and the intention of the existing workflowprocess and taking that intention and dropping it into the AI to iterate the new process.
I often think about
StephenCovey101, starting with the end in mind, what's the result?
And it seems like starting with the intention is a flavor of that, distinctly different.
(10:05):
Agree, disagree.
Yeah, so I mean, we follow the Clayton Christensen kind of approach of what's the job tobe done.
What is it that you're trying to really accomplish?
eh And then what's the output that you need to accomplish that?
So it's almost, and there's a big philosophical thing here where human intent and machineexecution in a way, and then human checks.
(10:35):
Mm-hmm.
a way to think about it where there's...
um
You know, like, I'll give you a very clear example to kind of make this all real.
(10:55):
One of the workflows that we began with was independent medical examinations in the US,where it's a mission-critical kind of high-cost-of-error workflow in the sense that
there's a human life on the line.
And it's usually when...
ah your uh insurance company refuses to pay your insurance claim uh and then you sue.
(11:20):
So the law in almost every state is uh they subpoena all your providers.
You can get your entire medical history in this Frankenstein document that's like 30,000pages.
And then an independent medical examiner must adjudicate whether they should pay you ornot.
And it's usually being litigated.
So it's a pretty high cost of error.
(11:43):
If you can't hallucinate, you can't have all these kind of risks there.
And it's around a $5 billion industry.
And it's a private equity and a BPO darling.
uh
because it's a cash cow as you can imagine.
uh the majority of the time the man hours spent in the industry is organizing pages.
(12:07):
Because you can imagine, because you have to organize then summarize.
And no human can summarize before they organize a 30,000 page PDF.
You can't scroll up and down.
You have to put all the, in overall chronological order, in types, in various different,all duplicates must be removed.
(12:28):
So, on the 48 hours that it takes to have this organized and summarized, BPO firms chargeyou 18 to 36 cents a page.
and there's about 90 % accuracy.
We came in and then initially we were doing it the old way without inverting the questionand we got great results.
(12:52):
We were 30 % faster, 30 % cheaper, it was good.
But then we generated the workflow from the intention of I want to know if this insuranceclaim should be paid or not.
That's the intention.
The output is a sorted medical document with a summarization of all of it and medicalcodes, but the whole point is, is this person sick enough that they should stay in
(13:15):
hospital?
Or can they go back to work?
And then what Opus did was when it generated the workflow, it removed the organized phase.
It just summarized it.
The entire organizing of pages was a human limitation of our eyes.
(13:35):
and mental capacity to scroll up and down.
So it went from 48 hours to 4 minutes and it went from 36 cents a page to 1 2 cents a pageand it went from 90 % accuracy to 95 % accuracy.
Just because humans make mistakes you can't do 30,000 pages without making a mistake.
(13:56):
So who pays that cost?
Ultimately all of us do in our insurance premiums.
Mm-hmm.
So there's this global administration attack.
So that's kind of why intention versus just the output is a more effective means togenerate.
That's a really good example and I appreciate you walking us through there.
(14:18):
One of the things that I've been finding in my conversations with folks that aredeveloping, deploying AI technology, as well as working with organizations to help
accelerate their AI transformation is what I would characterize as inherent pushback byhumans against mistakes made by AI and a greater level of grace afforded to humans that
(14:39):
make mistakes, right?
By other humans.
And
Fascinating.
that you just gave, I would imagine that even in an ideal situation with a human reviewerand a human organizer, there will be multiple mistakes or errors of some sort made in that
final product.
Yet we afford that level of grace because it's a human.
(15:00):
And I'm wondering, have you experienced any pushback if an error has been found after thefact.
when it's been organized and summarized by the AI.
Have you found that there's any inherent pushback out there right now about this, or doesthe speed and the decreased cost offset that?
uh
I think it's a very deep, deep point.
(15:25):
So one of the other realities and the bad news that we kind of share when we do thesethings is one of the blockers of adoption is sabotage.
AI is scarier than other humans doing it because your livelihood is on the line.
Like people look at it as a direct threat.
(15:46):
So they look for errors.
as a way to basically do a gotcha moment and have the pilot fail and say, he was better.
uh So we've also seen pretty sophisticated sabotage in the sense of having the pagesmissing in the document or white pages.
(16:07):
So to make the AI make a mistake and then point at it and be like, that didn't work out.
uh Which, by the way, is not.
I understand it, it's not that malicious, it's just human nature to realize what's coming.
So way that we tackle that is, one, we call it supervised automation, where every singledocument has had the human review.
(16:33):
em the AI does the carbs and the human does the protein, where it's almost like you just.
focused on the edge cases, on the parts that require judgment, on the part that the AI haslow confidence in.
uh And this way, the human fingerprints are still on it.
Someone holds liability.
(16:55):
You have a throat to choke in the sense of this person made the mistake.
uh So when the kind of end user sees that, they don't see the AI's making a mistake.
They see this human, this manager hadn't checked.
the work of the AI.
So that kind of empathy is still there.
(17:15):
But also, we are on the side that humans will stay employed.
And the future is this marriage between AI and humans.
And this way, we keep the human relevant.
So it sounds like you've found less, I'm going to call it organizational inertia andresistance with respect to this type of sabotage than individual resistance.
(17:41):
Is that fair or do you find an equal amount of individual and organizational resistanceand inertia?
It goes back to one of the earlier points on if you don't include the business processowners as part of the solution, then they see it as the solution being imposed on them.
And then you see individual sabotage and kind of organizational.
(18:04):
We call it an antibody reaction.
That comes up.
But when you actually give them agency and you give them a sense that they're building it,
ah It's kind of the Ikea effect, ah where now you feel like it's your workflow.
You built it.
It understands you.
ah And then we see a big decline in sabotage.
(18:29):
Got it.
And it's your fault when the table collapses, right?
Okay.
You didn't build it well enough or you forgot to that screw in.
Got it.
So let me ask one other question on what you just outlined before we switch gears.
I was very interested and I think you had in our uh pre call talked about mid mile versuslast mile checking and I think that's what you were just describing right now.
(18:53):
And what I'm interested in is how do you insert the mid mile versus last mile checking ina continuous oversight model without creating new bottlenecks where the humans become
those bottlenecks?
Yeah, so I realize that this is not as intuitive for lot of listeners, so let me just get,we kind of use this.
(19:19):
But the main metric is increasing human productivity, which is output per person per day.
We saw when the human review was only at the end, last mile, like the end of the workflow,AI does everything, and then the human checks, productivity started going down sometimes.
uh because of a compound error problem, where you weren't sure where the probabilisticcode made a small error.
(19:46):
So you have to go back 36 agents and you spend more time fixing than actually just doingit.
So the fundamental issue with agentic workflows, as you know, uh probabilistic code and uhdeterministic outcome requirements.
having the human reviews actually mid-mile, which means like in the middle of theworkflow, this this dance, AI works, human works, AI works, human works, you kind of reset
(20:19):
the probabilistic drift in the middle of these workflows, and you catch the errors muchearlier.
So then your productivity really goes up.
Now, your question is very,
a very correct chat because you now have to kind of leave the workflow and wait for thehuman.
(20:39):
That's also a very expensive thing because that pipeline was spun up in the cloud and youcan't wait four or five hours.
So we actually have to spin that whole pipeline back down, wait for the human input, spinit back up, run the next thing and spin it back down or else the cost will be kind of
astronomical to have all these pipelines kind of waiting.
(21:02):
uh So we actually within Opus have a system where it goes out and every human has theirreviews for the day.
So you're notified, it knows which human is online, which one's offline, uh almost like anUber model.
uh And then they kind of, and you add a deadline, say four hours, three hours, two hoursfor a response.
(21:23):
uh And then we have a dispatch kind of engine, essentially.
that will call a human at right time.
That was very helpful area and I said we're going to shift gears, but it actually unlockedanother question I want to drill into here.
It what you just outlined there reminded me a lot of the conversations I've had with folksregarding software development and how a lot of the coding is now moving into an automated
(21:51):
realm and coders are needing to become more editors uh of the code, right?
Very different skill set and what you just described is uh human checkpoints that are not.
necessarily in production, they are more in an editor role, for lack of better word.
So here's my question.
Are you finding that the production level skills translate fairly well to an editor role,or are there new skills that those folks are needing to learn as they become that mid-mile
(22:20):
checkpoint?
So I will give you a quick answer and one that we are very passionate about.
(22:41):
The humans are moving from a maker to a checker, to use banking terminology.
We're all becoming managers.
We're all supervisors, in a way.
And we're managing AI agents.
The issue is we can't walk the halls.
(23:03):
So in that, it's a challenge of design.
And the onus is on us to kind of showcase the AI recommendation in a way that the humancan easily interact with it.
So the training is more, how do I figure out what I'm looking at?
How does the AI communicate to me what it's sure, what it's not sure of?
(23:25):
And how do I optimize the UI UX in a way where I can respond very easily and quickly?
I don't have to review a 30,000 page file every time there's a checkpoint.
That defeats the purpose.
So there's an element of...
uh of human-computer interaction design, uh which we think is the singular long-term thingto optimize.
(23:51):
That's one piece.
Number two.
So this is, I think, a very existential point on a species level, where for labor to beable to compete in a post-AI world, it needs to reprice itself or re-unitize a unit of
(24:13):
labor from time to output.
Where the old paradigm is, why would I pay you more if it takes you two hours?
I'll pay you more if takes you two minutes.
And that paradigm has there's a lot of retraining.
(24:34):
yeah, and I think by the way, that already happened in the transportation industry, right?
um So I do think there'll be a new breed of lawyer that will charge you by output and giveyou faster results.
So the consumer will win, but
(24:55):
uh In a world of abundance, the thing that's most scarce, the only thing that may bescarce is time.
uh And why are we selling that, essentially?
uh So I think there will be a world in which you have to incentivize the increasedproductivity and not penalize it.
(25:24):
If I'm getting paid for an hour, why would I do 30 times more cases?
I'll do my three or four.
uh So the productivity gains that uh these kind of agentic workflows unlock must be sharedbetween the capital owners and the humans.
ah Or else, I think, and that's a major retraining of not only the human.
(25:50):
in the loop, also the enterprise and the payment apparatus around that.
Only when that's done properly, uh then I think you will see massive increases inproductivity.
(26:12):
So let's stick with this thread for a few moments because I think.
This goes directly to something you and I had previously talked about.
I actually I wrote this down.
So I'm going to quote something you said that directly pertains to your answer here.
You had said to me during our prior call, if we don't share productivity gains with labor,there will be no place for capital owners to hide.
(26:32):
And I couldn't agree more with that, quite frankly.
You even referenced the idea that it took two world wars for us to digest the last majorproductivity gains from mechanization and industrialization.
Yeah.
think that's a pretty stark warning and I'm bringing this up right now because I thinkyou've outlined where the nature of productivity is going.
You talked about output versus hours worked, right?
(26:54):
And I think you can't have this conversation without a larger conversation here about whatthe impact is on capital owners and business leaders.
And I'm interested in exploring this a little bit.
uh I know it's not a direct takeaway maybe for folks in our audience, but this is wherethe puck is moving to quote Gretzky.
I don't think there's any question.
Yeah, we're going through a major shift and we don't entirely know what it's going to looklike on the other side.
(27:19):
So from your point of view, you're on the front lines, you're developing this technology,you're working with a lot of folks, you're seeing the impact it's having now, you're
projecting forward.
What does sharing those productivity gains look like?
And is this something that self-interest is going to drive or is there another motivatingfactor here?
(27:45):
So I think we had a town hall yesterday and I kind of walked through this with the team.
I think if we don't get this right, nothing else kind of matters.
ah Where, you know, the last two world wars and kind of the rise of fascism...
(28:07):
and communism was literally, if you read those books again, the alienation of labor, thesurplus profit, a critique to capital owners, these are all the reasons why we had these
battles these days.
And um we also fought over the source of the gain of productivity, which is energy andoil.
(28:32):
So this gets very real very quick.
Yeah.
Um
But in that case, with hydrocarbons and electricity, we still had minimum wage.
And there were still uprisings.
In this case, there's a good chance to have no wage.
(28:55):
And in particular, in democracies, you're going to have some major disruption to theprocess.
And I do think it may be of interest to your audience too because on the supply side, thecost of delivery of the goods and services will go down with automation.
(29:23):
So we can increase our gross margins, we can do all the good things to kind of increaseefficiencies.
But if no one's on the other side to buy them, so on the demand side, it doesn't matter.
uh
if prices equally drop, automatically.
Or if there's no safety, where do you sell it to?
So I think it's extremely important that we get this right.
(29:46):
um And there are things that kind of block us that we've to go through, like minimum wagelaws.
And there's a lot of these kind of issues where we're mandating to pay people by time.
ah And then a lot of this work goes offshore.
ah
So there is this kind of on-shoring of knowledge work.
(30:08):
There's a billion knowledge workers in the world earning 12 trillion in salary.
The mass majority of is outside the US.
I think if we, to all folks who can, and who hear this, whether you're regulators or theprivate sector, public sector, figure out a way to share these gains of productivity with
(30:43):
humans.
uh And I think that, in fact, we may regulate ourselves into relevancy in the sense thatit's not a free market.
the government is still run by humans.
uh So there's no AI representation in the regulator.
So we could force this down, uh which we should do sooner rather than later.
(31:09):
Indeed.
That's a very interesting perspective that we can regulate ourselves into relevancy.
I appreciate you taking time to unpack that and walk through that with us because one ofthe things that's very clear from the history of technology and its impact on society is
that when you've got a new general purpose technology and AI, however you want to defineit, is definitely a general purpose technology.
(31:32):
It reshapes everything from the language that we use to the way that economies operate.
And I don't feel that that conversation is being had enough within the business community.
So thank you for taking time to walk through that with me right now.
Let's return to something a little bit more practical.
uh Less next year, next decade.
uh Let's talk about your rise framework.
(31:55):
Redesign, integrate, supervise, evolve.
If I'm a mid-market company, maybe I've done some AI experiments, but I haven't reallytransformed my operations.
How does this apply?
How can I use it?
Where should I start?
Does this work?
Yeah.
Yeah, so em I think the R we've covered in the fact that eventually you have to take thepain of business process re-engineering, whether you like it or not.
(32:24):
The issue is do you do it now or do do it later?
Everything but that is a band-aid.
So you need to reimagine these SOPs and processes that you have.
um
and not just one time, but for all of them.
And what we've launched now is this enterprise digital twin, which is now, imagine walkingthe halls of a hybrid workforce where you can see where you can get to.
(32:49):
And then we map here as is and have a kind of roadmap to get there.
So that's the kind of R.
The I is, uh
to restack your kind of thing is just way too painful.
So we think an orchestration layer can sit on top of your tech stack.
(33:10):
So essentially just integrate it with all the ERP and the various things that youcurrently use.
And uh that's still, by the way, hand-to-hand combat.
So that's where you'll have the most pain.
But I think the idea is you do it once.
uh You do it once and then we call that liberation from the tyranny of point solutions.
(33:32):
eh We're working with a bank here who has 3,400 point solutions and we're just like, canyou imagine upkeeping all the credentials and all databases?
So that's the I phase.
(33:53):
The S is the kind of execution phase, is supervised automation, where now you're justwatching it work and making sure that it's auditable, transparent, it has full human
supervision, so you have that reliability that you can run with it.
The E for us is actually the most important.
um We think enterprises will become kind of self-evolving in the sense of
(34:20):
their gain share, which is called in the old world.
The existing mechanisms around RPA and a few other things are pretty brittle.
I've never seen a workflow that stays the same for a year or two.
Things change all the time.
And then some folks say, we spend more time fixing the RPA than we do doing the work, sowe just do the work now.
(34:44):
So having this kind of self-evolving workflow where
It learns from the edge cases.
It learns from the advancement of them.
And you have this kind of dynamic, living enterprise that you can view and can evolve withyou that's not so brittle and rules-based.
uh And that's how you kind of, we call this potentially the last digital transformationbecause it can evolve itself now.
(35:13):
Once you have all the intentions and all the policies and all the constraints.
uh and humans in the right places, you can self-evolve.
And I think I'm of being ready for that, where you don't have to do this again in fiveyears.
That's fascinating and it begs so many questions I could keep you here for another twohours.
(35:38):
So I'll just pick one question out of the 30 that just popped into my head here.
And that is what you're describing reminds me in many ways of how easy it is now to spinup a website versus 20 years ago, right?
Website in a box and
As you have enterprises in your words that go through the different stages and get to apoint where the software and the process and the automations become self evolving, I
(36:09):
wonder how difficult it is going to be to have, uh let's say, a distribution center in abox, right?
Where you've got
preformatted, tested, battle tested in your language workflows and automations that folkscan either rent or just kind of plop in almost like the preformatted website templates,
(36:30):
right?
That changes the calculus and it removes the advantage that enterprise scale organizationshave against small to upstart, right?
um Yeah.
very interesting.
we call that, there's two parts of the product.
One is the rewire, which is the RISE framework.
We have this other part called wire.
(36:51):
So if you're a startup now, you can like buy back office in a box.
You can buy finance workloads.
You can buy your procurement.
You can buy this, you can buy that.
And actually be AI native from the beginning.
Which I think is quite interesting because you have more degrees of freedom.
Mm-hmm.
in the sense that you can really optimize this and then you can start hiring for the rightroles, right, up front.
(37:13):
And you save on the retraining, you save on a lot of the other kind of legacy issues,which gives new companies a massive advantage that others don't have.
uh So I think it makes the imperative for the rewire just that much more important, whereyou will have structural disadvantage if you don't.
And uh the cost of
(37:36):
iterating goes down dramatically where imagine you can launch 30 companies for the cost ofone in one year and see which one works.
So the cost of testing out in the market, the same way you can have 30 e-commercecompanies for the same price as 15 years ago to build one.
And then you just see which one works.
(37:57):
All the pieces are automated.
Shopify, delivery is all done.
AB testing of company concepts.
Actually even more than AB testing it A through Z, right?
Wow, that is crazy.
uh We say crazy because it's never been possible before, uh but we are seeing all kinds ofthings open up as possibilities now that we previously did not have uh within our
(38:23):
vernacular or even access to.
Yeah, yeah, 100%.
Okay, where are you seeing right now the highest level of demand come from for the type ofwork you talked about the sectors that you work with, but I'm sure you're also fielding
calls from other industries.
um So I'm interested.
Where are you seeing the most interest?
(38:44):
So we started with the eat the frog first approach of what's the hardest thing to do, andthat's the highly regulated industries.
that's kind of our initial focus.
If we get that right, then it's easier to do low cost of our workflows.
So we're doing banking, insurance, health care, pharma, energy, gaming, anything whereit's the work that we depend on.
(39:14):
to go right.
um We are seeing some in-bounds from Series A, a company that just want to do the wireproduct, how do I just build the entire back office of Opus?
uh And for that, we've launched a self-serve version where these will be a full kind ofAPI and dev site where you can kind of use it yourself.
(39:36):
um I think a challenge we have now is
Kind of within enterprise adoption is a top-down is a bottom-up is it side in you knowthat It is playing out the majority of things we're seeing is pretty top-down um But again
(40:00):
one last minor point this this is funny and I shouldn't be saying this is The quality ofthe product is much higher when it's maximally horizontal
Hmm.
uh So the same way in LLM is better at answering a radiology question than a radiologyspecific model because it knows more about uh where it disciplines and outperforms.
(40:37):
So the more we have more industries in our work knowledge graph, the more best practices
Mm-hmm.
processes we learn.
So the quality of the generation is going up dramatically.
However, from a sales perspective, people want to buy extremely vertically.
And not because the quality is better, product, but just because the credibility ishigher.
(41:00):
I want to buy for somebody in my industry, in my region.
So it's an interesting psychology we've seen.
That is.
Well, I also think that we've been conditioned to look for things vertically as well asopposed to horizontally.
yeah, that's true.
You mentioned uh top down versus bottom up.
I'm also, because we haven't really touched on this, but you're working in a lot ofindustries where there's a high level of regulation and the stakes are quite high.
(41:28):
And we haven't really talked other than a little bit on a side trail we took aboutregulators and compliance officers.
So I'm interested, how do you satisfy those two categories of folks who I would imagineare naturally skeptical of AI and what you're bringing to the table?
Yeah, so we actually call it compliance as code, where we think this can actually really,really help those folks.
(41:54):
Because in Opus Now, it actually went into production yesterday.
Every single workflow is rated on how much is compliant with your policies of yourenterprise.
Yeah.
uh So now you can actually have recommendations on to improve it or where you're kind ofleaving policy.
So the visibility of a compliance leader is now much, more.
(42:16):
Plus, when you update a policy, which happens all the time, you know which workflows arekind of, uh you know, inherit that policy.
So in the sense that it can just update all the workflows at once with your approval.
uh Whereas back in the day you had to go and hunt for what work, which processes were inthe business were linked to that policy and kind of catch people who would break the speed
(42:44):
limit.
Whereas now it's like in uh autonomous driving where you you can't exceed that limit.
uh So we see an uptick in compliance with these tools.
ah
But that still takes explanation.
There's still a lot of questions around IP data privacy and these things where, which iswhy we offer this kind of like self-hosted solution just to appease any worries about
(43:17):
things leaving the instance.
Okay, Ari, I want to tap into your perspective.
So shifting gears slightly here a little bit.
You see an awful lot and I love talking to folks that get to see a lot inside the walls ofdifferent organizations.
And I would imagine that you've seen quite a few uh pilot production fails, right?
(43:40):
There's a lot of impressive demos out there that just they don't survive the real world.
So.
What is, or maybe what are some of the biggest mistakes that you have seen companiesmaking when they try to move from a pilot to production with AI?
And of course not with your system, right?
Yeah, yeah, yeah, yeah, I think we've been very lucky with that.
(44:01):
I think we've had over 80 % of our pilots go to production.
I think, so a few things I've realized is, very early on we said this, is why we're theworld's most boring company, because we're focused on like real ROI and real results at
low hanging fruit that's measurable, quantifiable.
So there's a lot of, and by the way, there was disagreement with our tech team, like no,we shouldn't do that, we should just do software, we should be a purist.
(44:29):
And I said no, there needs to be, the application needs to be, have a very clear measureof all return.
And so for example, there's a lot of pilots that do enterprise search.
It's valuable, but it's very hard to quantify the value.
the enterprise machinery of procurement and all these kind of balances, chosen balances ofbudget, have a hard time approving big expenses that aren't linked to direct gains.
(44:58):
uh So that's one, uh a clear kind of ROI.
Then two is just mapping out all the stakeholders where the head of legal will haveconcerns, the head of finance will have concerns, the head of HR will have concerns, the
CEO will have concerns.
So kind of really kind of
addressing all the bad news upfront and say, here's how this is going to go wrong.
(45:24):
And just be very clear about it.
um Lastly, I think we've touched on most of these points already, which is the risk ofsabotage and the antibody reaction, how to address that.
um Having a kind of
(45:45):
sustainable solution with business process re-engineering as opposed to a band-aid.
Getting ownership from the edge of enterprise and not those in the ivory tower only.
yeah, I think these are the current pieces.
We'll obviously make more mistakes in this current round.
(46:07):
And our fourth version is coming out in April, which will incorporate any learnings wehave this year.
Excellent.
What without giving away any secret sauce since you mentioned your fourth version comingout in April 26, what's the biggest change or evolution from version three to version four
ah to the extent that you're able to speak about it?
(46:28):
I think it's going to be quite exciting.
uh
I can't say much about it, I will say, imagine an enterprise digital twin, where theoutput of every workflow is the input of another.
And you can kind of, for the first time, see your enterprise.
(46:50):
Hmm, beautiful.
That sounds pretty cool.
All right.
A couple other questions here as our time over the next 10, 12 minutes here starts to drawto a close.
You're somebody that impresses me uh with the ability to look around corners and we'vealready talked about one or two of the corners that you're looking around.
(47:10):
What are some of the other corners that you're looking around and what are you seeingeither from a practical perspective uh or a longer term almost intellectual academic
perspective?
So something practical and something a little bit more uh high level.
(47:32):
I there's, I don't know how controversial you want to go here, Chad, but let's open it up.
I think...
uh
There's a...
(47:52):
There's going be massive geopolitical changes that I think we're just seeing start.
uh the first country to have, it's almost like the Manhattan Project all over again, butwhoever gets AGI to run a military autonomously, I think will have phenomenal power.
(48:22):
globally.
And that's gonna be a big part of the next couple years.
think the issue that we spoke about, I think it was Elon who called it, we're in thesupersonic tsunami.
(48:43):
ah Where I think there's gonna be serious
unemployment issues to worry about, I think would cascade very quickly into meaning.
A lot of us get our meaning from our work.
um So I think there's big spiritual component to what happens, what we do.
(49:11):
And there are two ways that can go.
we spiral into addiction and entertainment and that world ah or rise into judgment and ahI think building the culture to get to one of those is very, important.
ah And what do we celebrate more um in the culture?
(49:37):
ah I think there's a self-governance question that's going to come to the fore.
Mm-hmm.
the previous models from a first principles perspective will get challenged.
Even like the basis of economics is pricing scarcity.
(50:00):
And in a world in which our productivity goes up, and if production of GDP is just numberof people multiplied by the productivity.
Hmm.
massive increase in production will mean abundance.
And then what is scarce?
How do we do pricing?
What is the need for money?
There's kind of these very, a uh lot of the pillars of what we built our global ahframeworks on have to be questioned.
(50:33):
um
and I've got where I read this but it may have also been Elon but he said that kind of theultimate irony may be that to get to universal high universal basic income you need
maximal capitalism to get there whereas with socialism you'll get to universal low basicincome so the ultimate irony may be that to have you know the
(51:09):
socialist dream materialized the answer was maximal capitalism which is the irony.
God.
No, I think I appreciate you laying out those uh those different ideas because we are at aperiod in time right now where I don't feel enough conversations are happening about where
(51:33):
we're going and what the possibilities are.
There's a lot of people that are very excited about the tech.
There's a lot of folks that are very wrapped up in the changes as.
what we have experienced for the last 40, 50, 70 years in many instances comes apart atthe seams.
There's no doubt in my mind that we're going through a transition period.
This is to use construction terminology and interstitial space.
(51:56):
is the space between.
folks like you are working very diligently to invent the future.
And I'm wondering based on that perspective, do you have any advice for mid-market execs?
that are listening right now about what they can do to help shape and invent the future asopposed to just react to it.
(52:19):
I would say don't wait for the adopt AI immediately.
For example, one of the hospital folks we spoke to said, why don't I just wait a few yearsto see who wins the race and then use that version?
What's the rush?
And I said, but that point is too late.
Make the mistake now.
Even though it may cost you one or two pilots or three things, but build the muscle.
(52:42):
It's an existential imperative.
ah
and then really look to how you can reward your folks working with you to share the upsidein some way, where it is a zero-sum game.
(53:04):
I think that's, as a mid-market manager, I that's what I would focus on the most.
and then see if you can like harness the productivity gains instead of cost reduction intocapacity growth and earn much more.
(53:25):
So there's also an element of, so that's one.
Two.
See you then.
Make sure your kids have the right education and are ready for the world to come.
ah
(53:49):
And that's the whole conversation that we can have afterwards.
But get the next generation ready.
That's good.
And for those interested in what that might look like, uh the episode with Brent Orle fromthe American Enterprise Institute, we talk a little bit about that.
uh He's got some opinions and some thoughts and I think his thinking on that is sound.
So thank you for your perspective, Aria.
(54:12):
Aria, if someone wants to learn more about your organization, about how they can engagewith applied AI, about how they can connect with you, where should they start to look?
ah So opus.com is our website, opus.com, and you can email me directly at a.opus.com andwe're very available.
(54:35):
Very good.
As you mentioned earlier in our conversation, time is a finite resource and it may becomeone of the most valuable commodities out there.
Your time is valuable.
I appreciate you spending some of your time with us today.
I really enjoyed our discussion.
So you bet.
All right.
And for our audience, thank you for once again listening to AI for the C-suite.
(54:57):
If this episode was useful, subscribe wherever you get your podcasts.
Follow us on LinkedIn.
and check out AIforthecsuite.com.
Until next time, keep your algorithms running, your leadership evolving, and your AI incheck.
Take care.
Thank you.
Popular Podcasts
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.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.