- 5 months ago
For years, AI has been framed as a productivity tool. But a new generation of enterprise agents is beginning to change the relationship entirely. These systems are no longer just assisting individuals, but executing workflows, retrieving context, coordinating tasks, and increasingly operating like members of a team. In this conversation, Dust co-founder and CTO Stanislas Polu explores the emergence of AI coworkers and what they could fundamentally change about the structure of work itself. What happens when AI becomes embedded into everyday collaboration rather than isolated tasks? Which roles are most exposed to this transition? And as AI agents become more capable, what remains uniquely human inside the workplace?
Category
🤖
TechTranscript
00:11To be continued...
00:30Good afternoon, everyone. I'm Helen O'Reilly, Deputy Editor of the French Tech Journal.
00:35For a few years now, AI has helped people to do their jobs faster, but things are moving fast.
00:43Instead of just helping with isolated tasks, today, a new generation of AI agents are beginning to do whole parts
00:52of the job themselves.
00:53So what happens when your next co-worker isn't human?
00:57To discuss that, I'm delighted to be joined by Stanislas Pallou, co-founder and CTO of Dust.
01:05So Stan, to get the ball rolling, why don't you tell us a little bit about yourself and what you're
01:09doing at Dust?
01:10Sure, thank you. Thank you for having me and thank you for that chat.
01:14So Dust basically sits at the product layer and really helps companies deploy agents.
01:21We're really the scene product layer that helps the collaboration between humans and the AI models.
01:28These days, these things are called a harness.
01:32There are many types of harness and we really see ourselves as the company harness.
01:37That seems that product that really helps that collaboration between human intelligence and artificial intelligence.
01:43Okay, but enterprise AI is becoming an incredibly crowded space.
01:49So concretely, what specific problem were you trying to solve when you built it?
01:54Yeah, so I think very different from when we started and where we are right now.
02:01When I started working on Dust, I spent some time at OpenAI before.
02:06It was pre-ChatGPD and it was kind of three years and six months ago.
02:12It's kind of mind-blowing to think about how little we knew about what would be doable with that technology,
02:19even if we were already building it.
02:21But today, the idea of Dust is really to really help for what we call multiplayer AI.
02:30And what we mean by multiplayer AI is that as the agents are taking longer and longer time horizon tasks,
02:38it will be the case that you will no longer be steering the agent yourself.
02:43You'll be much less collaborating with an agent on your own.
02:47But rather, if the agent is in charge of a task that is supposed to take multiple weeks or multiple
02:53days of human work,
02:56those tasks, because they are kind of, by essence, collaborative within the company and cross-functional,
03:03you'll have to steer those agents collaboratively.
03:06And so that's what we call multiplayer AI, that capability of having multiple people steering a set of agents
03:12to achieve a larger goal that works on a longer time horizon.
03:16Okay, so can you, to help people in the audience and myself really imagine it,
03:22can you give us a concrete example of an agent at work inside a company and what you might be
03:28doing?
03:28Yeah, sure.
03:29A good example, a kind of a toy example of multiplayer AI that is actually live in many of our
03:35users and at Dust as well,
03:37is when you think about the weekly team meeting.
03:42So the weekly team meeting, you have to prepare the contents.
03:46And generally, there's a content about finance, there's a content about the users, there's a content about sales, etc., etc.
03:52And preparing the team weekly used to be kind of a many hours job for somebody.
04:01You needed to get the frame of content ready, you needed to ping each of the area leads that were
04:07in charge of some parts of that meeting
04:10to give kind of their specific take for that week.
04:14If you have a deep dive on a subject that week, you would need to ping that person, get the
04:18slides, etc.
04:20Today at Dust, the task has been fully tackled by an agent.
04:25And so what happens is that an agent wakes up, so our team weekly is on Tuesday.
04:31An agent wakes up on Monday, knows kind of what's the common contents of each of those slides that's supposed
04:39to be every week.
04:40Like, go query Snowflake, go query UpSpot, go query GitHub to prepare the kind of default content that is supposed
04:51to be in each of those slides.
04:52Then it creates one session, which is each of the persons that are supposed to fine-tune that content and
04:59say,
04:59Hey, I've prepared your slide.
05:01It has what we generally put in the team weekly every week.
05:04What do you want to talk about this week?
05:06Do you want to change it a little bit, etc.?
05:08Each of those humans collaborate with the agent locally on each of those slides.
05:13Then the agents will say, Oh, this week we have a deep dive on that part of the product, so
05:18I need to ping that person, Daphne, as an example.
05:21And pings Daphne and say, Hey, you're supposed to do a deep dive at Team Weekly this week, so do
05:26you want to walk on the slides with me?
05:27And they start walking on a slide and it's there.
05:29And by Tuesday noon, the agent looks at all of those sessions and kind of aggregates together the presentation, maybe
05:37make a remember to some humans, and finally get the presentation ready and pings the person that is in charge
05:43of the Team Weekly to say the presentation is ready.
05:46And so the person that is in charge at the top probably checked that everything was going right, but didn't
05:52do any work.
05:53And in that case, you see, the agent has the goal of preparing the Team Weekly slides, but collaborated and
05:58was steered by at least eight different humans in the process.
06:02Okay, so it sounded a little bit like the boss was being bossed around by a bot.
06:08So that's what the technology does.
06:10But the bigger question is perhaps why we're seeing this now, because AI has been around for years, yet enterprise
06:18adoption seems to have accelerated dramatically over the past 18 months.
06:24Well, I think it's a matter of models capability that are still evolving really quickly, and there's been multiple events
06:33that have triggered a lot of adoption.
06:36The latest one was probably the release of Entropic Opus 4.5 in November.
06:41It was kind of a real step change, at least for developers, which kind of spearheaded a large adoption even
06:46beyond the developer personnel.
06:49And so why the adoption was not that great the years before, I mean, two years ago when we were
06:57in the markets, we would hear about Glean, Glean, Glean, Glean, Glean, Glean, Glean all over the place.
07:02And Glean is great at enterprise search, does a great job as Q&A on the company data, and that's
07:08what people imagine AI would do, answer questions.
07:10And as the kind of agent capabilities emerge, the ability to execute tools, the kind of standardization around MCP, the
07:19protocol that make that a reality, enable new use cases, and like, make those agents more useful.
07:26And so I think we are really in a phase where each step of the, at each time the technology
07:32evolves, the adoption follows.
07:34What we're seeing, I mean, there's been a real acceleration, because there's an acceleration, you see the acceleration in the
07:42revenue on Tropic or OpenAI, and you might wonder whether it's only developer-related acceleration or adoption.
07:49It happens that on Dust, you mostly do non-developer work, kind of all of the rest that you could
07:55do in a company.
07:56And we've seen a massive inflection in terms of token consumption per user, which has been fantastic to see.
08:04So there's really more work being done by agents all over the place in companies that, because the models are
08:12getting better, I think it's the core of it.
08:13Okay, so speaking about adoption and what you've just said, are companies buying productivity, transformation, or are they simply buying
08:22insurance against being left behind?
08:26Probably a bit of the three.
08:29I think we're still in a market where you need to provide a fair amount of transformation.
08:39And you can see that in multiple ways.
08:42There are many companies that are going to market to very large enterprise with FDs, and those FDs will build
08:48custom solutions in there.
08:49And that's a way to adapt the solution to the specific need and helps with the transformation.
08:55When our platform is kind of a very product-led and horizontal, we don't really build custom code on top
09:02of it.
09:02But we definitely need to kind of kickstart the usage within a company.
09:08We definitely need to help them find the use case, educate the champions that we call AI operators.
09:15And so adoption today is definitely tied with transformation.
09:22Okay.
09:23I'm going to take you on to the theme of sovereignty now, which is a theme that we've heard repeatedly
09:27about this week at VivaTech.
09:30And while that debate often happens at the level of governments, it ultimately plays out inside companies that are making
09:39purchasing decisions every day.
09:44So do you think that sovereignty is a real issue?
09:49Do you think that it's something that companies are really concerned about when they're making buying decisions or not really?
09:58I think it's a growing concern, but there's a lot to say on sovereignty.
10:02Let's forget about the kind of state sovereignty.
10:05We can dive into it if you want.
10:08But when you think about the input to a company or the input to an economy, but I think the
10:14two works the same way.
10:17You have the stuff, you have the energy, and you used to have the human.
10:22And the human input was mostly, at least in kind of white-collar companies, was mostly intelligence.
10:29And so your intelligence input was your own by default, because it was your human.
10:35If you're a company, it's your employees.
10:36If you're a state, it's your people.
10:39As you have an emergence of artificial intelligence, so part of that intelligence becomes an input to your company or
10:47your state.
10:49And because it's artificial, it comes from a provider.
10:52And so every way in which we've thought about energy as something that has to be sovereign,
10:59I think we should think about artificial intelligence in the same way.
11:04And so there's a big importance of sovereignty there.
11:08We do believe that the first step of sovereignty is local to the enterprise.
11:12And for us, that's the fact that they should have the choice of switching between intelligence providers freely.
11:20If you use the parallel with energy again, when you buy tokens from OpenAI and you buy codex at the
11:27same time,
11:27or when you buy tokens from Anthropic and you buy Cloud Cowork at the same time,
11:31it's really as if you were buying your energy somewhere and your machines at the same place,
11:36and the plug would only work with the energy provider.
11:41And if you wanted to switch to a different energy provider, the plug would not fit.
11:44You could be switching.
11:46That makes no sense, right?
11:48You want to buy your machines, and you want to buy your energy,
11:50and you want to put the energy in the machines, and you want to be able to switch providers freely.
11:53And so I think here it's really the same.
11:55You want your harness to be capable of routing to whatever provider you want,
12:00because models change, the external world change, some models get stopped by states,
12:06some of the models pop up, capabilities are great there, maybe not so great there.
12:10And so having that freedom of change is the first step towards sovereignty,
12:14at least at the enterprise level.
12:15And I think that argument resonates quite well with companies,
12:18and it's part of a value that does provide.
12:22Okay, so we've talked about the technology, the market, and the infrastructure,
12:27but I think ultimately what people care most about is work,
12:31and I think a lot of people are wondering what all of this means for the way they work.
12:36So you've described Dust as an orchestra conductor,
12:40so conducting and organizing information across teams,
12:44effectively replacing the need for one-to-one communication.
12:48And what impact do you think this is going to have on workplace relationships?
12:57I have no clue.
12:59I have no clue, but I guess we can take the developer job as a proxy to that,
13:05because it was kind of deeply transformed over the past six months,
13:10like deeply, deeply transformed.
13:12We used to code most of the time, use AI a bit to answer questions,
13:16and write some pieces of code.
13:19And comes November 2025, Opus 4.5,
13:23waits three months, and there's been a real phase transition,
13:27and now we code maybe 20% of the time,
13:31and the rest is written by the machine.
13:32And for the ones that have tried Fable last week,
13:40almost 100%.
13:41It felt really, within our team,
13:43we're switching pretty freely between OpenAI models
13:45and Anthropik models when it comes to code,
13:48and the kind of shared feeling was,
13:51oh, shit, it's November 2025 all over again.
13:55And so that's interesting how the work has been completely changed,
13:59and we should very quickly not code anymore at all.
14:02We'll be supervising, we'll be looking at the code,
14:04we'll be coaching the model,
14:05but there's very few lines of code we'll be writing.
14:09And so what has been the, going to your question,
14:11don't worry, I'm getting there.
14:13What has been the consequence on the job of a developer?
14:17I think it actually went opposite ways of what you may fear,
14:23where you wouldn't talk anymore.
14:24I think there's much more collaboration between developers,
14:27because we spend so much less time actually writing the code.
14:31It displays the work into deciding what to do,
14:35which is something that is closer to product,
14:37and is much more,
14:39kind of requires more collaboration,
14:43more chatting,
14:43more working together.
14:46So I think there, for developers,
14:48it shifted in that direction.
14:49And so we might think that actually,
14:53for many other jobs,
14:54you'll go much more into collaboration between humans,
14:57that's going to be the bulk of what you do,
14:58to decide what to do,
15:00and much less into doing the stuff.
15:02Okay, so you think that these agentic bots
15:05are going to actually help to strengthen communication in many ways.
15:08Okay, but still about jobs.
15:11So one area that comes up repeatedly in these conversations
15:14is the impact on junior talent.
15:16So historically,
15:17juniors learn through research,
15:20note-taking,
15:22internal coordination,
15:23and those are the very tasks
15:25that agents are starting to perform.
15:28So what will the impact be on them?
15:31Yeah, so
15:33it's a very interesting question
15:36if you think of,
15:37I mean,
15:39if you think of the bar of talents
15:41that you,
15:43basically,
15:44when the AI is doing most of the work,
15:49the importance is knowing really what to build.
15:52It's not anymore knowing how to build,
15:53it's knowing what to build.
15:55And it requires a form of talent
15:58that is hard to have without experience.
16:03And so
16:04a year ago,
16:05you would hire somebody out of school
16:07and they would be dependable and fast.
16:10That would be enough to convince you
16:13that they would bring a ton of value to the team.
16:16And it's true that today,
16:17somebody being dependable and fast in coding
16:19is almost not anymore.
16:21What you want is agency,
16:23what you want is ownership,
16:25what you want is bending the arc,
16:27all of those kind of soft skills
16:29that are a bit kind of baked into your person
16:34and but also a bit something
16:36that you learn with experience.
16:38So that's a great question
16:40and something that is quite mind-boggling.
16:44So that's for the kind of a risk side.
16:48On the brighter side,
16:50so at Dust,
16:50in France,
16:51we have high school internships,
16:54as many of you know.
16:56These are one week or two weeks internship
16:59for people that are 15 or 16 years old.
17:02And since day one,
17:04every year,
17:05we took one, two, or three high school interns.
17:08And we give them access to all the tools we have.
17:11So whether that's Dust
17:14and Codex
17:16and Codex
17:16and Cloud Code
17:17and you name it.
17:18And it's been incredible
17:19to see the slope
17:22of what they're able to do
17:23in a week or two.
17:26Like three years ago,
17:27you take a high school student,
17:29doesn't know how to code,
17:30you put it in front of the computer,
17:31you give it,
17:31this is the model,
17:33do stuff.
17:34They would kind of manage
17:36to self-teach themselves
17:38a little bit of coding.
17:40Already awesome.
17:41You don't give them any instructions.
17:43You say,
17:44try to learn JavaScript
17:44and they manage to learning on their own,
17:46asking questions to the model.
17:48That was three years ago.
17:50Two years ago,
17:51the intern comes in,
17:52I give him a paper,
17:54don't realize that he doesn't speak English,
17:56the paper is in English.
17:57He translates the paper entirely
17:59and implements the thing entirely
18:01and gets to a working implementation
18:05autonomously
18:05by asking questions to agents.
18:08He was still very well aware of the code
18:10and it was easy to see
18:11that he had a hard time
18:14understanding everything,
18:15which was worrying,
18:16but also the outcome
18:17was very exciting.
18:19One year ago,
18:22the student comes in
18:23and we give him what he want to do.
18:24He says,
18:25I want to do a game.
18:26I say,
18:26go ahead,
18:27do a game.
18:28And so he does a game for a week.
18:30We don't much,
18:31I mean,
18:32we help him a bit
18:33and at the end of the week,
18:34he presents the game to the team
18:37and he shows the game,
18:38says blah,
18:39and there was a guy walking,
18:413D guys,
18:41and we ask him,
18:42how did you create that 3D model?
18:44He's like,
18:45ah,
18:45Claude downloaded something
18:47somewhere on the internet,
18:48okay, blah, blah, blah.
18:49It was kind of exciting
18:50and we talk and we talk
18:51and at the end,
18:52we ask him,
18:53can you show us the code?
18:56And Ian says,
18:57the what?
19:01And so he walked on an artifact,
19:03which is a file of code
19:05for a week
19:06through an agent
19:08and didn't realize
19:09that that artifact
19:10has an existence.
19:12That's like creating
19:13a black box,
19:14isn't it?
19:15Yeah,
19:15but at the same time,
19:15he built a freaking game.
19:18Fine.
19:19And so the question there is,
19:22do he need to have,
19:23and does he understand
19:25what he built?
19:25Not that much.
19:27The question in the question is,
19:28will he need in the future
19:29to understand that?
19:31Maybe not.
19:32I was about to say,
19:33please say yes.
19:35Because otherwise,
19:36we're just handing over
19:37all the power to the bots.
19:39And speaking of that,
19:42you know,
19:42maybe let's finish
19:43by looking ahead.
19:44The pace of change
19:45has been extraordinary
19:46over the past two years
19:48and no one really knows
19:49where we're going.
19:50But I'm going to ask you
19:52to make a prediction
19:53anyway.
19:55Five years from now,
19:57what do you think
19:58will seem completely
20:00normal at work
20:01that still sounds
20:03completely crazy today?
20:06Five years is just
20:08not possible.
20:09I cannot predict that far.
20:11Things are going too fast.
20:12I think the image
20:14that many people use
20:15that is an interesting one
20:17is if you think about
20:19our great,
20:20great grandparents,
20:23if they look at
20:23what we do at work
20:25because of one or two
20:28industrial revolutions
20:29went through
20:30between them and us,
20:32they would look at us
20:33and say,
20:34you're not working there.
20:36You're just sitting
20:37on a chair slacking around.
20:38That's not work.
20:39And so given the speed
20:41at which it goes,
20:43assuming there is no
20:44kind of backlash
20:46or et cetera,
20:48and we can chat
20:48about that obviously,
20:50in five years,
20:51it's 100% sure
20:52we'll be looking
20:53at people walking,
20:54which will be us,
20:55and we'll be,
20:56that's not work.
21:00Where it's going to be,
21:01I cannot tell you.
21:02Looks like we're entering
21:03a brave new world.
21:08So how optimistic are you?
21:12And I guess like,
21:13I'm going back
21:14to the job question again,
21:16but it is a recurring theme.
21:17How optimistic are you
21:18that the future
21:20is going to be bright
21:21with all these AI agents
21:23and humanoids
21:24and goodness knows
21:25what else are running around?
21:28Pretty optimistic,
21:30I think.
21:30I think the times
21:33are daunting,
21:34not frightening.
21:37So really,
21:39it's daunting,
21:40the speed at which it goes.
21:43Not necessarily yet frightening.
21:47I mean,
21:48at the end of the day,
21:49we're still pretty much
21:50in control.
21:53whether you're a doomer
21:55or not,
21:55et cetera,
21:56will give you
21:57different point of views.
21:59I'm personally
22:01definitely optimistic
22:02that it will go
22:04the right way,
22:05but at the same time,
22:08it's cautiously optimistic.
22:09So I think you have to
22:10stay very well aware
22:11and try to think
22:14about that a lot.
22:15I mean,
22:17the Fable 5 ban
22:18by the US,
22:19unclear what are
22:20the underlying reasons.
22:22Is it because
22:24Antropic
22:25kind of
22:25hyped their model
22:26around cyber
22:27and so
22:28they're just
22:29getting
22:30what they ask for,
22:31which is
22:32the state saying,
22:33wow,
22:33if they're really
22:34that good at cyber,
22:35it's probably an arm
22:36or weapon,
22:37sorry.
22:38Is it because
22:39there's something
22:40political behind that
22:41or is it because
22:42the US
22:43actually think
22:44that it's
22:44some form of
22:45intelligence
22:46that is just
22:46too dangerous
22:47for the world?
22:48We don't know.
22:49But the slightly
22:51scary thing is that
22:52for the past couple
22:53years,
22:53if you look at
22:54any blog post,
22:55any podcast,
22:56any kind of
23:00predicting
23:01story about
23:02how AI
23:03can go wrong,
23:04they all start
23:05with an event
23:06that looks like
23:07last week.
23:09So that one
23:10is a little bit scary,
23:10but hopefully
23:11we'll go around
23:12and we'll find
23:13the right way.
23:15Okay, Stan,
23:15so we're coming
23:16to the end
23:16of our time now
23:18and I think
23:19I'll probably conclude
23:20by saying that
23:20it seems the real
23:21question is no longer
23:22whether our next
23:24co-worker is arriving
23:25but how we choose
23:26to work alongside them.
23:29Stanislas,
23:29thank you very much.
23:30Fascinating.
23:33And thank you.