• 5 months ago
AI is advancing rapidly across healthcare, but deployment remains far more difficult than the hype suggests. Behind nearly every promising use case lies the same challenge: fragmented data, legacy infrastructure, regulatory complexity, and systems that were never designed for AI at scale. For organizations moving beyond experimentation, the question is no longer what AI can do—but whether the underlying data can support it reliably in real-world environments. What changes when AI becomes embedded into everyday workflows across R&D, operations, manufacturing, and support functions? And how do companies build systems that are secure, interoperable, and robust enough for one of the world’s most regulated industries?

Category

🤖
Tech
Transcript
00:18Hi, everybody. Thank you very much for still being at VivaTech. I know Friday, it's hard,
00:24so you're the bravest one. Thank you very much. And thank you, both of you, for being here to talk
00:30about this drug topic. We're going to talk about AI in the health industry. So I have with me
00:36Emmanuel Frenera. You are the EVP Chief Digital Officer at Sanofi. And next to me, Love Kotari,
00:44your AI lead at Snowflake. So maybe just a first question for you, because everybody knows,
00:51of course, Sanofi, but can you just explain us what Snowflake does? Because it's very,
00:56very technical. I don't want to say a bad thing. Happy to. First of all, very excited to be here,
01:04especially with Emmanuel, who we have been partnering very closely with for many years now.
01:09Snowflake, some of you may know, some of you may not, is a data platform that's powering more than
01:1514,000 enterprises worldwide on basically the data and AI needs. The data estates are on top of
01:23running on top of Snowflake. The business knowledge, the analytical workloads are being powered by
01:29Snowflake. And name an industry, name a geography, Snowflake has a customer there.
01:35Thank you very much. So Sanofi develops, of course, life-changing medicines for patients. That's funny
01:43because we were on stage last year. Last year, yeah. Talking about not the same topic, but I'm sure
01:49you have other things to share. So I imagine the journey from the lab to getting a drug to a
01:57patient
01:57is incredibly complex, of course. So what's been one of the biggest challenges for you in that
02:03journey? And what role does data play in all of that? It's a great question. So it's not just a
02:10complex journey. It's also a very lengthy one. It takes 10 to 12 years to get a molecule to a
02:15patient.
02:17And in our industry, you get 90% failure rate. So it's not a great bet. So if you want
02:22to go to
02:23the casino with those odds, they're not great. So whatever we can do in terms of the data continuum
02:28is what's important is how to ensure that we bring the data as close as possible as we can
02:34to where we can consume it in a normalized way and then use it with AI to create simulations. So
02:41in our,
02:41I think in the future, what we're going to be doing, a lot more will be done in simulation before
02:47we get
02:47to the patient to reduce the timeframe and improve the odds of a successful drug.
02:52Okay. As enough, like you work with a lot of mathematical companies, do they have specific issues?
03:00Yeah, I think one of the most important things, especially in healthcare, is about data governance.
03:06I mean, that's something we see in highly regulated industries a lot. And that's the work that we're
03:11doing, building out that data platform that's really governed and has security built in from the
03:17get-go. That is a key piece that we're seeing, especially when you're rolling out AI at the
03:23enterprise level and creating AI transformations across the organizations. It's very important
03:28that you have a governed layer for data, and then data is being powered through AI after that.
03:36One of the things we're seeing, especially in last six to nine months with the rise of agents,
03:41is that the future of work is going to completely evolve for all of us, right? It is no longer
03:46going
03:46to be humans querying systems or pulling data from one place and putting it somewhere else.
03:52A lot of it is going to be humans managing autonomous agents, governing them, guiding them,
04:00defining the guardrails under which these agents will work. And these autonomous agents will be able
04:04to look at the past, look at the present, and then look at possible futures. And it's just that
04:09journey together. But the key piece in this is creating an environment in which humans and
04:16agents can work together. They have the shared understanding, shared context on which they can
04:20make decisions and take actions.
04:23So you're both on stage because you work together. How long have you been working together?
04:29So together, together, it's been short. But as a company, so five years ago,
04:34we made a conscious decision to move away from what we used to call having data lakes,
04:40but they were data dumps. We had very fragmented data across the company to move everything in
04:44Snowflake. So we made a bet about five and a half years ago to say we wanted to bring all
04:48the data
04:49to a place because data has gravitas. And I joked with Love and the senior leadership at Snowflake to say
04:56it was a ton of work. It was a big investment. And at the end of it, what we got
05:02were dashboards.
05:03We could see across the company for the first time because they were no longer isolated islands of
05:08data, but we had dashboards. But in 2026, when 2025, when Love and I started working together,
05:14we moved away from just dashboard to something far more profound in terms of how we exploit that data
05:20because Snowflake is no longer just a repository of data that separates compute from storage.
05:25It's also an intelligence behind the data. So that's where this is all headed.
05:30Yeah. And we have a very unique way we are partnering with Sanofi. Folks in the Valley,
05:37Silicon Valley, that's where Snowflake is, hear this word called forward deployed engineering.
05:42This is a buzzword right now in AI. And that's exactly the way we have been working with Sanofi.
05:49We have built out a team of AI forward engineers. These are AI researchers with, you know, PhDs in
05:57machine learning and AI. And then domain experts who have seen across industries how to take AI to
06:04production. And we take these engineers, practitioners, researchers, work closely in a highly embedded
06:10manner with organizations like Sanofi where our engineers, our folks are working alongside Emmanuel's
06:18team and thinking about building agent experiences across our organization. What that does is it's no
06:25longer a SaaS vendor, software vendor, and an enterprise relationship. This is an environment where not
06:33only we have access to their systems and their challenges and the restrictions regulations, but we are able to
06:40bring that AI expertise, combining with the domain expertise that Sanofi has, and then together we are able to create
06:47a
06:47transformation that is happening at an unprecedented space because of how we are bringing it together.
06:54That embedded model, Charlie, that is important. There's no distinction between Snowflake people and Sanofi people or as a company
07:03as well that we use it, it's called Elementum. There's three parties in what we build.
07:06You will not know who's who. They share the same goals, the same understanding, the same, and the same ways
07:12of working. And that's been quite exceptional because
07:15Snowflake is a pretty big tech company now. We're a fairly large company. And so big companies tend to not
07:22be so agile, but we found a way to be very agile together.
07:25Yeah. And that shared outcome between, like, you know, we're working towards the same goal, and Emmanuel's done a phenomenal
07:31job of setting the vision and then defining real outcomes, right?
07:35A lot of the times you see in AI that you're defining POCs that are not going anywhere. But being
07:40more outcome forward and being very almost maniacal about creating that outcome with AI is one of the key things
07:48that you could do.
07:48And bringing that shared understanding and shared outcome between the two teams makes you go very, very fast.
07:53So you think the forward-deployed engineers are going to be the new norm in the AI industry to work
08:02with big companies like that? That's the only way so big companies can transform themselves?
08:09That's definitely one of the ways, right? Because, as you're seeing, the pace at which AI is moving is incredible,
08:16right?
08:17And there are a few companies in the world, Snowflake is definitely one of them, that is at the frontier,
08:21gets to see what it takes to build AI and take AI to production across different industries and across different
08:29use cases.
08:30That understanding, that vast knowledge is something we want to bring to all of our customers, and we build a
08:37really compelling, robust AI platform, but there is always that last mile, and forward-deploying engineering sort of puts it
08:43all together, and then, again, back to the shared outcome, and then having that understanding of what we're trying to
08:48go towards together is the model that we're seeing a lot of success, and you'll see that shortly in, like,
08:54one of the demos that Emmanuel is going to do, how that's all coming together in a very short amount
08:58of time.
08:59Okay. So, we are in Europe, so I have to talk about regulation. Sorry, love. But how can you balance,
09:10Sanofi, global data standardization with local regulations like we have in Europe, like GDPR?
09:19You're right, we're in Europe, and we're right, we love our regulations. I think if there's a plea you have
09:24to make, it's time to simplify.
09:27We see that markets that are more competitive with AI adoption are doing it because they're probably more risk-aware
09:34and risk-tolerant.
09:36So, there's a plea to say we need to simplify things in Europe. Now, of course, we want to comply,
09:41and there's nothing more sacred than patient data for us, how we anonymize it, how we store it, how we
09:48safeguard.
09:48So, we comply with regulations because companies like Snowflake are available and hosted in Europe, so we already have the
09:56geographical premise, if you want.
09:58And then there's enough that we do in terms of data encryption, data protection, and safeguarding the access.
10:04And it all goes back to governance, who has access to what data. So, we're fairly rigid on one thing.
10:11It's great to have all your data in one space, but it's also, you could say it's a risk to
10:15have all the data.
10:16We ensure that we safeguard the access through a governance layer in there. And working with Snowflake, we've been able
10:24to implement that.
10:26So, we've been talking a lot about agentic AI for the past year. So, last year in Vivatech, we were
10:33only talking about generative AI. Today, it's agentic AI.
10:36So, is this something you're working on at Sanofi already, or it's just the beginning?
10:42So, it's something that we're very much working on. And to go back to the parallel, what I said, we
10:46used to do dashboards on the data.
10:48We moved away from that. With Elementum, who builds workflows on the data lake, so Snowflake is the data lake.
10:56With Elementum, we build workflows on the data lake, and we use agents, or sometimes we don't.
11:01Not everything has to be a probabilistic answer. Some things can be very prescriptive.
11:06So, we build workflows and new engagement workflows on, if you want, as close as possible to the data.
11:13History has been, we use systems of records. Take any system, a HR system, for example.
11:20And you move data into it to be able to do your workflow, and then you move data out of
11:25it to do your reporting.
11:26That used to be a cycle that was well understood. It's actually a very tedious and costly cycle.
11:32We decided we're not going to do it anymore. We're going to build workflows away from those systems, directly on
11:38the data.
11:38So, we've moved everything down as close as possible to the data. And so, we're not moving data up and
11:42down.
11:43The data, if you want, is resident, and we're reading and writing from the data lake, which is a little
11:48bit of a difference.
11:49And that's what, with their work and with Elementum's work, we've been able to do.
11:53And there's a lot of engineering that went into doing it.
11:57And you're bringing AI next to the data, right? That's what you're doing, versus there is a paradigm where you're
12:02trying to take data out to AI models.
12:06But that doesn't retrieve high-quality data, and it's also very inefficient in the way.
12:12So, the approach we've taken together with Sanofi is to actually bring that AI and the business knowledge, business workflows,
12:20that context that the models need to create the level of impact we need from AI right next to the
12:26data in the data lake itself.
12:27And so, that's the work from Snowflake from the AI platform that we've been doing with our customers.
12:34And doing it in a way that is, again, governed. It has pretty much an audit trail, right?
12:39These agents, you're giving them kind of superpowers. You have given a role. They have given an identity.
12:45They will go out and look at the data and make decisions and take actions on your behalf.
12:50What is the right governance layer so that there is a clear, you know, handoff between humans and agents, and
12:58agents have the power to go take those actions.
13:01So, there's a lot of work in that infrastructure that has gone in to ensure that there is security, safety,
13:07and governance is built into the platform itself that allows you to then scale these agents to a large organization
13:14like Sanofi.
13:16So, I think it's time to have a demo. I don't know which one of you.
13:21I'll be the one. So, I'm going to give everybody a little bit of context before going to it.
13:26Okay.
13:27So, this is something that we built with Elementum and Snowflake. And what I'm going to do now is I'm
13:32going to pretend for the sake of these next few minutes to be a sales rep in a medical company,
13:38ours.
13:39This is the production system I'm going to call. This is all live. This is not a demo. And I'm
13:44going to just do the activities that our sales rep would do, but in 2026.
13:49So, this is in production in the hands of some of our sales reps. What we build at Sanofi is
13:55we build a concierge. We call it concierge.
13:59Much like if you were going to a hotel, you know what a concierge service is. They guide you, they
14:05help you, they help you find answers, maybe a place to go out for dinner if you need to.
14:09So, this concept of the concierge is something that we build for our sales teams. So, our commercial workforce is
14:15now leaving the office with their phone and with concierge with them.
14:20So, I'm going to try to do it. There's a fair amount of background noise and the AI will pick
14:25the background noise.
14:26So, I'm going to do my best at giving you a view of what it is to be a Sanofi
14:32sales rep.
14:41Concierge for field.
14:42Hey concierge, listen, I'm about to head to meet Dr. Hannah Clevenger. Could you help me prepare my pre-call
14:49plan?
14:51One sec.
14:54I've got your pre-call plan for Dr. Hannah Clevenger. The biggest opportunity here is to focus on advancing treatment
15:00earlier in the disease course, especially for patients stuck on steroids or just PPI alone.
15:07This doctor recently started their first Dupixen patient but has a flagged patient with persistent steroid use that's likely causing
15:14frustration.
15:14Hold on, you said that they have a Dupixen patient. How many Dupixen patients does this doctor have?
15:20This doctor has recently initiated treatment for their first Dupixen patient and is actively managing a growing number of patients,
15:27including those still in PPI monotherapy, who could be candidates for Dupixen initiation.
15:33Okay, got it. What would be her objections to using a biologic instead of a steroid?
15:41The most common objection you might hear is that steroids or PPIs are working fine for now. The doctor may
15:48hesitate to switch because they see those treatments as familiar and sufficient.
15:51I'm about to get there. I don't feel so confident, concierge. What would be a great icebreaker for Dr. Clevenger?
16:01A good way to start could be by asking how she decides when a patient on PPIs or steroids needs
16:07to move to advanced therapy like Dupixen. Want me to help you handle the likely objection?
16:13No, can you email that to me, please?
16:18Preparing that for you now with the email I have for you on file.
16:22I'm almost there.
16:23I've emailed the plan to you.
16:24Um, wish me luck.
16:25Got it. Let me know if you want...
16:28Good luck! But with a million-dollar smile like that, you won't need luck. Go get him, Tiger!
16:39So, what you just experienced is what a number of our sales reps today have in their hands.
16:46This is the production capabilities.
16:48This is the workflow I was talking to, Elementum, on top of the raw data from Snowflake.
16:53This is the intelligence that we're bringing back.
16:55And you could see that I've taken a number of ways, I've interrupted the AI, and every time it was
17:01able to respond.
17:02So, that level of velocity is quite unique, and a lot of it is kudos to what Snowflake has done.
17:07Yeah.
17:08Yeah.
17:09And it's real production data that's running on. It's taking on a role.
17:13If I were to invoke it, it's going to be a very different answer.
17:16I might not be able to get exactly the same answer there because I don't have the same privileges and
17:21accesses that my agent would have versus what your agent would have.
17:24That's right.
17:25So, what lessons from this partnership can apply to other industries?
17:32And maybe if you can tell to entrepreneurs and innovators in the room, yeah, take away, you know, for the
17:40development of, yeah, in AI for the next decade.
17:44I'm happy to take it, and I would love your opinion as well.
17:47For me, I think really kind of three things.
17:50First is, as we've been hearing in this conference this week, the tech has gotten easier, very easier to go
17:56build things now with AI.
17:59Everything else is the hard part, especially the change management.
18:02So, first and foremost, you need something from the top.
18:06You really need a direction, a vision, and a mandate to really bring AI at the enterprise level.
18:14And I really want to commend Emmanuel for what you have done at Sanofi.
18:17Right?
18:18You have led with the vision, something that was, you know, not possible, not even fathomable just a few years
18:24ago.
18:25And then I see you working with the team all day long.
18:28You're the one doing the demo and not me.
18:31It's incredible to see the vision, the execution, and the velocity at which you're moving the team.
18:36So, I think that is something that every organization needs to move at the pace of AI.
18:43Second piece, I think we've touched on this quite a bit, which is building the shared context and understanding across
18:49the organization
18:50and feeding that into their AI models is critical.
18:55And Sanofi did the hard work of, you know, five years it took for them to consolidate
18:59and have one data estate and one data lake on which now they're able to accelerate their AI journey.
19:04And that hard work is absolutely critical for you to accelerate.
19:09And I would say the last thing is don't look at where the technology is today.
19:13Be, create, have your imagination, and then be very outcome oriented.
19:18This is another thing Sanofi has done really well, is to set out their goals.
19:22Everything that they're doing is very clear business outcome attached to it.
19:27And that is very critical.
19:28You're not doing science projects.
19:29You're not doing toy experiments and POCs.
19:32You have a very clear idea of where you want to be, and then technology is going to help you
19:36get there.
19:37At Sanofi, where else are you deploying AI and maybe what's the ultimate vision for you in the next five
19:47years?
19:48The ultimate vision, I don't think we have it yet, Charlie.
19:50I think it's happening by the day.
19:51But what I would say is, much like what Love was saying, we can't use, in my mind,
19:59agentic AI as an automation tool.
20:03It's misuse of the technology for the power of the technology.
20:07And let me explain that in a second.
20:09Humans are very good at improving existing processes.
20:12We're amazing at improving what we already know.
20:15We make it faster, cheaper, better, more resilient.
20:19We're actually not very good at reinventing processes.
20:22And AI is the tool to reinvent the process.
20:25Here we're reinventing how our commercial workforce goes.
20:28We could have taken what they have today and said, we're going to make it simpler for them.
20:31We said, no, how do we make it everything new and truly with them at the center?
20:37And use agents when you need them and use no AI when you need it.
20:40So what else we're doing?
20:41We're doing that.
20:42You saw it for commercial.
20:43We're doing that for IT support in the company.
20:45It'd be done the same way.
20:47We're doing that for HR support in the company.
20:49It would be done the same way, same principle.
20:51We're doing that for procurement in the company.
20:53We're reinventing everything that we do with procurement.
20:55And you're going to go inside concierge.
20:57You're going to describe what you need.
20:59Concierge know who you are, Charlie.
21:00Concierge knows which call center you belong, which function, which geography.
21:04Already known.
21:05Why do you have to go into a system?
21:07Enter forms, drop down menus, complicated pick lists, get lost three or four times before
21:12you do it.
21:13Describe in natural language what you need, what product you need to solve, and let the
21:19technology do its magic behind the scenes.
21:21So that's, for us, the bold ambition is there'll be only one entry point for an employee inside
21:27Sanofi.
21:28Concierge.
21:28And then you'll do all your activities through that one Trojan horse.
21:32Okay.
21:33One last question.
21:36Will, is health data ready for AI be still a thing of the, is it going to be a thing
21:43of
21:43the past?
21:44Or will it still be, we'll still be talking about cleaning Excel spies next year on this
21:51stage?
21:52Oh, we're going to be cleaning, we're going to be cleaning data for a, for a long, long
21:55time.
21:56And I think we're going to, my view is that, and Kauta, who's in the audience, who's not
22:01paying attention right now, and I were discussing that just last evening.
22:06Because our demands are going to grow, this concept of the AI ready data would continuously
22:11be something we'll chase.
22:13Because as you get this and you give it in the hands of people and say, well, I need more
22:17data on this.
22:18I need more.
22:18We don't have it.
22:19Where is it?
22:20It leaves in your Excel files, Charlie, that we have to.
22:24So I think it's kind of an endless pursuit, but an endless pursuit for a positive outcome.
22:29Yeah.
22:30And also you, Concierge is a great example again, right?
22:33Not only saving, what, 40% of bandwidth for your field reps, right?
22:38That's right.
22:38Pretty large scale.
22:39But the key is getting insights from those interactions.
22:43Now it's gotten so much better and easier with AI and bringing that back to, to Manuel's
22:48point earlier, to make better decisions for the company.
22:51Like that is, that's the next layer of data that you're going to unearth as you roll out
22:56these agents into the wild.
22:58Well, thank you very much.
23:00A great conclusion.
23:01Thank you, love.
23:03Thank you, Manuel, for being at Vivid Seiko again this year.
23:06Thank you for having us.
23:08Thank you for having us.
23:08Thank you for having

Recommended