March 10, 2025 • 39 mins
Get featured on the show by leaving us a Voice Mail: https://bit.ly/MIPVM
FULL SHOW NOTES
https://www.microsoftinnovationpodcast.com/663
Step into the world of AI with our latest episode, where Dewain Robinson from Microsoft unpacks the transformative power of Copilot Studio and intelligent agents in 2025. As organizations gear up to integrate advanced AI solutions, the discussion delves into what this shift means for everyday tasks and overall productivity. From understanding the distinction between copilots and agents to exploring real-life case studies demonstrating innovation in action and unexpected value, listeners will gain actionable insights into the future of AI.
This vibrant discussion is not just about technology; it’s about shaping the future of work—witnessing AI evolve from a mere tool to an essential colleague. Grab your headphones and join us to prepare for the AI revolutions that lie ahead. Don’t forget to subscribe, share your thoughts, and leave a review if you enjoyed the episode!
TAKEAWAYS
• Understanding Copilot Studio and its capabilities
• The rise of autonomous agents in 2025
• User-friendly AI systems boosting creativity over simplicity
• Case studies demonstrating the transformational value of AI
• Preparing for the future: AI as an enabler, not a replacement
This year we're adding a new show to our line up - The AI Advantage. We'll discuss the skills you need to thrive in an AI-enabled world.
Accelerate your Microsoft career with the 90 Day Mentoring Challenge
We’ve helped 1,300+ people across 70+ countries establish successful careers in the Microsoft Power Platform and Dynamics 365 ecosystem.
Benefit from expert guidance, a supportive community, and a clear career roadmap. A lot can change in 90 days, get started today!
If you want to get in touch with me, you can message me here on Linkedin.
Thanks for listening 🚀 - Mark Smith
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Mark Smith (00:01):
Welcome to the Copilot Show, where I interview
Microsoft staff innovating withAI.
I hope you will find thispodcast educational and inspire
you to do more with this greattechnology.
Now let's get on with the show.
Today's guest is from NationalTennessee in the US.
He works at Microsoft as aPrincipal Program Manager for
(00:23):
Copilot Studio andConversational AI.
He helps people leverageCopilot Studio to build amazing
copilots.
You can find links to his bioand socials in the show note for
this episode.
He was last on the show forepisode 497, and we talked about
Power Virtual Agents, so thatwas a wee while ago, but welcome
(00:43):
back, Dewain.
Dewain Robinson (00:45):
Thanks, I'm glad to be here, mark.
It's awesome to get to talk toyou again.
We just were kind of talkingbefore that.
It's been so long and Icouldn't even believe it was.
Yeah, it's hard for me tobelieve it's even been that long
.
It just seems like yesterdayyou and I were chatting about
this.
Mark Smith (01:00):
Things are changing so, so rapidly.
If you look at you know we'rerecording this at the start of
2025.
I feel we've come off an epicyear where corporations are
really starting to go.
What does AI look for us?
And then the last three monthsI feel we were heavily there was
heavy marketing, not just fromMicrosoft, but all the major
(01:23):
players in the world in thegenerative AI space LLMs, et
cetera.
Around that, 2025 is going tobe the year of agents.
So, with that backdrop beforeus, when you look at this year
ahead, what excites you rightnow the most?
Dewain Robinson (01:40):
I think the thing that is going to be very
interesting in the CopilotStudio world, right, is, I would
say, that there's thisperception of AI and what it can
do, and then there's thereality of what it could do and
how easy it is to do that withthat AI back last year.
(02:02):
Right, you could really buildsome really cool stuff With that
AI back last year.
Right, you could really buildsome really cool stuff.
And I think what's happening isthe commoditization, the
roughing of the edges that werea little bit in there.
And so I think that in itselfor what I would call like a
synchronous or a, you know, kindof like a synchronous co-pilot,
where you talk to it and ittalks back to you is going to be
(02:24):
one of the things that is goingto be one of the things that is
going to be the majorinnovation.
It's not like oh, we're goingto see something that we never
thought of.
It's just going to be thethings that you are wanting to
do will become very easy andcommoditized, and the
integrations to be able to tiethat in with Azure and other
pieces are just going to becomeeasier and easier, and you could
(02:45):
do it before, but now it'sgoing to be like oh, there's a
button and it's an easy buttonto do it.
And then I think the other thingI would say, mark, is that as
we walk into the autonomousagent space, that's going to
really change the perception.
It's beyond the concepts ofwhat you can do with AI by
(03:08):
itself.
It's now AI working inconjunction with business
process automation to be able todo something you could never do
before.
And I think the combination ofthese two things are really
going to be a year.
That's just insane.
And, as you're aware, I meanthe innovation on the AI side
and the AI model side is justgoing to keep going too.
(03:29):
So, when we talk again and youknow, because we're not going to
wait so long, right, mark sothen when we talk again next
time, I would anticipate there'sprobably going to be some
innovation in the model spacethat you can do and stuff like
that, because we already startedseeing some of that.
Like you can now stream audioin and you're not having to do
text to speech.
Yeah, like that's a big deal,right.
(03:51):
And so I think there's going tobe a lot of interesting things
when we start going beyond justchatware.
Yeah, I think chat stuff isgoing to get super easy to do,
but then these other things arejust going to really take it to
a level that people reallydidn't think about, and I think
that's the beauty of it.
And this is the year that Ithink it's not just about open
AI.
It's about open AI's integrate,how Microsoft's going to take
(04:14):
open AI and combine it with theMicrosoft asset pool, and then
you're going to get somethingthat's completely different,
that you couldn't do before,which I think is really awesome.
Mark Smith (04:24):
Have you seen, I suppose in 2024, looking at the
year in reverse, withoutrevealing any customer details
or anything did you see anythingthat was quite you're like?
You know that was cool, that'sa cool use case, or that was a
cool way to use a technology?
Because you mentioned voicethere and I think voice is a
massively underexploredterritory in how we go forward.
(04:46):
I just had a customer the otherday say listen, when I fill out
my CRM system, I really justwant to be able to open the
record and say the first name isBob, the second name is Frank
and it just types it in justwith me talking and you know,
auto filling out, that type offunctionality which we get that
degree coming in in the Dynamicssuite of products.
And but it was interesting.
(05:07):
This is an old guy, but his wayof how you know he would like
to use AI combined with what hewas doing and but was there
anything that really kind ofstood out top of mind story that
was highlights for you lastyear?
Dewain Robinson (05:23):
I had a couple of things, that kind of top of
mind, and I will say there wasthis thing that when I think
back where I got value that Ididn't expect.
So it's not necessarily likethe use case in this scenario.
Well, I'll say this use casewas fairly interesting because
and everyone's heard about theuse case, because it's the
(05:45):
Holland America one that we talkabout publicly, right but the
thing that was about that Ididn't expect.
You know, we had done businessvalue assessments and all this
stuff about what we expected,the return and how this thing
was going to perform andeverything.
And we were expecting to getlike a return from like over in
(06:07):
this area, you know on the leftand what also, but we didn't
expect this value area on theright.
We didn't even know that therewould be impact there.
And and what it came down to isthat if you think about it,
think about I have an API or Ihave something that you want to
do, and in the case of this it'slike cruises I want to find a
(06:27):
cruise.
So when we were doing this, thecustomer had an API that had
here are all the differentcruises that you might want, and
so we needed to go search for acruise that would be the right
cruise for you cruise that wouldbe the right cruise for you.
(06:49):
And what we found and thisbecame similar because it
wouldn't matter that it wasfinding cruises it could be like
you wanted to searchServiceNow's knowledge base
right For knowledge articles.
What we didn't expect at thetime and it makes logical sense
now that this would be the caseis what we were doing is using
an AI model to look at thecontext of the whole
(07:11):
conversation and generate a lereally improved search results,
coming back from an existingsearch service that then could
(07:31):
be consumed and consumed by theAI with the context as well, but
knowing that the search qualitywas going to go up and the
search results were going to bebetter off of your API.
That was just a standardlexical search.
I didn't expect that.
I didn't think that that wouldbe something that was the case.
But we saw that not only withthe Holland America example, but
(07:55):
we also were working withanother company and they were
trying to do something withServiceNow knowledge articles
and we saw the same thing and wewent oh man, we didn't think
about that.
And so there's actually someyou know features and stuff that
we started putting into CopilotStudio because we realized
after we saw that that that'sgoing to mean that a lot of
(08:15):
people are going to want to beable to do that and we want to
make that easier to do and allof that.
And so I think that was kind oflike a really aha moment for me
when I was looking at that.
And then the second thing thatwas like really cool is I did a
video that showed the neworchestration engine.
Mark Smith (08:36):
I was just going to ask you about it.
Dewain Robinson (08:37):
Yeah, yeah, and I did one where I do one with
the PO, like I'm creating a PO,where I do one with the PO, like
I'm creating a PO.
And what a lot of people don'tknow is that I'm building this
thing.
And I get on the phone and Icall up Gary Pretty, which is
the lead over it, and I said,hey, I need you to tell me what
are all the things that I needto show to demo this thing in
(09:00):
the best light.
And so I built this thing withno intention, right?
I was just wanting to checkthis box, check this box, check
this box.
And I went through it allmarket and I built all this
stuff with all the differentpieces to do it.
And then, out of a fluke, I justwent, wait a minute, if it has
all that, if I say, create metwo POs with this one with this
(09:21):
and this one with that andeverything.
And it just did it.
It literally figured out allthe assets and did it and
dynamically assembled it and didit not just once, but did it
for two different things,summarize it all out and spit it
out.
And that demo and that videothat you saw from that stems
from me building that as a wayto demonstrate it on a stage, to
(09:43):
show people this new feature.
But then I'm trying to dosomething real world and then it
occurs to me that wait a minute, this thing can do more than
what I really thought about whenI was building it, which
showcases the fact that if youdo these things right and you
modularize it right and yourmodel descriptions are good,
it'll actually do stuff that youdidn't even intend it to do,
(10:06):
because it's flexible enough tofigure it out.
And that's kind of a really bigaha moment.
We even had one of our guysthat was demoing on a stage and
he was showing something and wehad negative feedback on him
because he said on the stage hewent oh wow, get a feedback on
him.
Because he said on the stage hewent oh wow, that worked like I
(10:27):
didn't expect, and so, likethose things were really
impressive to me as I was goingthrough.
So those are the standout onesfor me right now.
Mark Smith (10:34):
I like it, I like it .
And now, in 2025, and we seethis concept and what I'd like
you to do actually start withyour description of what agents
are what is semi-autonomous,what's autonomous and, let's say
, what's fully autonomous orreally you know, set and forget
type agent scenarios.
I think orchestration playsinto that mix, because you're
(10:56):
going to have these agents thathave deeper expertise, but
you're going to need anorchestrator to know which agent
to call or which one to fire ina scenario.
Let's start with how do youexplain the concept of agents
that we're hearing about now?
Dewain Robinson (11:13):
Yeah, so there was a little bit of a rebrand.
That happened, and Microsoftwere great at marketing and
rebranding things without beingclear that we did it.
So the word co-pilot let'sstart there.
Co-pilot was being used likeeverywhere.
Yes, right.
(11:33):
And so we started to get aproblem, which was nobody
understood what the wordco-pilot meant, and then we
started using co-pilot and theword agent, and then we started
getting into apilot and the wordagent, and then we started
getting into a new one which wasgoing to be a function, an
extension and a plugin.
And all of these words startedbecoming a problem because
(11:55):
people were using theminterchangeably and I literally
had two sales guys fighting infront of a customer on what
co-pilot did, and they didn'teven realize they were both
right, they were talking aboutdifferent products.
And so the thing that if westart out what is a Copilot now
by branded by Microsoft, you'regoing to see that it's going to
have like a big C in the front,and the reason behind that is
(12:17):
Copilot is a product fromMicrosoft, it is a product that
you configure, and sure youcould say it's an agent that's
prepackaged and available to you, right?
And so that's why you're goingto see this word co-pilot.
Now, an agent is going to bedifferent and you will see an
agent with a small A, not with abig A, and the reason behind
(12:42):
that and this is, by the way,I've had to have a long talk
with marketing to get this, sojust trust me, it's not like
this is common nature.
The word agent is meant to sayit is a conversational AI
implementation that usesgenerative AI that you're in
control of that you build andthat you are in control of the
(13:03):
orchestration.
When you go over into, likeCopilot for M365, you don't have
full control of theorchestration engine.
Now you can throw something inand you can extend it and you
can do all that, but you can'tmake it not do something that
it's programmed to do, whereasan agent means you are in
control of it, and so if youthink the word co-pilot agent
(13:26):
big C, little A means that'swhat an extension is of a
co-pilot product is a co-pilotagent.
And so if you're saying agent,then we're just talking about
something you built.
Now how you build that thing,it can be whatever.
And so, yeah, we always thoughtit was funny and you'll get a
kick out of this Mark is that wechanged the name from Power
(13:48):
Virtual Agents to Copilot Studio.
But now, copilot Studio doesn'tbuild copilots, it builds
agents.
Mark Smith (13:54):
Yes.
Dewain Robinson (13:55):
So I had fun with marketing on that one, yes,
but the thing is, is so inCopilot Studio, what are you
going to do?
You're going to build agents.
Right Now, can you build agentsin different ways?
Absolutely, you can, butfundamentally understand that an
agent that is a realintelligent agent, that is
(14:16):
something bigger than just amodel, is going to need to have
an orchestration engine.
And that is what Copilot Studiois.
It's an out of the box, it is aSaaS-based conversational
orchestration engine.
Right, and we saw the evolutionof that to incorporate using not
a traditional language model,and for those that don't know
(14:36):
what I mean by that, it used tobe that you would have to train
an NLU model.
You'd say here's six differentways of saying this thing, and
that way I can say this is theintent.
And then you could do the samething with entities, which, if
you want to understand thedifference between intents and
entities, an intent is if I saidI want to order a large
pepperoni pizza.
The intent is that I want toorder something, then the
(15:00):
entities are large pepperonipizza.
The entities are largepepperoni pizza, and so, being
able to do that, you had totrain that into a natural
language understandingimplementation and traditionally
, this is where large languagemodels came along, because you
don't train one Not typicallyyou don't train one and, matter
of fact, to train a largelanguage model from the ground
(15:22):
up takes like six months to dowith a massive data center.
It's worth the GPUs to be ableto get it done.
So, when we're looking at these, you don't do that anymore.
How you use these is you hadthe innovation that happened
where I don't need to do intentrouting or determining what tool
to use.
I can use a orchestrationengine that's not powered by a
(15:45):
traditional NLU.
I can do it with a largelanguage model, and this is what
you heard us talk aboutgenerative actions and
everything, and you're going tosee like you'll hear us refer to
this as generativeorchestration now, because we
realize like actions, which isthe wrapper around an API, we
needed to separate that from theorchestration engine capability
, and I'll explain why in just asecond.
(16:06):
So, once we've done this, thisis where it really gets into
what you're talking about.
What is an agent?
Okay, so now we can start toclassify them.
Well, and there's differentagent types.
Now you have what I refer to asa synchronous agent.
We just call it an agent, whichis what you traditionally think
(16:27):
of as like a chat, gpt orbasedorchestrator that could
understand the context of theconversation, and or the Power
Automate flow, to go create a PO, and here's the one to go get
(17:04):
an approval, and oh, here's oneto go get information about our
HR policy and all that.
So what you're seeing is it'sbasically acting as like a
router, a conversational router,and that, by the way, is what
you know.
Conversational orchestration isyes.
So that's when we started to seethat you're going to start to
see these other agent types, andI think this leads into a thing
(17:29):
that I know many people getconfused about, which is what is
the difference between whatyou're hearing us talk about
autonomous agents.
And why is that?
Because people think about itand they go oh well, smart,
business process automation orsomething like that and they go
well, wait a minute, how is thatdifferent than Power Automate?
And because now we're startingto see these classifications
(17:51):
where it's not you're talking toit and it's talking back.
Now it can be that somethingelse triggers it and then it
does something, and a lot ofpeople go well, that's Power
Automate.
So Power Automate can havesomething that triggers it and
then it does something and a lotof people go well, that's power
automate.
So power automate can havesomething that triggers it and
then you can follow a businessprocess and everything.
So here's the difference, mark,and this is why it was so
(18:12):
important to land aconversational orchestrator
based off a large language model.
Because now imagine back up toCopilot Studio for a second.
Imagine if I was going to hiresomeone.
I'm going to write a jobdescription.
I take that and I say here'swhat your job is, here are the
rules you should follow, hereare the tools at your disposal
(18:33):
to be able to do your job.
And this is what your job is.
And when should you do your job?
Well, when you receive an email, when someone puts a record in
a CRM, when someone fills out aform like whatever, right.
So there's this thing.
That's not that someone'stalking to you, but guess what?
Those autonomous agents cantalk to you, but they don't have
(18:55):
to, because they can go andlook at it and imagine I have
six different power automateflows.
I explain what they all do.
I come over here and I say, oh,here's some knowledge with like
information on products andinformation about business
processes and how those thingsshould, what those policies are
and things of that nature.
(19:16):
And then I have, like, allthese different tools, and even
a topic could be a tool, yeah,so now I'm triggering those
things and it can do whatever Ineeded to go do based upon
something going on.
So one of the demos that we didI mean I didn't show this one
publicly.
I can't remember if I made thevideo for this one or not.
(19:36):
I have to go back and look whatI did.
I think this is the video I'mworking on right now.
But the problem, marcus, it'sgetting harder to build the
crazy complexity that's going tomake everybody go wow, because
I know I can do it, but I haveto go build it all.
But what I did at thatparticular thing at the CIO
Summit is we took and made aform that you could go ask it
(19:57):
about what are cruise specials?
What are cruise specials?
And it would go look you upbased upon your email address
and figure out what you hadcruised in the past, make a
product recommendation, look atwhat the specials are, figure
out what you like, go in andcreate a CRM and implement you
know record for it, update theCRM record, then make you an
email and email it to the personand giving them the cruise
(20:19):
specials.
And then someone could emailback to it and be like, hey, you
gave me these three options,but I was thinking about I
wanted to cruise Greece thisyear Do you have any specials
for that?
And it could then converse withthem and then, and all the time
when it's doing this, it'supdating the CRM system for
people to know what was going onand it's helping with the
pipeline.
So the thing is, is that that'swhere you get into autonomous
(20:43):
agents?
I think everyone gets confusedbecause they make it too simple.
They think it's like justbusiness process automation or
that there's a step within thebusiness process automation
where I want to look at themodel and figure out what to do
next.
But it is literally like thebrain of the orchestration
engine being able to betriggered and then it understand
what its job is and the toolsit has, so that it can just
(21:04):
figure the brain of theorchestration engine being able
to be triggered and then itunderstand what its job is and
tools it has, so that it canjust figure out what's the right
thing to go do to be able tocomplete it.
And right now, I would say we'rein the infancies of what it can
do, so you might hit somelimitations and stuff, because
it was only in preview, but theintent is that it works like
that, and a lot of people thenfreak out when I talk about this
(21:24):
, mark, and they go oh my God,it's going to replace me for my
job and it's like no, it's goingto take the thing that you have
to do over and over and overagain and make you not have to
do that so you can work on thethings that you need to provide
the higher quality, and so Ireally think this is the thing
that's really going to beamazing in 2025.
Mark Smith (21:45):
If we look at that example that you've just given
and we say it's a jobdescription, right, if you look
at a job description for a role,it'll fit on one side of a
sheet of paper maybe.
Then you start doing the joband you realize it's really 300
sheets of paper because we don'ttell you that you're going to
have to be in these meetings andyou're going to have to, you
(22:07):
know, reply to an aggregatedemail.
We don't put all those nuancesof the job in the job
description.
But you start to pick it up andwhen I think of agents and the
element that I would love to seeis the concept of memory,
because with memory then we canlearn.
And in that example of thecruise scenario and like, I
(22:30):
don't see memory being somethingthat can inherently sit in an
LLM because, as you said, an LLMtakes six months to train and
it's ready to go.
So somewhere outside of that,we need some concept of being
able to remember past activity,et cetera, and also to learn
from error, right, right, howdoes a rocket get to the moon?
(22:53):
It makes lots of errors all theway that it self-corrects on to
get there.
Do you think that's somethingthat's going to sit in the camp
of Copilot Studio particularly,or is that going to be something
that's going to be more on thefoundry side of the game as we
move forward and that ability tostore more long-term memory?
Dewain Robinson (23:11):
So a combination.
It's going to be like acombination, so think of it that
you already have memory.
So Copilot Studio already hasmemory in two places, but
there's a third place that'scoming right.
Copilot Studio already hasmemory in two places, but
there's a third place that'scoming right.
So the first place is there hasalways been something that, if
you go all the way back into botframework, we used to refer to
(23:31):
this as the conversation state.
So the conversation state andyour ability to write something
that you're talking about intomemory is ability to do that,
and that's where variables comealong, right.
So, if you think about it, thisgoes back to okay, everybody's
like oh, I'm going to go and useAI Studio or I'm going to go
use OpenAI's implementation andI'm just going to.
(23:53):
Everything's just in this onemodel.
Well, that's where you don'thave a conversational
orchestration engine or adialogue management system in a
sense, right, and that means youdon't have a conversation state
, so you can't store somethingthat I don't want to take into
consideration in the next turn,but later I want to use it right
(24:14):
, and so that's something thatCopilot Studio has today and
that's why our variables come in, and there's been a ton of work
on variables that nobody evenpaid attention to, like being
able to store a record in avariable and be able to update a
record within a variable, sothat you can take a whole JSON
payload and drop it into avariable and then, just whenever
(24:35):
you wanted to use it, you couldjust concatenate that into the
query set and then provide thatinto the large language model
for context.
And so then the second part isgoing to be in the prompt, so
the prompt itself is going togive you the ability to pass
this in, and everyone kind ofunderstands that one probably if
they're following thisconversation to this level.
(24:57):
And then we've got the new partright, which is when OpenAI
started introducing the conceptof memory itself and having a
way to provide memory to themodel.
That is more beyond the prompt,and that is something that
you'll see Copilot Studioembrace Now, when you talk about
(25:19):
foundry and you start talkingabout this, this is just you're
going to build something to dothis, like to pass it in and to
be able to deal with it, and sothat's kind of the value of the
whole Copilot Studio thing isthat it's not that it's low code
type of thing I think a lot ofpeople get over the moon about
low code, it's the fact that itis a SaaS platform, so that you
(25:42):
don't have to go build all thisstuff.
So, whether you really want lowcode or pro code or whatever you
want to do, that's really not aseparation in Copilot Studio.
So much anymore.
Can you use Copilot Studio anddo low code Absolutely.
Can you do pro code Absolutely?
So the question really getsdown to do you want to SaaS or
(26:05):
do you want to build it yourself?
And then some of that's goingto be I want to build this one
little piece by myself, andthat's going to be where you're
going to see, like Copilot SDKand all of that.
So you're going to see anevolution of the bot framework
SDK happening which is going tomake it where it can handle
memory and it can do all ofthose things, and you're going
(26:27):
to see the same thing infoundation and all that, and so
all these things in our work,it's just what level of control
and how much depth do you want?
Copilot Studio?
It's not low code, pro code,it's SAS or no SAS.
Do you want SAS or do you wantPaaS?
That's really better the way tothink about it and I could have
my SaaS call for thisparticular thing.
(26:48):
I need to be in more control,so I'm going to build a service
in PaaS that I'm going to callfrom my SaaS, and that's
actually the most commonconfiguration that we see.
If you just build everything inPaaS, then your ROI is going to
suck pretty bad and yourmaintenance is going to be very
(27:09):
high.
So that's why I think it's acombination of these things and
I think you'll see this get alittle clearer to people,
because at Microsoft we'relearning, just like you guys are
learning right, and you findthe mistakes like we've already
made a few mistakes Like look atM365 Copilot.
Originally there was like thislittle dropdown box that you had
where you're going to build allthese plugins, and that thing's
being deprecated.
It's coming out Right and nowyou're seeing a Copilot agent
(27:32):
replace it, and it's because werealized people wanted to be
able to at mention a Copilotagent so that they can direct it
to what they wanted to talk to,and I also think you'll
probably see in this spacethat'll help you on this mark as
well.
When it comes to memory, is Ithink we might with the
(27:53):
combination of memory, thecombination of a platform that
has a conversation state.
That can be because we're in aSaaS.
We can standardize theconversation state to make it
look standardized so that youdon't have to worry.
I can build things that can useit because it's standard the
way that it works, where, if Ilet you have an SDK, I can't
(28:17):
make everybody follow the samestandard for the way we're going
to store variables.
So now, with those thingshappening and memory coming in
and everything, I think thismight actually open the door for
something that I think you'rekind of alluding to, which is
imagine an agent that can talkto another agent or understand
(28:39):
what another agent can do andbeing able to orchestrate
between agents in a way thatdoesn't make the user really
irritated or cause a massivenightmare.
So I think there's apossibility that that might be
viable with this whole memorything coming out.
Mark Smith (28:58):
So I like it.
Dewain Robinson (29:00):
Long-winded answer, but there you go.
Mark Smith (29:01):
It's good.
It's's good we're at our finalquestion.
We've run out of time.
I have the ui ai.
Co-pilot is the ui for ai, andI see it as a dimension probably
beyond what microsoft arecurrently talking about it right
, which is it's a greatcatchphrase, whoever the
marketer you came up with it.
(29:21):
Fantastic what I've noticed inthe enterprise customers that
are implementing Copilot.
Or actually, let's back up, Anyenterprise customer has a lot
of software let's say, 200 to300 different applications that
do something in thatorganization, from the big
vendors like SAP Pega ServiceNow, these type of things Microsoft
, of course, like SAP PegaServiceNow, these type of things
Microsoft, of course.
(29:42):
And people have fatigue insideorganizations around application
switching.
I need to do something aroundinventory, so I'm going to jump
into my SAP system.
I'm going to do somethingaround my holiday leave, so I'm
going to jump into my HCM system.
Blah, blah, blah, blah Right,blah, blah, blah, blah, right.
And each of these productmanufacturers have now come up
(30:03):
with their AI interface to theirsoftware.
And what I see happening is nowpeople are going okay, which AI
do I use and how does that AIdiffer in how I prompt that one
to how I prompt that one?
And so when Microsoft came outwith this UI for AI, I'm just
like awesome.
We don't care where the datasits in your organization that
(30:26):
we're going to call as part ofthis AI process.
If I have a single interface, Ican build out my skill at
working with Copilot, the UI forAI and, like you're saying, add
the agents in as long as I'vestarted developing.
Hey, there's an agent that cando this and I can add them in,
but I don't have to keep going.
(30:46):
Am I in the right AI interfaceto do what I need to do?
And I see the UI for AI as acompelling story by lighting up
the agent experience that youcan do with Copilot Studio, by
bringing in the ServiceNow data,but I don't need to use the
ServiceNow AI.
I don't need to use the SAP AIif I basically can consume it
(31:07):
through APIs.
And this is where orchestrationcomes back into play, because
when I'm talking about purchaseorders or order numbers, we're
talking about a connection toSAP here.
I'm going to go and retrievefrom that as part of what I do.
Dewain Robinson (31:22):
So I think it's close, but I think where this
is going to this is a guess,right?
So, just to be clear, this ishow I'm thinking.
This is based upon what I'mseeing, I think how we get there
.
It's not that I need to do itLike, let's say, servicenow has
got this really cool AI thatthey come out with that does a
really great job over ServiceNowdata, because they know their
(31:44):
data structure, they know how tosearch it, they know how to
query it, and so they canprobably write a really
compelling AI for their self.
But what they can't do is alsobring in Salesforce.
No, salesforce is the expert AIover there, right?
And so the problem we ran intois I've said this many times I
(32:04):
said the answer isn't one AI.
It's not to build Iron Man's AI.
Why can't I remember TonyStark's AI?
So forgive me about myinabilities to be geeky here,
but the idea is you're not goingto have that, because humans
(32:25):
speak with context.
That's not spoken so, like if Isay what is the finance rate?
If I'm talking to an HR person,I mean something different than
if I'm talking to someone inthe finance department for the
organization.
One's corporate tax rate andthe other is personal tax rate,
(32:46):
and if I don't know the context,then it doesn't work.
This is why co-pilot agents hadto be incorporated with, where
you can do the app and talk to aco copilot agent so that you
can say I'm asking this to theHR thing, I'm asking this to
that.
So that's how we're handlingthe unspoken context, right.
(33:08):
But if you think about it, thesesub-AIs there's no reason that
an AI can't ask another AI aquestion and then use that as
how to respond.
So imagine if an AI under itsfunction was to and it goes back
to orchestration again, yeah,imagine that we have an AI that
(33:30):
is a master orchestration enginethat can also have good
connections to call the API toanother AI to ask it a question.
Yes, okay, not to access thedata, but to just ask it a
question.
I love it.
I love it.
Now I get a response from theAI who really understands its
(33:51):
data and its stuff, and it comesback with it and imagine, then
I ask the question in order forme to answer that question, I
actually have to go to threesystems, three different other
AIs, to be able that it has togo there to go figure it out.
So what I'm think where we areright now is we're in the
thousands of AIs, right, but thereality is we don't need one
(34:16):
because the human context,unspoken context, problem.
But what we might need is a fewthat that's the HR one, that's
the corporate one or whatever.
So imagine, now that CopilotStudio starts to slot itself
into a place where it is theconversational orchestrator for
Microsoft that is configurable,is the conversational
(34:39):
orchestrator for Microsoft thatis configurable.
And think about what I saidabout the explanation of all the
different tools, and the AI hasjust become yet another tool,
and now you can say I couldactually create this.
The reason I think it's goingto work this way, mark, is I
think this is what I told you atthe very beginning if we do a
callback to the beginning, thething that makes me think this
is the thing that was theunexpected value prop, which was
(35:04):
even a lexical search, gotbetter when I was able to have
an AI build the query with thecontext.
So imagine now how good an AIcould build a question that it
needs to ask another AI to getwhat it needs from the response.
And so I think, if you thinkabout it, that's where Copilot
(35:25):
Studio is going to become likethe superpower, because it will
be the AI of AIs.
Yeah, it will be the thing thatyou explain how to go get all
these things and how they allcome together, and you're not
going to program that in C sharp.
You can program that in Englishor Dutch or whatever your
spoken language is.
This is why a lot of peopleneed to be thinking about prompt
(35:50):
engineering.
They need to be thinking aboutthat new skillset, which is
programming in natural language.
Being like a lawyer, how do Iwrite an explanation that's very
crisp and take out all theloopholes so that AI behaves how
I want it to behave, because inthat orchestration world, it's
(36:12):
going to be super important.
So this is the part that'sactually going to become very
interesting, and it needs to bea SaaS, because you need a
runtime engine that's going toexecute consistently, and so
this is Copilot Studio.
Mark Smith (36:28):
I like it, dwayne, this has been awesome.
I'm excited about this year.
It's going to be an epic year.
Dewain Robinson (36:33):
Well, mark, I think everybody's going to think
we pre-planned that questionand I just want to go on the
record and say, guys, actuallynot, this is actually just a
really good riff with you,because it's interesting how all
this starts just comingtogether, you know.
Mark Smith (36:47):
Yeah, I love it.
Thank you, sir.
I hope to see you in, probablyMarch MVP Summit.
If you're around, I'll see youthere.
Dewain Robinson (36:53):
I will be there and I am looking forward to
getting to see all my wonderfulMVPs, like yourself and you know
.
If there's anybody that'swanting to kind of follow me,
feel free to follow me onLinkedIn or, I think, twitter.
I'm like Dwayne 76 and Ispelled Dwayne really weird.
So look at the show notes andif all else fails, you can go to
CopilotStudioDudecom, andthat's where you'll find the
(37:21):
YouTube channel that we'vetalked about, and so there's a
lot of stuff I'm going to beputting on that.
I'm going to kind of doubledown on that this year mark, so
I'm definitely interested inseeing stuff, and you need to
give me feedback.
By the way, you and all yourcommunity need to get youtube.
Mark Smith (37:32):
Stuff is amazing.
Right, it's the co-pilot sourceto go to and to as your videos
drop.
It's epic.
Thank you, sir.
Dewain Robinson (37:41):
Well, I appreciate it.
You are too kind, sir.
You are too kind, yeah Well,like I said, exciting 2025, and
we can't make it this long againBetween the next time I talk to
you.
Agreed, there you go.
Mark Smith (37:55):
Awesome, I love it, I love it, thank you.
There you go.
Awesome, I love it, I love it.
Thank you, sir.
Bye-bye.
Hey, thanks for listening.
I'm your host, mark smith,otherwise known as the nz365 guy
.
Is there a guest you would liketo see on the show from
microsoft?
Please message me on linkedinand I'll see what I can do.
Final question for you how willyou create with copilot today?
Welcome to the Copilot Show, where I interview
Microsoft staff innovating withAI.
I hope you will find thispodcast educational and inspire
you to do more with this greattechnology.
Now let's get on with the show.
Today's guest is from NationalTennessee in the US.
He works at Microsoft as aPrincipal Program Manager for
(00:23):
Copilot Studio andConversational AI.
He helps people leverageCopilot Studio to build amazing
copilots.
You can find links to his bioand socials in the show note for
this episode.
He was last on the show forepisode 497, and we talked about
Power Virtual Agents, so thatwas a wee while ago, but welcome
(00:43):
back, Dewain.
Dewain Robinson (00:45):
Thanks, I'm glad to be here, mark.
It's awesome to get to talk toyou again.
We just were kind of talkingbefore that.
It's been so long and Icouldn't even believe it was.
Yeah, it's hard for me tobelieve it's even been that long
.
It just seems like yesterdayyou and I were chatting about
this.
Mark Smith (01:00):
Things are changing so, so rapidly.
If you look at you know we'rerecording this at the start of
2025.
I feel we've come off an epicyear where corporations are
really starting to go.
What does AI look for us?
And then the last three monthsI feel we were heavily there was
heavy marketing, not just fromMicrosoft, but all the major
(01:23):
players in the world in thegenerative AI space LLMs, et
cetera.
Around that, 2025 is going tobe the year of agents.
So, with that backdrop beforeus, when you look at this year
ahead, what excites you rightnow the most?
Dewain Robinson (01:40):
I think the thing that is going to be very
interesting in the CopilotStudio world, right, is, I would
say, that there's thisperception of AI and what it can
do, and then there's thereality of what it could do and
how easy it is to do that withthat AI back last year.
(02:02):
Right, you could really buildsome really cool stuff With that
AI back last year.
Right, you could really buildsome really cool stuff.
And I think what's happening isthe commoditization, the
roughing of the edges that werea little bit in there.
And so I think that in itselfor what I would call like a
synchronous or a, you know, kindof like a synchronous co-pilot,
where you talk to it and ittalks back to you is going to be
(02:24):
one of the things that is goingto be one of the things that is
going to be the majorinnovation.
It's not like oh, we're goingto see something that we never
thought of.
It's just going to be thethings that you are wanting to
do will become very easy andcommoditized, and the
integrations to be able to tiethat in with Azure and other
pieces are just going to becomeeasier and easier, and you could
(02:45):
do it before, but now it'sgoing to be like oh, there's a
button and it's an easy buttonto do it.
And then I think the other thingI would say, mark, is that as
we walk into the autonomousagent space, that's going to
really change the perception.
It's beyond the concepts ofwhat you can do with AI by
(03:08):
itself.
It's now AI working inconjunction with business
process automation to be able todo something you could never do
before.
And I think the combination ofthese two things are really
going to be a year.
That's just insane.
And, as you're aware, I meanthe innovation on the AI side
and the AI model side is justgoing to keep going too.
(03:29):
So, when we talk again and youknow, because we're not going to
wait so long, right, mark sothen when we talk again next
time, I would anticipate there'sprobably going to be some
innovation in the model spacethat you can do and stuff like
that, because we already startedseeing some of that.
Like you can now stream audioin and you're not having to do
text to speech.
Yeah, like that's a big deal,right.
(03:51):
And so I think there's going tobe a lot of interesting things
when we start going beyond justchatware.
Yeah, I think chat stuff isgoing to get super easy to do,
but then these other things arejust going to really take it to
a level that people reallydidn't think about, and I think
that's the beauty of it.
And this is the year that Ithink it's not just about open
AI.
It's about open AI's integrate,how Microsoft's going to take
(04:14):
open AI and combine it with theMicrosoft asset pool, and then
you're going to get somethingthat's completely different,
that you couldn't do before,which I think is really awesome.
Mark Smith (04:24):
Have you seen, I suppose in 2024, looking at the
year in reverse, withoutrevealing any customer details
or anything did you see anythingthat was quite you're like?
You know that was cool, that'sa cool use case, or that was a
cool way to use a technology?
Because you mentioned voicethere and I think voice is a
massively underexploredterritory in how we go forward.
(04:46):
I just had a customer the otherday say listen, when I fill out
my CRM system, I really justwant to be able to open the
record and say the first name isBob, the second name is Frank
and it just types it in justwith me talking and you know,
auto filling out, that type offunctionality which we get that
degree coming in in the Dynamicssuite of products.
And but it was interesting.
(05:07):
This is an old guy, but his wayof how you know he would like
to use AI combined with what hewas doing and but was there
anything that really kind ofstood out top of mind story that
was highlights for you lastyear?
Dewain Robinson (05:23):
I had a couple of things, that kind of top of
mind, and I will say there wasthis thing that when I think
back where I got value that Ididn't expect.
So it's not necessarily likethe use case in this scenario.
Well, I'll say this use casewas fairly interesting because
and everyone's heard about theuse case, because it's the
(05:45):
Holland America one that we talkabout publicly, right but the
thing that was about that Ididn't expect.
You know, we had done businessvalue assessments and all this
stuff about what we expected,the return and how this thing
was going to perform andeverything.
And we were expecting to getlike a return from like over in
(06:07):
this area, you know on the leftand what also, but we didn't
expect this value area on theright.
We didn't even know that therewould be impact there.
And and what it came down to isthat if you think about it,
think about I have an API or Ihave something that you want to
do, and in the case of this it'slike cruises I want to find a
(06:27):
cruise.
So when we were doing this, thecustomer had an API that had
here are all the differentcruises that you might want, and
so we needed to go search for acruise that would be the right
cruise for you cruise that wouldbe the right cruise for you.
(06:49):
And what we found and thisbecame similar because it
wouldn't matter that it wasfinding cruises it could be like
you wanted to searchServiceNow's knowledge base
right For knowledge articles.
What we didn't expect at thetime and it makes logical sense
now that this would be the caseis what we were doing is using
an AI model to look at thecontext of the whole
(07:11):
conversation and generate a lereally improved search results,
coming back from an existingsearch service that then could
(07:31):
be consumed and consumed by theAI with the context as well, but
knowing that the search qualitywas going to go up and the
search results were going to bebetter off of your API.
That was just a standardlexical search.
I didn't expect that.
I didn't think that that wouldbe something that was the case.
But we saw that not only withthe Holland America example, but
(07:55):
we also were working withanother company and they were
trying to do something withServiceNow knowledge articles
and we saw the same thing and wewent oh man, we didn't think
about that.
And so there's actually someyou know features and stuff that
we started putting into CopilotStudio because we realized
after we saw that that that'sgoing to mean that a lot of
(08:15):
people are going to want to beable to do that and we want to
make that easier to do and allof that.
And so I think that was kind oflike a really aha moment for me
when I was looking at that.
And then the second thing thatwas like really cool is I did a
video that showed the neworchestration engine.
Mark Smith (08:36):
I was just going to ask you about it.
Dewain Robinson (08:37):
Yeah, yeah, and I did one where I do one with
the PO, like I'm creating a PO,where I do one with the PO, like
I'm creating a PO.
And what a lot of people don'tknow is that I'm building this
thing.
And I get on the phone and Icall up Gary Pretty, which is
the lead over it, and I said,hey, I need you to tell me what
are all the things that I needto show to demo this thing in
(09:00):
the best light.
And so I built this thing withno intention, right?
I was just wanting to checkthis box, check this box, check
this box.
And I went through it allmarket and I built all this
stuff with all the differentpieces to do it.
And then, out of a fluke, I justwent, wait a minute, if it has
all that, if I say, create metwo POs with this one with this
(09:21):
and this one with that andeverything.
And it just did it.
It literally figured out allthe assets and did it and
dynamically assembled it and didit not just once, but did it
for two different things,summarize it all out and spit it
out.
And that demo and that videothat you saw from that stems
from me building that as a wayto demonstrate it on a stage, to
(09:43):
show people this new feature.
But then I'm trying to dosomething real world and then it
occurs to me that wait a minute, this thing can do more than
what I really thought about whenI was building it, which
showcases the fact that if youdo these things right and you
modularize it right and yourmodel descriptions are good,
it'll actually do stuff that youdidn't even intend it to do,
(10:06):
because it's flexible enough tofigure it out.
And that's kind of a really bigaha moment.
We even had one of our guysthat was demoing on a stage and
he was showing something and wehad negative feedback on him
because he said on the stage hewent oh wow, get a feedback on
him.
Because he said on the stage hewent oh wow, that worked like I
(10:27):
didn't expect, and so, likethose things were really
impressive to me as I was goingthrough.
So those are the standout onesfor me right now.
Mark Smith (10:34):
I like it, I like it .
And now, in 2025, and we seethis concept and what I'd like
you to do actually start withyour description of what agents
are what is semi-autonomous,what's autonomous and, let's say
, what's fully autonomous orreally you know, set and forget
type agent scenarios.
I think orchestration playsinto that mix, because you're
(10:56):
going to have these agents thathave deeper expertise, but
you're going to need anorchestrator to know which agent
to call or which one to fire ina scenario.
Let's start with how do youexplain the concept of agents
that we're hearing about now?
Dewain Robinson (11:13):
Yeah, so there was a little bit of a rebrand.
That happened, and Microsoftwere great at marketing and
rebranding things without beingclear that we did it.
So the word co-pilot let'sstart there.
Co-pilot was being used likeeverywhere.
Yes, right.
(11:33):
And so we started to get aproblem, which was nobody
understood what the wordco-pilot meant, and then we
started using co-pilot and theword agent, and then we started
getting into apilot and the wordagent, and then we started
getting into a new one which wasgoing to be a function, an
extension and a plugin.
And all of these words startedbecoming a problem because
(11:55):
people were using theminterchangeably and I literally
had two sales guys fighting infront of a customer on what
co-pilot did, and they didn'teven realize they were both
right, they were talking aboutdifferent products.
And so the thing that if westart out what is a Copilot now
by branded by Microsoft, you'regoing to see that it's going to
have like a big C in the front,and the reason behind that is
(12:17):
Copilot is a product fromMicrosoft, it is a product that
you configure, and sure youcould say it's an agent that's
prepackaged and available to you, right?
And so that's why you're goingto see this word co-pilot.
Now, an agent is going to bedifferent and you will see an
agent with a small A, not with abig A, and the reason behind
(12:42):
that and this is, by the way,I've had to have a long talk
with marketing to get this, sojust trust me, it's not like
this is common nature.
The word agent is meant to sayit is a conversational AI
implementation that usesgenerative AI that you're in
control of that you build andthat you are in control of the
(13:03):
orchestration.
When you go over into, likeCopilot for M365, you don't have
full control of theorchestration engine.
Now you can throw something inand you can extend it and you
can do all that, but you can'tmake it not do something that
it's programmed to do, whereasan agent means you are in
control of it, and so if youthink the word co-pilot agent
(13:26):
big C, little A means that'swhat an extension is of a
co-pilot product is a co-pilotagent.
And so if you're saying agent,then we're just talking about
something you built.
Now how you build that thing,it can be whatever.
And so, yeah, we always thoughtit was funny and you'll get a
kick out of this Mark is that wechanged the name from Power
(13:48):
Virtual Agents to Copilot Studio.
But now, copilot Studio doesn'tbuild copilots, it builds
agents.
Mark Smith (13:54):
Yes.
Dewain Robinson (13:55):
So I had fun with marketing on that one, yes,
but the thing is, is so inCopilot Studio, what are you
going to do?
You're going to build agents.
Right Now, can you build agentsin different ways?
Absolutely, you can, butfundamentally understand that an
agent that is a realintelligent agent, that is
(14:16):
something bigger than just amodel, is going to need to have
an orchestration engine.
And that is what Copilot Studiois.
It's an out of the box, it is aSaaS-based conversational
orchestration engine.
Right, and we saw the evolutionof that to incorporate using not
a traditional language model,and for those that don't know
(14:36):
what I mean by that, it used tobe that you would have to train
an NLU model.
You'd say here's six differentways of saying this thing, and
that way I can say this is theintent.
And then you could do the samething with entities, which, if
you want to understand thedifference between intents and
entities, an intent is if I saidI want to order a large
pepperoni pizza.
The intent is that I want toorder something, then the
(15:00):
entities are large pepperonipizza.
The entities are largepepperoni pizza, and so, being
able to do that, you had totrain that into a natural
language understandingimplementation and traditionally
, this is where large languagemodels came along, because you
don't train one Not typicallyyou don't train one and, matter
of fact, to train a largelanguage model from the ground
(15:22):
up takes like six months to dowith a massive data center.
It's worth the GPUs to be ableto get it done.
So, when we're looking at these, you don't do that anymore.
How you use these is you hadthe innovation that happened
where I don't need to do intentrouting or determining what tool
to use.
I can use a orchestrationengine that's not powered by a
(15:45):
traditional NLU.
I can do it with a largelanguage model, and this is what
you heard us talk aboutgenerative actions and
everything, and you're going tosee like you'll hear us refer to
this as generativeorchestration now, because we
realize like actions, which isthe wrapper around an API, we
needed to separate that from theorchestration engine capability
, and I'll explain why in just asecond.
(16:06):
So, once we've done this, thisis where it really gets into
what you're talking about.
What is an agent?
Okay, so now we can start toclassify them.
Well, and there's differentagent types.
Now you have what I refer to asa synchronous agent.
We just call it an agent, whichis what you traditionally think
(16:27):
of as like a chat, gpt orbasedorchestrator that could
understand the context of theconversation, and or the Power
Automate flow, to go create a PO, and here's the one to go get
(17:04):
an approval, and oh, here's oneto go get information about our
HR policy and all that.
So what you're seeing is it'sbasically acting as like a
router, a conversational router,and that, by the way, is what
you know.
Conversational orchestration isyes.
So that's when we started to seethat you're going to start to
see these other agent types, andI think this leads into a thing
(17:29):
that I know many people getconfused about, which is what is
the difference between whatyou're hearing us talk about
autonomous agents.
And why is that?
Because people think about itand they go oh well, smart,
business process automation orsomething like that and they go
well, wait a minute, how is thatdifferent than Power Automate?
And because now we're startingto see these classifications
(17:51):
where it's not you're talking toit and it's talking back.
Now it can be that somethingelse triggers it and then it
does something, and a lot ofpeople go well, that's Power
Automate.
So Power Automate can havesomething that triggers it and
then it does something and a lotof people go well, that's power
automate.
So power automate can havesomething that triggers it and
then you can follow a businessprocess and everything.
So here's the difference, mark,and this is why it was so
(18:12):
important to land aconversational orchestrator
based off a large language model.
Because now imagine back up toCopilot Studio for a second.
Imagine if I was going to hiresomeone.
I'm going to write a jobdescription.
I take that and I say here'swhat your job is, here are the
rules you should follow, hereare the tools at your disposal
(18:33):
to be able to do your job.
And this is what your job is.
And when should you do your job?
Well, when you receive an email, when someone puts a record in
a CRM, when someone fills out aform like whatever, right.
So there's this thing.
That's not that someone'stalking to you, but guess what?
Those autonomous agents cantalk to you, but they don't have
(18:55):
to, because they can go andlook at it and imagine I have
six different power automateflows.
I explain what they all do.
I come over here and I say, oh,here's some knowledge with like
information on products andinformation about business
processes and how those thingsshould, what those policies are
and things of that nature.
(19:16):
And then I have, like, allthese different tools, and even
a topic could be a tool, yeah,so now I'm triggering those
things and it can do whatever Ineeded to go do based upon
something going on.
So one of the demos that we didI mean I didn't show this one
publicly.
I can't remember if I made thevideo for this one or not.
(19:36):
I have to go back and look whatI did.
I think this is the video I'mworking on right now.
But the problem, marcus, it'sgetting harder to build the
crazy complexity that's going tomake everybody go wow, because
I know I can do it, but I haveto go build it all.
But what I did at thatparticular thing at the CIO
Summit is we took and made aform that you could go ask it
(19:57):
about what are cruise specials?
What are cruise specials?
And it would go look you upbased upon your email address
and figure out what you hadcruised in the past, make a
product recommendation, look atwhat the specials are, figure
out what you like, go in andcreate a CRM and implement you
know record for it, update theCRM record, then make you an
email and email it to the personand giving them the cruise
(20:19):
specials.
And then someone could emailback to it and be like, hey, you
gave me these three options,but I was thinking about I
wanted to cruise Greece thisyear Do you have any specials
for that?
And it could then converse withthem and then, and all the time
when it's doing this, it'supdating the CRM system for
people to know what was going onand it's helping with the
pipeline.
So the thing is, is that that'swhere you get into autonomous
(20:43):
agents?
I think everyone gets confusedbecause they make it too simple.
They think it's like justbusiness process automation or
that there's a step within thebusiness process automation
where I want to look at themodel and figure out what to do
next.
But it is literally like thebrain of the orchestration
engine being able to betriggered and then it understand
what its job is and the toolsit has, so that it can just
(21:04):
figure the brain of theorchestration engine being able
to be triggered and then itunderstand what its job is and
tools it has, so that it canjust figure out what's the right
thing to go do to be able tocomplete it.
And right now, I would say we'rein the infancies of what it can
do, so you might hit somelimitations and stuff, because
it was only in preview, but theintent is that it works like
that, and a lot of people thenfreak out when I talk about this
(21:24):
, mark, and they go oh my God,it's going to replace me for my
job and it's like no, it's goingto take the thing that you have
to do over and over and overagain and make you not have to
do that so you can work on thethings that you need to provide
the higher quality, and so Ireally think this is the thing
that's really going to beamazing in 2025.
Mark Smith (21:45):
If we look at that example that you've just given
and we say it's a jobdescription, right, if you look
at a job description for a role,it'll fit on one side of a
sheet of paper maybe.
Then you start doing the joband you realize it's really 300
sheets of paper because we don'ttell you that you're going to
have to be in these meetings andyou're going to have to, you
(22:07):
know, reply to an aggregatedemail.
We don't put all those nuancesof the job in the job
description.
But you start to pick it up andwhen I think of agents and the
element that I would love to seeis the concept of memory,
because with memory then we canlearn.
And in that example of thecruise scenario and like, I
(22:30):
don't see memory being somethingthat can inherently sit in an
LLM because, as you said, an LLMtakes six months to train and
it's ready to go.
So somewhere outside of that,we need some concept of being
able to remember past activity,et cetera, and also to learn
from error, right, right, howdoes a rocket get to the moon?
(22:53):
It makes lots of errors all theway that it self-corrects on to
get there.
Do you think that's somethingthat's going to sit in the camp
of Copilot Studio particularly,or is that going to be something
that's going to be more on thefoundry side of the game as we
move forward and that ability tostore more long-term memory?
Dewain Robinson (23:11):
So a combination.
It's going to be like acombination, so think of it that
you already have memory.
So Copilot Studio already hasmemory in two places, but
there's a third place that'scoming right.
Copilot Studio already hasmemory in two places, but
there's a third place that'scoming right.
So the first place is there hasalways been something that, if
you go all the way back into botframework, we used to refer to
(23:31):
this as the conversation state.
So the conversation state andyour ability to write something
that you're talking about intomemory is ability to do that,
and that's where variables comealong, right.
So, if you think about it, thisgoes back to okay, everybody's
like oh, I'm going to go and useAI Studio or I'm going to go
use OpenAI's implementation andI'm just going to.
(23:53):
Everything's just in this onemodel.
Well, that's where you don'thave a conversational
orchestration engine or adialogue management system in a
sense, right, and that means youdon't have a conversation state
, so you can't store somethingthat I don't want to take into
consideration in the next turn,but later I want to use it right
(24:14):
, and so that's something thatCopilot Studio has today and
that's why our variables come in, and there's been a ton of work
on variables that nobody evenpaid attention to, like being
able to store a record in avariable and be able to update a
record within a variable, sothat you can take a whole JSON
payload and drop it into avariable and then, just whenever
(24:35):
you wanted to use it, you couldjust concatenate that into the
query set and then provide thatinto the large language model
for context.
And so then the second part isgoing to be in the prompt, so
the prompt itself is going togive you the ability to pass
this in, and everyone kind ofunderstands that one probably if
they're following thisconversation to this level.
(24:57):
And then we've got the new partright, which is when OpenAI
started introducing the conceptof memory itself and having a
way to provide memory to themodel.
That is more beyond the prompt,and that is something that
you'll see Copilot Studioembrace Now, when you talk about
(25:19):
foundry and you start talkingabout this, this is just you're
going to build something to dothis, like to pass it in and to
be able to deal with it, and sothat's kind of the value of the
whole Copilot Studio thing isthat it's not that it's low code
type of thing I think a lot ofpeople get over the moon about
low code, it's the fact that itis a SaaS platform, so that you
(25:42):
don't have to go build all thisstuff.
So, whether you really want lowcode or pro code or whatever you
want to do, that's really not aseparation in Copilot Studio.
So much anymore.
Can you use Copilot Studio anddo low code Absolutely.
Can you do pro code Absolutely?
So the question really getsdown to do you want to SaaS or
(26:05):
do you want to build it yourself?
And then some of that's goingto be I want to build this one
little piece by myself, andthat's going to be where you're
going to see, like Copilot SDKand all of that.
So you're going to see anevolution of the bot framework
SDK happening which is going tomake it where it can handle
memory and it can do all ofthose things, and you're going
(26:27):
to see the same thing infoundation and all that, and so
all these things in our work,it's just what level of control
and how much depth do you want?
Copilot Studio?
It's not low code, pro code,it's SAS or no SAS.
Do you want SAS or do you wantPaaS?
That's really better the way tothink about it and I could have
my SaaS call for thisparticular thing.
(26:48):
I need to be in more control,so I'm going to build a service
in PaaS that I'm going to callfrom my SaaS, and that's
actually the most commonconfiguration that we see.
If you just build everything inPaaS, then your ROI is going to
suck pretty bad and yourmaintenance is going to be very
(27:09):
high.
So that's why I think it's acombination of these things and
I think you'll see this get alittle clearer to people,
because at Microsoft we'relearning, just like you guys are
learning right, and you findthe mistakes like we've already
made a few mistakes Like look atM365 Copilot.
Originally there was like thislittle dropdown box that you had
where you're going to build allthese plugins, and that thing's
being deprecated.
It's coming out Right and nowyou're seeing a Copilot agent
(27:32):
replace it, and it's because werealized people wanted to be
able to at mention a Copilotagent so that they can direct it
to what they wanted to talk to,and I also think you'll
probably see in this spacethat'll help you on this mark as
well.
When it comes to memory, is Ithink we might with the
(27:53):
combination of memory, thecombination of a platform that
has a conversation state.
That can be because we're in aSaaS.
We can standardize theconversation state to make it
look standardized so that youdon't have to worry.
I can build things that can useit because it's standard the
way that it works, where, if Ilet you have an SDK, I can't
(28:17):
make everybody follow the samestandard for the way we're going
to store variables.
So now, with those thingshappening and memory coming in
and everything, I think thismight actually open the door for
something that I think you'rekind of alluding to, which is
imagine an agent that can talkto another agent or understand
(28:39):
what another agent can do andbeing able to orchestrate
between agents in a way thatdoesn't make the user really
irritated or cause a massivenightmare.
So I think there's apossibility that that might be
viable with this whole memorything coming out.
Mark Smith (28:58):
So I like it.
Dewain Robinson (29:00):
Long-winded answer, but there you go.
Mark Smith (29:01):
It's good.
It's's good we're at our finalquestion.
We've run out of time.
I have the ui ai.
Co-pilot is the ui for ai, andI see it as a dimension probably
beyond what microsoft arecurrently talking about it right
, which is it's a greatcatchphrase, whoever the
marketer you came up with it.
(29:21):
Fantastic what I've noticed inthe enterprise customers that
are implementing Copilot.
Or actually, let's back up, Anyenterprise customer has a lot
of software let's say, 200 to300 different applications that
do something in thatorganization, from the big
vendors like SAP Pega ServiceNow, these type of things Microsoft
, of course, like SAP PegaServiceNow, these type of things
Microsoft, of course.
(29:42):
And people have fatigue insideorganizations around application
switching.
I need to do something aroundinventory, so I'm going to jump
into my SAP system.
I'm going to do somethingaround my holiday leave, so I'm
going to jump into my HCM system.
Blah, blah, blah, blah Right,blah, blah, blah, blah, right.
And each of these productmanufacturers have now come up
(30:03):
with their AI interface to theirsoftware.
And what I see happening is nowpeople are going okay, which AI
do I use and how does that AIdiffer in how I prompt that one
to how I prompt that one?
And so when Microsoft came outwith this UI for AI, I'm just
like awesome.
We don't care where the datasits in your organization that
(30:26):
we're going to call as part ofthis AI process.
If I have a single interface, Ican build out my skill at
working with Copilot, the UI forAI and, like you're saying, add
the agents in as long as I'vestarted developing.
Hey, there's an agent that cando this and I can add them in,
but I don't have to keep going.
(30:46):
Am I in the right AI interfaceto do what I need to do?
And I see the UI for AI as acompelling story by lighting up
the agent experience that youcan do with Copilot Studio, by
bringing in the ServiceNow data,but I don't need to use the
ServiceNow AI.
I don't need to use the SAP AIif I basically can consume it
(31:07):
through APIs.
And this is where orchestrationcomes back into play, because
when I'm talking about purchaseorders or order numbers, we're
talking about a connection toSAP here.
I'm going to go and retrievefrom that as part of what I do.
Dewain Robinson (31:22):
So I think it's close, but I think where this
is going to this is a guess,right?
So, just to be clear, this ishow I'm thinking.
This is based upon what I'mseeing, I think how we get there
.
It's not that I need to do itLike, let's say, servicenow has
got this really cool AI thatthey come out with that does a
really great job over ServiceNowdata, because they know their
(31:44):
data structure, they know how tosearch it, they know how to
query it, and so they canprobably write a really
compelling AI for their self.
But what they can't do is alsobring in Salesforce.
No, salesforce is the expert AIover there, right?
And so the problem we ran intois I've said this many times I
(32:04):
said the answer isn't one AI.
It's not to build Iron Man's AI.
Why can't I remember TonyStark's AI?
So forgive me about myinabilities to be geeky here,
but the idea is you're not goingto have that, because humans
(32:25):
speak with context.
That's not spoken so, like if Isay what is the finance rate?
If I'm talking to an HR person,I mean something different than
if I'm talking to someone inthe finance department for the
organization.
One's corporate tax rate andthe other is personal tax rate,
(32:46):
and if I don't know the context,then it doesn't work.
This is why co-pilot agents hadto be incorporated with, where
you can do the app and talk to aco copilot agent so that you
can say I'm asking this to theHR thing, I'm asking this to
that.
So that's how we're handlingthe unspoken context, right.
(33:08):
But if you think about it, thesesub-AIs there's no reason that
an AI can't ask another AI aquestion and then use that as
how to respond.
So imagine if an AI under itsfunction was to and it goes back
to orchestration again, yeah,imagine that we have an AI that
(33:30):
is a master orchestration enginethat can also have good
connections to call the API toanother AI to ask it a question.
Yes, okay, not to access thedata, but to just ask it a
question.
I love it.
I love it.
Now I get a response from theAI who really understands its
(33:51):
data and its stuff, and it comesback with it and imagine, then
I ask the question in order forme to answer that question, I
actually have to go to threesystems, three different other
AIs, to be able that it has togo there to go figure it out.
So what I'm think where we areright now is we're in the
thousands of AIs, right, but thereality is we don't need one
(34:16):
because the human context,unspoken context, problem.
But what we might need is a fewthat that's the HR one, that's
the corporate one or whatever.
So imagine, now that CopilotStudio starts to slot itself
into a place where it is theconversational orchestrator for
Microsoft that is configurable,is the conversational
(34:39):
orchestrator for Microsoft thatis configurable.
And think about what I saidabout the explanation of all the
different tools, and the AI hasjust become yet another tool,
and now you can say I couldactually create this.
The reason I think it's goingto work this way, mark, is I
think this is what I told you atthe very beginning if we do a
callback to the beginning, thething that makes me think this
is the thing that was theunexpected value prop, which was
(35:04):
even a lexical search, gotbetter when I was able to have
an AI build the query with thecontext.
So imagine now how good an AIcould build a question that it
needs to ask another AI to getwhat it needs from the response.
And so I think, if you thinkabout it, that's where Copilot
(35:25):
Studio is going to become likethe superpower, because it will
be the AI of AIs.
Yeah, it will be the thing thatyou explain how to go get all
these things and how they allcome together, and you're not
going to program that in C sharp.
You can program that in Englishor Dutch or whatever your
spoken language is.
This is why a lot of peopleneed to be thinking about prompt
(35:50):
engineering.
They need to be thinking aboutthat new skillset, which is
programming in natural language.
Being like a lawyer, how do Iwrite an explanation that's very
crisp and take out all theloopholes so that AI behaves how
I want it to behave, because inthat orchestration world, it's
(36:12):
going to be super important.
So this is the part that'sactually going to become very
interesting, and it needs to bea SaaS, because you need a
runtime engine that's going toexecute consistently, and so
this is Copilot Studio.
Mark Smith (36:28):
I like it, dwayne, this has been awesome.
I'm excited about this year.
It's going to be an epic year.
Dewain Robinson (36:33):
Well, mark, I think everybody's going to think
we pre-planned that questionand I just want to go on the
record and say, guys, actuallynot, this is actually just a
really good riff with you,because it's interesting how all
this starts just comingtogether, you know.
Mark Smith (36:47):
Yeah, I love it.
Thank you, sir.
I hope to see you in, probablyMarch MVP Summit.
If you're around, I'll see youthere.
Dewain Robinson (36:53):
I will be there and I am looking forward to
getting to see all my wonderfulMVPs, like yourself and you know
.
If there's anybody that'swanting to kind of follow me,
feel free to follow me onLinkedIn or, I think, twitter.
I'm like Dwayne 76 and Ispelled Dwayne really weird.
So look at the show notes andif all else fails, you can go to
CopilotStudioDudecom, andthat's where you'll find the
(37:21):
YouTube channel that we'vetalked about, and so there's a
lot of stuff I'm going to beputting on that.
I'm going to kind of doubledown on that this year mark, so
I'm definitely interested inseeing stuff, and you need to
give me feedback.
By the way, you and all yourcommunity need to get youtube.
Mark Smith (37:32):
Stuff is amazing.
Right, it's the co-pilot sourceto go to and to as your videos
drop.
It's epic.
Thank you, sir.
Dewain Robinson (37:41):
Well, I appreciate it.
You are too kind, sir.
You are too kind, yeah Well,like I said, exciting 2025, and
we can't make it this long againBetween the next time I talk to
you.
Agreed, there you go.
Mark Smith (37:55):
Awesome, I love it, I love it, thank you.
There you go.
Awesome, I love it, I love it.
Thank you, sir.
Bye-bye.
Hey, thanks for listening.
I'm your host, mark smith,otherwise known as the nz365 guy
.
Is there a guest you would liketo see on the show from
microsoft?
Please message me on linkedinand I'll see what I can do.
Final question for you how willyou create with copilot today?
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.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Special Summer Offer: Exclusively on Apple Podcasts, try our Dateline Premium subscription completely free for one month! With Dateline Premium, you get every episode ad-free plus exclusive bonus content.