Good Morning. [applause]
Good morning. It's fantastic uh to be in
Madrid uh and in Spain and uh to be in a
theater. In fact, I was just thinking of
the last time I was in a theater like
this this big was uh growing up in India
watching some Bollywood flick uh as a
middle schooler and being you know
completely taken aback by the size of
the theater. Man, this is huge. Uh so
it's fantastic uh to be with all of you
um especially at a time like this right
when there is so much you know talking
about you know um popular culture who
would have thought uh that you pick up
any paper you pick up any news feed and
all you hear about is AI everywhere um
and it's exciting uh and so what I
wanted to do today was to spend some
time perhaps trying to give our point of
view on what I think collectively what
this technology means to us and what we
can do with it. Uh to us it always
starts with our mission. Uh our mission
as a company is about empowering people
and organizations
to achieve more. Right? That's where we
start from. And whenever there's been a
new technology, however powerful, we
have always tried to focus on can that
technology realize that mission. uh
which is can we put it in the hands of
people and in the institutions that they
build and can that organization that
person uh really change the frontier of
what they do right in fact we have this
entire description of what a frontier
firm is right because if you think about
it right any technology is an input uh
ultimately what we try to do with it is
change some output um and that's what
the movement of the frontier is uh and
in practice practical terms for us u you
know in the past we talked about digital
transformation and now we're thinking
and talking about uh AI transformation
how can an organization use AI to change
the customer experience the employee
experience the internal operations bend
the curve of innovation right whatever
goods or services you're producing can
you now do it with more throughput a
much more expans with much better speed.
Uh that really is the bending of the uh
frontier. Now the interesting thing with
AI is that you now have to perhaps you
know take a step back and think a little
bit on how to structure yourself for
this new world. And I think it starts
with a mind uh mindset shift because AI
in some sense is a little bit of a
horizontal general purpose technology
that cuts across roles, cuts across
functions. Uh so you want to think end
toend workflows, end-to-end process,
end-to-end task trajectories, right?
That mindset is slightly different
because we've built up all of what we
have done in our institution in our
organizations over the decades on a
particular workflow. In fact, the last
time you could say the workflow really
changed especially in the enterprise or
in the in the at the work setting uh was
when the PCs became standard issue.
Right? Sometimes I always describe this.
Imagine in the 80s uh you were a company
trying to do a forecast. Uh how would
you do it? Right? you would sort of have
faxes going around and you would have
inter office memos and then uh and then
ultimately there would be some kind of a
paper forecast and then suddenly email
became standard issues. Excel was there
and you would send an Excel spreadsheet
and email and people would enter numbers
and you would have a forecast, right?
They just changed the work the workflow
and the work artifact all
simultaneously. And that's what I mean
by really rethinking how the work
workflow and the work artifact change
and the mindset shift. So that means you
need to be able to think about new
skills that you need uh to do that new
tools to do that just like you know
standard issue in the past uh was all
these office tools right I mean I one of
the other things I think about is if
somebody had come to us in the 80s and
said you know three billion people are
going to get up or four billion people
are going to get up every morning and
start typing and we would have said why
like why would we need so many typists
uh except That's what we do and we
describe it as knowledge work. uh right
and that's what now the new skill set
with the new tools is going to allow us
to do and one thing that I think we all
are going to learn is how to measure the
quality of the outputs we create right
so this eval is something that AI people
talk about right whenever you're
building a new model what you're doing
is hill climbing on an eval or an
evaluation criteria and all of us so
this is not just for labs or the hypers
scalers or product makers but it's for
every one of us in some sense to have
that high taste high bar for what is the
evaluation of what we are creating as an
organization and to get clear about it
and then use AI to transform that eval
to a new frontier or hill climb. So
that's a bit of how we think uh one has
to think about AI transformation.
And to that end, what we're doing is
building a full stack. And I'm going to
spend the rest of my time uh today just
walking through uh this stack. Um but
before I do that, I wanted to mention
that perhaps it's important to think
about two very important criteria.
uh when you think about AI since it's a
learning system of course it means it's
about intelligence and the intelligence
that you as a company you as a person
are going to build with AI right so
that's clearly uh going to be the case
but one of the things that you want to
make sure is top of mind for all of us
is that it's our intelligence whether
it's as an individual level or it's an
organization. Right? This is not about
celebrating some intelligence that's
outside.
It is about us being able to take the
latest model but turning that into more
tacit knowledge, more knowhow inside the
company. In fact, I'll talk about
sovereignty at the very end. Of course,
everybody's talking about sovereignty at
the comp country level, but I would say
one of the most important topics is
perhaps sovereignty of a firm. Right?
After all, what is a firm? A firm is
about the tacet knowledge a company has
thanks to the human capital that is
there uh and even the digital systems.
And so in the AI age, you want to think
even more carefully that you are
building that tacet knowledge and even
embedding that tacet knowledge and
perhaps a set of weights [snorts] that
you control, a set of agents or models
that you control. Right? So it's not
about celebrating anybody else's model.
It is about your model that you have in
your firm that makes you a unique
participant in the economy. Right?
That's at the end of the day uh the real
currency for all of us and that comes
with this second consideration that
means trust. Trust everywhere. There's a
technical aspect to it, right? For
example, how do you make sure that
there's real governance around your
data? Uh how do you make sure that uh
you have security around anything that's
happening which agents are accessing. Uh
but the biggest trust is going to be a
trust in ensuring that your enterprise
value that you have built over the
decades inside of your firm doesn't leak
out uh to somebody else's uh model or or
somebody else's business and that I
think is going to be the two things that
are going to become very very important
as you think through your AI
transformation journey. So with that as
a bit of a backdrop I want to start with
uh the first uh part of our sort of AI
stack at the experience layer. In fact
if you look at the best way for us to
introduce these powerful new AI tools is
to put them where you already are
working and then transform that work and
that workflow. Right? So that's the
approach we have taken with all of this
co-pilot layer. I describe copilot as
the UI for AI. Um, it's about really
taking the human and the human agency
and the human ambition
and really amplifying it uh by stitching
in AI uh and then letting you then use
AI to do things that you never imagined
even that you could do before, right? Um
in fact the the way I think about
sometimes is we've had these very
powerful metaphors uh that have been
there with computing right Jobs had this
of the bicycle for the mind which I I
think was a very powerful one. Bill in
the mid '9s came up with this metaphor
where he talks about he talked about
information at your fingertips. Uh that
was another very sort of approachable
way to think about computers especially
with the internet. uh and now I feel
like we're at that a you know the
beginning of a new era where the
metaphor could be that we have infinite
minds with all these agents that each of
us can use to shape what we're doing the
artifacts we create the work we do and
so on and so that's what we are trying
to get done whether it's for you know
information work whether it's in coding
whether it's in security whether it's
even in science and so that's sort of
the co-pilot so If you take information
work, what we've done is built this
co-pilot layer right into Outlook into
teams into word, excel, powerpoint and
also built the copilot app, right?
Effectively think of it as a new module
even and this is a graph that's
connected. It's stateful. So that means
you can start one place and go to
another place and that intelligence
layer continues to compound build uh so
that it helps you with what you're going
to do next. Uh right it doesn't stop in
any one app. It's not captive. In fact
that's perhaps the biggest sort of
misnomer. It's not about any one
application. It is intelligence that is
sort of ambient across all applications.
Right? In fact, you can even think about
like one of the metaphors we're playing
with is think of it as a composer that
sits above all apps uh right and it's
there everywhere and that's really
literally what at least copilot uh is in
its first version of it right so you can
imagine uh where the the direction of
travel there is so and and this these
agentic workflows are already there
right we started with chat right the
first couple of years of uh copilot was
all about chat chat chat but Now
increasingly if I look at even my own
daily usage it's got many agents that
I'm using all the time and the most
popular one that at least all of us I
think have gotten used to in the last
what 12 months or so is researcher. So
that ambient video shows you I go t you
know essentially assign a research task
in this case I think it's some kind of a
product strategy uh document I want the
researcher to create it goes off and by
the way it's not just doing reasoning
over all the data but it's also using
the computer in fact it has a computer
use agent um that it's using as a tool
uh in order to complete the task I
assigned it right so that's kind of an
example uh of researcher we built
Another agent called analyst, right?
This is like having a data science
analyst with you 24 by7. Um, in fact,
this one I think is doing um some kind
of a campaign that I'm u analyzing and I
load up a bunch of spreadsheets and it
goes and does all of the analysis that
is needed and gives me insights uh like
a data scientist would do. Uh and the
third one is uh the Excel uh agent. Uh
so we've not only uh built these agents
in copilot but we have also built agent
mode into word, excel, powerpoint so
that you can create these new artifacts.
In fact you know artifact creation
has become just easier because you can
single shot it. But the more interesting
thing is what you do with the artifact
by using this agent mode and iterating
on it continuously refining it. Right?
when I talked about that eval right you
want like like a great DCF uh you want a
spreadsheet that does a fantastic cash
flow uh or what have you and you can get
that out and then iterate on it refine
on uh refine it and that's sort of the
new way to think about even the workflow
because if you you you know previously
spent a lot of the time building the
spreadsheet you now spend a lot more
time doing the analysis with the
spreadsheet uh with agent mode and so
this is the transformation information.
And to show you some of this at work, I
wanted to invite up on stage uh my
colleague Anne. Anne, take it away.
[applause]
>> Thank you, Satia. Hi, everyone. I'm so
excited to show you how the entire
Copilot stack is transforming the way
people work. I start my day in Copilot,
and the first thing I'm going to do is
look at all of these models that I have
to choose from. I'm going to pick GPT
5.2 into quick response because I've got
a live demo and a lot to show you. Then
I'll ask Co-Pilot to prep me for my day.
Co-pilot uses Work IQ to look at my
meetings, emails, and chats. And it
looks at its previous history of
interaction with me to know what I
prioritize the most and come out with a
customized tailored response. So, let's
see what's going on today. First up, I
have a time block for this demo here.
So, I'm glad I'm in the right place with
you all. And then if we keep going, we
see at the top of my inbox, I have an
ongoing customer conversation about a
pending contract. And part of my job as
an accounts receivable manager is to
make sure our customers pay their bills
on time. And to incentivize them to do
so, I need to offer potentially a cash
discount for early payments. And again,
I have lots of options I can offer for
such a cash discount, but I want to use
co-pilot to make sure that I'm following
a policy compliant approach. So, I'll
ask Copilot, what are my options to
offer this customer a discount? Now,
this is where the power of work IQ
really shines. Copilot's looking at our
policy documentation. It's looking at
the context and history of that customer
and this particular contract. And it's
applying my role as an accounts
receivable manager to drive what it is
I'm specifically entitled to author. And
here we go. We see how all of these
things come through in the response.
First up, I can see that we did ini
indeed find that policy documentation. I
can verify that the source is our
finance SharePoint. So, we're accurately
pulling those details. And then if I
look through copilot's given me this
great summary of all the different
options, I can provide anything that
requires additional approval beyond my
manager level remit. And it ultimately
summarizes we should start with our
standard 2% cash discount for any
payments made within 10 days of invoice.
Now, that makes sense to me, but before
I go off and add this to the contract, I
need to assess the impact to our
profitability. And what better place to
do that than Excel? Here in Excel, I
have a workbook that I've already built
with agent mode that models the revenue
on the contract as it currently is
stated. And I want to add this nuance of
how a discount impacts it. Right from
here, I'll go to co-pilot and turn on
agent mode. So, I don't have to do all
this work manually. I'll have the agent
take care of it for me. I'll ask it to
create a new sheet that models the
impact of offering that 210 cash
discount. And because the agent's
grounded in the same work context I am,
we'll give it access to that policy
documentation file. And it can
extrapolate additional details from that
context. Then we'll kick it off. What
the agent is going to do here is work
directly inside my workbook just like I
might or a co-orker might. Unlike us, it
goes quite a bit faster, but it still
takes a little bit of time, which means
now I have time I can use for other
activities.
All these insights that are helping
drive business decision making, whether
it's in Copilot, in Outlook or Excel,
they come from our underlying financial
data estate. So for our business
insights to be valuable, our data needs
to be accurate and up-to-date. Dynamics
365 is our platform for intelligent
business applications, which enables us
to do just that. Here in Dynamics, when
I log in, I see I have a notification
from the account reconciliation agent.
That's a pre-built agent that comes out
of the box in Dynamics, and it's letting
me know there's a couple discrepancies
in our books that need my attention. So,
let's take a look. For each of our
reconciliation exceptions, I get a
natural language summary that
immediately tells me exactly what's
going on. In this case, a transaction
was posted directly to our ledger, and I
can see those posting details here, and
it missed our subleddger. And if you're
not a finance person, don't worry. Just
know this is not ideal. We need
transactions in both places. And it's
not just insight that the agent is
giving me, but it's also recommending
actions. It knows finance best
practices. It knows our best practices.
And in this case, it recommends we
reverse the transaction so we can post
it properly through the subleddger.
From there, all I have to do is confirm
and commit the transaction. And I want
to highlight what's happening here
because it's so important. This fix is
being logged directly in our financials.
So our source of truth can remain just
that. There we go. We got a confirmation
the reversal was completed. And for any
of my audit folks out there today,
everything the agent does, everything I
do gets logged in our audit trail. So we
have that critical accuracy and
accountability we need in finance
processes.
Okay, we took care of that discrepancy.
Let's go back and check on our co-worker
agent in Excel.
And what we can see already is how we
are on a brand new sheet that the agent
created and it's making edits live right
in front of us. We can even follow its
thought process as it goes step by step
in the pane over here. It looks like
it's reassessing. And because we're
doing a live demo and we've only got so
much time left, I'm going to jump ahead
to the same prompt that I ran earlier
just to show you that output of what it
looks like. So, we can see we have our
discount situation broken out. It's
pulled that detail of the 2% 10 days
from the prompt and additional context
from the document that I attached to it.
It's looked at our monthly impact to
revenue and ultimately we should have
the chart I asked for as well over here
on the side. What I want to stress here,
it's not just putting static text in
cells. These are dynamic formulas and
visuals that work with me and the agent
as we continue to work together and
refine. And most importantly over here I
have full visibility into all the data
sources, calculations and assumptions
the agent used to build out this
context. So to recap, what did we just
see? We saw how co-pilot with work IQ
helps me work not just more efficiently
but more effectively. We saw how agent
mode in Excel is helping me very quickly
strategize and make decisions that drive
profitability of a contract. And we saw
how agents in Dynamics 365 are helping
us create a financial data estate that
enables the whole business to make great
decisions. And this is not just for
finance personas. Everything I showed
you today is 100% available. I hope you
go home and try it out in Copilot, in
Excel, and in Dynamics 365. Thanks. And
back to you, Satia.
>> [applause]
>> you know, I was uh when I was seeing the
demo, I was thinking, God, I started,
you know, my career at Microsoft doing a
lot of Excel demos. I had a lot more
hair then and here I am um you know, in
2026 still looking at Excel demos. You
know, one thing that hopefully you capt,
you know, you got out of that what an
showed was, you know, the design of AI
systems, I think requires a particular
thought that it doesn't add more
cognitive load to you, right? I mean,
one of the things that could easily
happen is because you have this powerful
tool that can do things like create an
unbelievable spreadsheet or go off and
reason for hours and come back with some
answer. And if all it does is cause you
to then spend more time understanding
what it did, then that's not great
design. So one of the subtle things that
we are trying to make sure happens uh is
as we build out this co-pilot layer, how
it works with work IQ and how you
traverse uh your job and your task
trajectory, the cognitive load has to
come down. uh even the aesthetic like
one of the things that an was showing in
Excel in agent mode uh right
spreadsheets the reason why spreadsheets
have st stood the test of time uh from
whatever 40 I think we just celebrated
the 40th year of Excel uh is because it
has it communicates a lot cognitively
right the the fact that you have a
tabular data form um you can sort of
look at the formulas in fact I so I
remember the first time I sat next to
somebody you know one of these Excel ma
maestros was like sitting next to a
piano player uh because the way they
built the model uh it made sense and
when you see the output it made sense to
you as a consumer of that model uh and
that needs to be true even when it comes
to AI right so it can't be I single shot
at a model that nobody else can
understand but it has the answer and
then you're looking for hours as to how
the heck did they find the answer and so
that's the type of stuff uh that matters
in how AI gets designed into
experiences.
So we it's great to see the diffusion of
all of this technology right here in
Spain. I had a chance to meet with many
customers who are deploying you know
co-pilot uh using C-pilot. Um I had a
chance to meet this one you know uh
partner of ours uh illunion uh which is
you know doing it's a doing a fantastic
job of not only using this technology
but the social good uh you know how they
employ people with disabilities in their
workforce and empower them uh with
technology is fantastic to see they
built an agent in fact using an agent
builder inside a co-pilot uh for their
the law department to be able to reason
over complex PDFs Uh in fact that's kind
of one of the most favorite features of
copilot is man PDFs are tough to parse
uh and you can go you know assign them
to copilot and it can come back uh and
extract the right sort of uh reasoning
on top of it and give you information.
It's just great to see uh this kind of
diffusion. So that was the first layer
of the stack which is about building in
AI experience into the flow uh of human
activity and then really helping us
raise the human ambition. Uh the next
piece of the stack is the agent
platform. Um and when it comes to agent
platform the first place to start is the
data layer but more importantly not just
raw data but uplifting the semantics of
this data into what we describe as the
IQ layer right and an talked about this
right when when she was using co-pilot
she was essentially interfacing with
this thing that we describe as work IQ
uh and and and I would claim that one of
the most important databases inside of a
commercial entity is the information
underneath their Microsoft 365 tenant. I
mean think about it right inside of
Microsoft 365 you have all this very
rich information people their
relationships to other people like your
org chart uh not just org chart in its
formality but even who is working with
whom on what projects right I mean think
about how important that is for any
company uh you have all their calendars
you have all of their communications in
teams and outlook you have all their
documents in one drive and sharepoint
you have projects and project schedules.
So that is a very rich set of
relationship and entities uh that now
you can in fact use AI models to reason
over. That's what work IQ is. It's a
stateful reasoning agent essentially on
top of a very rich data model uh which
captures all this tacet knowledge of an
organization. The other place where
we've done in fact you all have created
a tremendous amount of semantic data
semantic information is PowerBI right I
mean if you think about the number of
dashboards over the years that have been
created someone took the pains to make
sure that all that data was annotated
with the right semantic um annotation so
that you could make sense of it right uh
this is some sales aggregate information
it is some forecast information it's
some supply chain information. Now all
that can be brought together uh in
fabric IQ. Uh and then foundry IQ is
another service we have built in to be
able to in fact have an outer
orchestrator of all these IQs.
Essentially you can have a plan that
gets created that orchestrates work IQ,
fabric IQ and any other document
database or any other thing that you
have. Uh and now you can build
essentially agents on top of all of this
IQ layer. Right? So data is the best
bootstrap but more importantly not just
raw data that you build a forward index
on or you do a search on but literally
have a reasoning layer plus data
available for any agentic system. So
that's sort of the first layer of the uh
the agent platform. Then once you have
this, the thing that you need is an app
server, right? I mean with any
generation of technology,
you always built a middle tier app
server so that you could build
applications on that technology, right?
It happened with the web, right? We had
web servers. Uh we had it with .NET and
Java. Uh we had that with cloud, right?
We had cloudnative app servers. So this
time around again for us it's foundry
where you have all the services to build
agent and multi- aent systems starting
with frameworks uh with you know
services for AI safety like guard rails
services for doing these eval services
to do fine-tuning what have you. So
these are a complete list of runtime
services that you need in order to build
agents and multi- aent systems. And on
top of that, you have models. In fact,
Foundry comes with 11,000 models. And
you would say, "Wow, why do I need
11,000 models?" In fact, it comes with
all the popular ones, right? You have
all the GPT family, you have the Claude
family, you have grot, you, you know,
you pick your, you know, frontier model
that you want to use. It's there. But it
also has in fact all the models in all
the domains that you want to be able to
access and use as part of your
applications. And in fact just to kind
of give you an intuition for why the
world needs more models not just one
model it, you know, we launched this
thing called gigat time I think a few
months back. Um and you know it was done
in in cooperation with Providence um and
a couple of other partners of our
University of Washington. Um and the
idea was if you look at cancer treatment
and imunotherapy, one of the challenges
is you never know if a particular
imunotherapy would work on a particular
tumor. And so in order to figure that
out, there's a test that one does. And
that test, you know, costs a lot of
money, takes a lot of time. And so the
researchers put together essentially a
spatial proteomics um solution. So
essentially using generative AI to
simulate that test, right? So instead of
doing the test, what if we could do a
simulation of that test and that way
then you can reduce the cost of it all
and make it available to every secondary
hospital in every uh country so that
they can now do that same test and you
know make a difference in oncology. So
to give you a flavor for this, let's
roll the video.
Imunotherapy is really the most
promising direction for taming cancer
once and for all. The whole idea is
using the immune system to fight cancer.
>> The big challenges we have in the field
right now is to understand which
patients will respond.
>> Facial proteomics can simultaneously
measure multiple protein and create
multiplex imunofllororescent images. It
can tell you whether the patient might
respond to imunotherapy or not. But
unfortunately to generate this kind of
images takes days and cost thousands
[music] of dollars.
In collaboration with Providence and
University of Washington, we developed
this multimodel AI called gigat time to
generate spatial proteomics at scale. So
we have trained the model using large
scale data analyzing millions of cells.
The input image has limited data. We
don't know where the protein is active.
Gigat time can convert [music] this
image to a colorful multiplex imunosence
image. The color is the protein
activation. They are the key factor
drive diseases especially in cancer.
This can help us to predict patient drug
response. We now taking the first
meaningful step to be able to generate
this kind of data for everyone in the
globe.
The gigat time model is now available to
everyone through Microsoft Foundry. Our
hope is to [music] empower every
secondary hospital in every tertiary
city to be able to now conduct their own
research.
>> We look forward to a future where we can
actually bring this to the clinic,
impact patients lives and bring the next
generation of therapies to places where
this would be otherwise impossible.
You know, it's a great example of um how
you translate sort of unique knowledge
that a institution in this case the
combination of quite frankly Providence
and University of Washington had into
something that now of course is going to
be a powerful tool and a powerful model
for the rest of the world to use. Uh but
when I talked earlier about that
intelligence in every company, that's
what it is. In fact, if some someone
asks me like how many models should
there be in the world, I would say the
number of the companies out there are in
the world because after all, every
company is effectively a model of some
unique tacet knowledge. Uh sometimes
it's proprietary. You don't want to put
it in foundry, but it'll be in your
company. But sometimes it may make sense
for you to then expose it to even the
world. And that to me is uh how to think
about models. is not about just one
model but it's about sort of the models
all of us build uh that truly increase
not only the intelligence in your
company but the collective intelligence.
So now we talked about two things,
right? We talked about that IQ layer. We
talked about the app server with
Foundry. And now it's time to build uh
agents. And to do that, we have a a full
tool chain, right? Starting with App
Builder. Uh for those of you who used
things like Power Apps before, App
Builder is the AI sort of era version of
it where I can now go in and just use
natural language to essentially build a
fullscale app. Um and in this case, I
think it's sort of um an app that was
builting building for a person a
personalized career path uh application.
Um and so they were able to just simply
use natural language to build a full
stack application that was generated. Um
a second uh solution we have is copilot
studio right so copilot studio allows
you to build agents uh so you can again
uh take any workflow and make it agentic
by grounding it in some knowledge uh and
again using uh this builder tool uh to
build a pretty sophisticated um
application for employee onboarding for
example um and so this is a second tool
that's available uh for us and of course
if you're a professional developer We
now have GitHub and Visual Studio and uh
you're able to use the sophestication of
those tools to build agent and
multi-agent systems. Uh in fact, it's
fantastic to see the growth of GitHub as
a community in uh uh is Spain. We have
two and a half million developers on
GitHub. It's grown 22% over year. Um and
it's just unbelievable like what's
happening uh in coding and the number of
repos that are getting created in public
and of course a lot more in private. uh
just given the power of these coding
agents and so uh it's great to see even
uh Spain participate in it. So one of
the core goals we have with GitHub is
not just to have your repo or your CI/CD
pipelines and all the other core
primitives of GitHub but it's also
becomes the organizing layer for all of
the coding agents. So we're designing in
uh right into GitHub uh you know
whatever your favorite coding agent is
is available right there. In fact, you
can go to your mobile app in the morning
and you essentially have a new inbox uh
which is your coding agents and their
work uh and you can go and keep track of
what's happening on your repo uh you
know by these coding agents. Uh so
that's what we describe as agent HQ. Now
the other thing that we've also sort of
um made sure is we build all the
different form factors right because if
you look at how coding has changed uh
the IDE still remains a very important
one in fact if you look back uh the
evolution the first thing where we used
AI was in the next that it suggests
right so you're in an IDE it's doing
essentially uh completions uh then we
put chat uh so that you could stay in
flow and you didn't have to go outside
to get answers and you could continue to
code. Uh then we had agent mode where
you could assign the agent a set of
tasks. Um and it could complete that. Uh
and then you have these fully autonomous
agents as well. Now so in fact putting
all of these form factors together,
right? The CLI or the command line
matters, the IDE matters, the autonomous
agent matters. The other thing we are
also making sure we do is all the models
are available to you. Right? It's not
it's multi-form factor multimodel uh and
in fact many many sub aents uh that you
may build for a variety of reasons for
profiling for security all are part of
the GitHub copilot ecosystem and to show
you all of this in action I wanted to
invite up on stage my colleague Marlene
Marlene take it away
[applause]
>> thanks Aia and hi everyone let's say I'm
a developer working at a toy company
called Tailspin Toys. A few days ago, I
got an email from the search product
management team asking me to add some
new search and filter features to the
site. They've given me a short checklist
of what I should build, including a
search bar with text search for simple
searches and Azure search for more
complex ones. They've also asked me to
add in a sidebar so customers can filter
by both category and price. I'd like for
Copilot to help me with this task. And
so to do that, I'm going to head to my
terminal
and I'm going to run the command to get
GitHub Copilot CLI started. With Copilot
CLI, I can now kick off any agentic task
directly in my terminal.
For C-pilot to know how it should help
me with my task, it needs to somehow be
able to get the information from the
email we got from the PM team and bring
that information into the terminal.
Earlier, Satia talked about work IQ,
which allows us to connect to the M365
suite and to bring that information here
into the CLI so Copilot can work with
it. You'll see at the top here. It's
doing a lot of thinking, but you'll see
that Copilot has already connected to
the work IQ MCP server. In my prompt,
I've asked Copilot to look for the PM
team's email and then to return a list
of the things we need to build based off
of that email. You'll also see that
copilot has invoked the work IQ skill
which under the hood uses the graph API
to be able to triage through my emails.
It's going to look through and hopefully
if the demo gods are on my side this
morning, it's going to find the email
that the PM team sent us and then we'll
list out the things we need to build.
While it's doing that, I'll also point
out that I'm currently connected to
Claude Opus 4.6. So, this is one of the
most popular coding models that we have
available, but we have a couple more
um as well. Okay, great. I can see that
Copilot has used Worth Q to find the
email and it's list out listed out the
checklist that was in the email we saw
earlier before. So, we're almost ready
to build. And like I said, I'm connected
to Claude Opus 4.6, but I'm not sure if
this is the right model for the job. So,
let's see what options we have. With a
single GitHub copilot subscription, I
now have access to all the latest
models, including like I can see here,
GPT 5.3 codecs. Now, I'm not sure how
many of you are chronically on social
media online like I am, but if you are,
you might have already heard about this
model. It just was released a few days
ago and it's already available for us to
use in GitHub Copilot today. So, this is
super exciting. I'm going to choose this
model and then I'm going to send through
and then I'm going to send through a new
prompt to be able to ask co-pilots to go
ahead and get started with building that
search and filter feature.
You'll see that the model begins by
thinking. It's going to analyze both my
prompt and the codebase around it. And
after that, it's going to create a plan
and then implement on that plan. This
process is going to take a few minutes.
And so while Copilot works in the
background, let's switch over to a new
terminal. Earlier today, I gave Copilot
the same prompt asking you to build that
search and filter feature, but I asked
it to add the changes to a PR. You'll
see that we are also connected to the
Playright MCP server. I'd like for
Copilot to use this MCP server to test
the search bar and the filter just like
a human user would.
Playright is an open-source testing
framework built by Microsoft that allows
us to be able to automate browser
interactions. So to be able to test an
app, Playright will begin by opening up
a browser and it will start running an
automated script. It's going to start
typing in inputs, navigating to
different pages, and all along the way
it's going to take screenshots. You can
see my hands are not on the keyboard.
And this is all co-pilot and playright
doing this work for us and testing that
our features are working as expected. We
just saw Copilot literally working
through four different functional tests
of our application. We can see that our
search bar and our filters are working
as expected. And this is perfect. Our
tests have passed. Now that we know that
the app is working, we're almost ready
to merge the PR. But before we do, let's
add in some documentation and head over
to VS Code for that.
For developers, good documentation has
always been important, but I would say
it's even more so now that we can
generate code faster than ever with AI.
To be able to automate the process of
creating better documentation, I've
created a PR visualizer app and I've
used the GitHub Copilot SDK to do this.
Copilot SDK gives developers the power
of GitHub copilot directly in their
apps. The Python code that you can see
on the screen creates a co-pilot agent
session. And in this session, I'm
passing through this PR data tool set.
And that set will be used to pull in
data from the PR we saw uh in the
terminal. And then we'll pass that data,
the key points to Excaliraw, which we've
connected to Copilot using MCP. And then
Excaliraw is going to generate a diagram
for us that will show us which changes
have been made to our code in that PR.
So I have the app running locally in
simple browser and I'm going to pass
through the link to the PR that I got
from Copilot and it's going to start off
by calling that tool we passed through
in the code and we'll use that tool to
pull in the information and then it's
going to combine that with the excal
draw format to understand how to then
generate a diagram like you can see on
the screen there to show us what has
been added to our code. So we can see
which parts the search bar the filter
have been added to the code and which
parts of our code base have been
modified as well. So in the past few
minutes we've seen Copilot with work IQ
pull in a spec from an email and build
it. We've also seen Copilot work with
Playright to be able to test an app
functionally. And finally we've just
seen Copilot SDK with MCP generate an
architecture diagram for us. So whether
you're working in the CLI, in the IDE,
or adding AI to your applications, give
GitHub copilot a try today. Over to you,
Charles. Thanks.
>> No, thanks. Thanks, Marlene. And um you
know, that kind of gives you a flavor
for
um you know, what's possible uh with
GitHub Copilot. And if you look at the,
you know, who would have thought again,
you know, you're back at the terminal
doing all these powerful things. And
that's because you can through natural
language get your tasks done, right? I
mean, you're not just writing code, but
you're able to literally reference a
team's meeting, your email discussions,
and have work IQ surface the plan that
you then go execute through the coding
agent and change your repo. I mean
that's the power uh of some of these AI
capabilities all compounding. Uh so work
IQ doesn't just remain in uh Office 365
or Microsoft 365. It now is available to
you even when uh you're coding and that
I think is a great example. The other
thing that Marlene showed which I think
is another very powerful thing for
developers is the entire platform or the
runtime underneath GitHub copilot is now
available for anybody to build their own
agentic system. Uh so you can now take
that access to multiple models. Right?
So when she did the PR visualizer that
was built using that SDK uh where
literally you have access to all the
models that GitHub copilot has. So your
app has that the skills uh and its
ability to call MCP servers. It has
that. So that way you can now build
applications that are you know maybe in
core banking or in insurance or in
healthcare that literally are also going
to be like GitHub copilot with variety
of different form factors multiple
models and access to all the MCP servers
and skills. So that's just a great
example. So that was the stanza where we
now said we can you can build you can
build all these agents and multi- aent
systems. In fact you can take what is
underneath even GitHub copilot and build
your own version of it. Now if you have
all these agents, the next big issue
that one has to deal with is how do you
govern them? And really to ensure that
you have a robust solution uh for both
observability and governance of AI
agents, we built agent 365. And to talk
about this very important initiative, uh
I wanted to introduce up on stage my
colleague Charles Lammana. Charles, come
on up.
>> [applause]
>> Thanks.
And like Satia said, we're going to have
a lot of agents moving forward. What
we're going to see is a need to have a
control plane to go manage and sit on
top of all of these different agents.
That way, we're going to be able to get
the right level of visibility and
understanding about what people are
creating and how those agents are
interacting with other people, with
data, and all the different systems
inside the enterprise. And this control
plane really enables three core things
across all of your agents. First, it
helps you with governance. You can
create policies and workflows so you can
manage exactly who creates your agents,
who consumes them, how they're shared,
and what systems they can work with.
Second, you can have perfect
observability. So, you can understand
what these agents are doing on a regular
basis. That way, if there's a risk or a
potential security problem or just
errors or cost issues, you can start to
have an understanding across all of your
different agents. And last but not
least, you can enforce security
policies. So no matter what these agents
are doing in terms of connecting or
authenticating or authorizing or even
having real-time protection for things
like jailbreaks and XPIA, you're able to
protect these agents across your entire
estate. And as all of this starts to
come together, we need to have a really
integrated experience. So, I'm going to
walk you through some of the new Agent
365 capabilities that we've been
releasing.
If we look at what Agent 365 enables you
to do is it's a one-stop shop to get a
view of all of those agents that are
created, it's available right inside the
Microsoft Admin Center. So the same
place where you already go today to do
things like manage your users or devices
or even your co-pilot rollout, you're
now able to see all of the different
agents that are active. And if I look at
some of this top level dashboard view,
you can get an idea of how many agents
there are. In this example, 128,000, as
well as how many users are interacting
with those agents, in this case, over
58,000. And we chose these numbers for
the demo to really highlight that ratio
that we expect. There are going to be a
lot of agents and every user is going to
need to interact with many agents to get
their job done most efficiently. Whether
it's inside of GitHub copilot or agents
that your IT team or dev teams have
created or even from third parties. And
that's one of the important parts of
agent 365 is it's not just for Microsoft
platforms or Microsoft agents. We of
course have registrations for Copilot
Studio and Foundry, but if your agents
are built and integrated with Workday or
Google or Service Now or Manis, you're
able to see these agents and what
they're doing. And you can track things
like trends as well as potential risks
as they show up. And the risk part is an
important item because if you have over
a 100,000 agents, you don't want to
click through each one of them to
understand what the problems are. So,
we're able to give you the top actions
as an admin. So, you can go and manage
things like which agents require review
or approval, where you have those
potential problems or agents that have
been orphaned and no longer have an
owner. All available inside that
console. And if I say click through into
agent risk, you can see here are some of
them that that have been identified. And
if I look at the comm's agent, I can
start to get a quick view of what its
problems are. I can see there are two
security risks that require my
attention. And if I go and look at it,
you can start to see that the agents are
integrating with the same systems we
already use to secure those users and
those devices. In this case, it uses
Entra to let me know that the agent has
an abnormal sign-in frequency. So that's
based on things like location, IP range,
how often it's logging in, it's been
flagged for further review. Or one of
the reasons you want to have your agents
and your users managed in the same
system is we can look and identify a
risky user who may have been compromised
is interacting with this agent in a way
that isn't normal. This can help us
elevate and flag across people and
agents and devices where that risk is
starting to show up. And in addition to
looking at the agents, you have to have
a process by which these agents get
deployed. So we have an experience where
you can approve each agent as it's
getting set up. So if I look at this
staffing agent which someone has
requested me to go provision, I can
start to publish it out to users and I
can manage do I want to do it to
everybody or do I want to do it to
specific users or groups? I can manage
who has access to it and importantly I
have this idea of a template and these
templates have built-in policies around
what the agent can do. So I can just
choose for example the staffing access
governance template to make sure the
agent's going to behave the way I expect
it to. And if I look here, I can also
review the permissions and the access
that this agent is going to take
advantage of. So I can very quickly
click through and get a good idea of
what these agents are doing across my
organization and start to scale that
agent practice. But it's not just about
the agents, it's also about the
intelligence, the systems it connects
to. So if I look here inside of the
agent 365 experience, we have an IQ
page. And this is important because the
thing that makes you unique is your
data, not the model. And here I can
understand things like how many emails
or how much emails or meeting contexts
are being used by agents on a regular
basis. And this lets you know how that
IQ is being put to work to go drive that
business outcome and business process
value. And this goes across all of the
important work IQ data and more. So what
agent 365 really starts to enable is
this idea that you can empower users to
create agents. You can have IT teams
that build their agents and you can also
work with third parties and other
companies to get those agents deployed
at work. This creates an incredible
agent ecosystem. We've already worked
with dozens of other companies like SAP,
Service Now, Workday, as well as
startups like Genpark and Manis or even
the AI labs like Anthropic and OpenAI to
integrate with agent 365. That way, as
you get these things deployed, you can
make sure you have a single simple way
to govern and control them across the
entire company. So, with that, I hope
you give a 365 a try. It's available in
the Microsoft admin center today and
I'll hand it back over to you Satia.
[applause]
>> Thanks.
>> Yeah, that it's it's a very important
sort of aspect I think of how the roll
out of these agents will happen. I mean
that thing that Charles showed the ratio
of people to agents by definition is
going to be enormous. uh it may be what
2:1 3:1 who knows uh but you really want
to have this governance system uh this
observability system and the assurance
of security so that you can confidently
deploy these agents inside of the
enterprise. You know the good news in
all of this again is there's fantastic
traction in building out these agentic
systems right here in Spain. And I had
this morning a chance to meet uh you
know a bunch of developers. Uh I had a
chance to meet the developers at
Factorial who are building out a full HR
agentic system. Uh it's fantastic to see
the success they're having uh their
design uh of how they're building the
system out. Um also had a chance to meet
with Ferovial. In fact obviously they're
some of the largest infrastructure
builders. They're building in fact our
Madrid data center expansion as well.
Um, and they were showing me how they've
built an agentic system that literally
allows you to effectively apply uh,
intelligence in real time to any project
of theirs, whether it's a data center
buildout, a highway, an airport. Um, had
a chance to meet with the Repol folks
who are deploying um, uh, these IT
workflows. Uh in fact they showed me an
agent that was doing automatic conver I
mean basically was able to connect a
front end using ABBA to their SAP and
doing the automation uh to of a workflow
and now they're scaling it beyond it. um
had a chance to meet with uh Sanitas
which is really built a voice system um
the which is a front end to how their co
customers are interfacing uh with uh
their uh their entire sort of mobile
application. Uh and so it's great to see
uh the level of innovation and the
diffusion of all of these technologies
uh and the developers right here in
Spain uh not only building for Spain but
for the world uh and uh and it just
showcases the power uh of the human
capital uh here and then the and in
their hands these tools get translated
into an unbelievable uh value. The last
sort of part of the talk is going to be
about the token factory. So, we talked
about AI um the experience layer, we
talked about the agent platform, and of
course, if you're going to have lots of
agents, you need infrastructure that
generates tokens most efficiently. Um
and it's all about this particular
formula, right? It's about performance
uh of tokens per dollar per watt. Um and
the goal is how do you really make sure
um uh you have the most efficient system
for it? and we're building out Azure as
the world's computer. Um we have uh 70
plus data center regions, 400 plus data
centers. We continuing to expand that.
Uh right here in Spain, we have two
regions uh both in central and north. Uh
the other thing that we have remained
very committed to is renewable. Uh and
we are 100% in fact we've achieved that
milestone where all of our data centers
have 100% of renewable power. In fact,
you can even think about our spend in
capital and our procurement and PPAs are
a great stimulant uh for essentially any
renewable financing uh and project
financing. And so it's great to see this
uh continue to be a focus for us and
everyone else who are our partners
there. The other aspect uh is the the
systems themselves. In fact, we are at a
fascinating time, right? We are at a
time where you are starting to see the
workloads somewhat get stable, right?
These atcale training workloads or
inference workloads. They're unlike
anything we've seen in the past, right?
These are large sync data parallel
synchronous um you know GPU workloads.
Uh but you're now learning that for
example take even an inference workload
that you can optimize it. uh you can
disagregate the prefill and the decode.
um you could in fact take even the how
you access memory whether it's high
bandwidth memory or SRAMM how do you
think about uh the dist like you can
have a draft model completely in SRAMM
so there's a complete renaissance of
systems in fact so you know starting in
December to now there has been a
renaissance uh in how you think about
even the systems for inference uh and so
the Microsoft approach is going to be to
have a a fleet that takes advantage of
the best technology and uses the power
of software to make it available to so
that you can have the best token
factory. Right? We're going to partner
with Nvidia. We're going to partner with
AMD. We're building our own systems. In
fact, Maya 200, which we just disclosed,
which is taking real traffic today, uh
is around 30 plus% better than any
leading node in any leading system uh in
Azure today. So we're going to continue
to innovate uh and really take advantage
of the ecosystem innovation so that we
can build out the best token factories.
Now we in addition to doing all this
work around building out the data
centers making sure that they're you
know got renewable sources of energy
innovation in the system side and the
silicon side. Uh the other aspect is
making sure we're committed uh to
Europe. Um and we made in April of last
year, we had these European commitments
uh around strengthening the digital
ecosystems here with local partnerships
and local innovation, privacy, security
and continuity uh and digital
resilience. Uh and these are commitments
we stand by and we've continued to work
and one of the aspects of this is
sovereignty. uh after all sovereignty is
a top consideration
uh for any country and uh in order to
address that what we've done is really
built a portfolio of solutions. In fact,
if you think about sovereignty as
essentially managing even risk uh the
way you manage risk in any other domain
is to have a portfolio. Um and that is
essentially what you should even think
of when it comes to sovereignty which is
some workloads using sovereign services
in public cloud. You can for example
encrypt all your data. You can manage
the keys that do the encryption. You can
use confidential computing uh in order
to ensure that the data that's being
used even in a public cloud is never
outside a confidential chamber. Right?
So effectively it is fully sovereign. Uh
so that's sort of one solution. On top
of that, we have a solution which is
fully local. Um and so when it comes to
uh we already have Microsoft 365 uh
local, we have Azure local and we're
very thrilled this week to announce
Foundry local. Right? So when we talked
about Foundry with all its models,
they'll be available now even locally uh
for you to be able to use in a private
uh setting. Uh and not only that, we
were also thrilled this week to announce
that we've taken Azure local which up to
now had a control plane uh which was in
the cloud but even the local with local
operations but now you can have fully
disconnected uh um control plane. So
that means you have a local cloud with a
local operation control plane versus
being in any hybrid uh mode uh at all.
And so these are all again um options
for you to think about the different
workloads and where you would like to
put them in order to have the best uh
efficient use of the cloud resources and
we also have many sovereign partners
locally uh who are helping us ensure
that all these portfolio of solutions
are available uh to all of you. The one
aspect I wanted to uh talk about though
at the very end of it is as you think
about sovereignty and risk, it's also
important especially for policy makers
to think about even cyber and cyber
security and cyber resilience because if
you think about cyber uh it's
fundamentally an intelligence game,
right? It's a global intelligence game,
right? you're only as resilient on cyber
to be able to sort of make tap into the
full intelligence the world has to
protect yourself. Uh because no cyber
attack respects borders. Uh and that I
think is a very important thing. So when
you think about a sovereign service or a
sovereign cloud and if for what whatever
reason it doesn't have the best cyber
protection that sovereignty
unfortunately could get compromised and
so therefore you really want to think
about sovereignty and cyber as two parts
of the risk uh spectrum that you want to
manage uh simultaneously and that's what
uh we have a commitment to ensuring we
are partnering with in fact local
providers uh in bringing that global
intelligence to all of these portfolio
solutions. Um, so I want to close out
where I started. Ultimately, to us, it's
about taking all of what we talked about
this power of this new technology, but
putting it in your hands so that you can
start to transform your organizations,
uh, your communities and all of Spain,
uh, through the power of these
technologies and really your ingenuity.
So, thank you all very very much for
your partnership and I look forward to
sort of seeing what you all build going
forward. Thank you and have a great day.
[applause]