🦄 Making great presentations more accessible.
This project aims to enhances multilingual accessibility and discoverability while maintaining the integrity of original content. Detailed transcriptions and keyframes preserve the nuances and technical insights that make each session compelling.
Overview
📖 AWS re:Invent 2025 - Can Your AI Show Its Work? Healthcare's Critical Imperative for Explainable AI
In this video, Jed from Dataiku and Kaoutar from Sanofi discuss how pharmaceutical companies leverage AI to reduce the 15-year timeline from molecule discovery to market. Sanofi uses AI across their entire value chain—R&D, manufacturing, and commercial—with 70% of small molecules now utilizing AI. Kaoutar explains their RAISE framework for responsible AI, emphasizing explainability in a regulated industry. Their AI Foundry combines AWS, Dataiku, and Snowflake to ensure AI-ready data with proper lineage, traceability, and bias control. She stresses the importance of central governance with executive accountability, requiring business sponsorship and three-month value demonstrations for all AI initiatives. Rather than following hype, Sanofi focuses on scalability, integrity, and impact, using both classical AI and Gen AI where each makes sense. The conversation highlights that in pharma, explainability is critical because patients need to understand why medicine works before putting it in their bodies.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
Sanofi's AI-Powered Vision: Accelerating Drug Discovery Across the Value Chain
Hi everybody, I'm Jed. I'm the SVP of AI and Platform at Dataiku, and I'm here today with Kaoutar from Sanofi. We're going to be talking a bit about explainability, about data pipelines, and how the pharmaceutical industry is leveraging AI today. So can you start a little bit by educating us on how Sanofi is using AI today and, more generally, what it means to be a big pharma? What are your goals in this industry?
Yeah, first, Sanofi's statement is that we are an R&D biopharma company powered by AI at scale. So basically, our ambition is to shorten the distance or the timeline between the therapy and discovery, because usually it's a little bit 15 years between when we start to search the molecule to have it in the market. So AI for us is one of the technology and opportunities to shorten this distance.
So what that really means in layman's terms is that you're having to make a really big bet on a couple of molecules, right, at any given time, and it takes a very long time to figure out if that bet is going to pay off. Speaking of Las Vegas and betting, this is a very bet-heavy industry. So how can AI help optimize that entire process? Where does it fit in?
So first, the value chain starts from, of course, the R&D to manufacturing supply, then the commercial side, and then employee experience. So our bet was first on AI. It's everywhere. We didn't choose to put it on R&D or on commercial or on manufacturing. It's all on that. And of course, you do have some quick wins where you see the outcome of AI quickly. For others, like the research, it's long term.
What's a good example of a quick win that you've seen? So quick wins, for example, can be on commercial or on manufacturing. If we can take the example, for example, on yield optimization, that's something that we see in real time using AI, how we can optimize our supply chain, how we can make a connection between scientists and the shop floor. Those kinds of things are quick wins, or even the sales reps, how they can make decisions faster, how they can push the right content to the right HCPs. And yes, for the long term, it's more on the discovery and the research of molecules. But even on that, today we are using AI on 70% of our small molecules, basically the chemical ones.
Wow. And when we say AI, are we talking about classical AI, Gen AI, both? Where's the delineation there? So it's all of them, basically, because for us it's not about hype. It's about scalability, integrity, and impact. So it's both where we can put the technology for the service of the impact and the outcome. So there are some use cases and some domains where the classical AI works well. Others, Gen AI, it's a huge opportunity, and AI agents are more better for other use cases.
Maybe, for example, I can take an example on Gen AI. Like everything relating to today regulatory submission, Gen AI is a huge opportunity. It accelerates filling documents. It accelerates putting the right information on the sections, and even the regulatory today, they are open to that. On others, like yield optimization, classical AI time series ML works well. So why go into Gen AI? Because the cost, the effort, and adoption are more complex.
So you've invested for 30 years into classical AI. There's no reason to throw all of that out just because Gen AI. Yeah, exactly. But it doesn't mean that I don't have requests coming from the business. Oh, we see opportunities on Gen AI. Can we use Gen AI? You have objectives that will help you care about if Gen AI is the most suitable thing or classical AI or even just advanced analytics, right?
Explainability and AI-Ready Data: Building Trust Through the RAISE Framework
And how does explainability come into play here? I know explainability is a very buzzed-up term, and of course in pharma, you really need to understand not just success but why success is happening. So what does explainability mean in pharma, and how important is it to you?
So there is a fundamental thing on the domain of pharma because it's a really regulated domain, and the first step of it is AI-ready data. It starts by the data itself, and then it goes, yes, to the outcome. These outcomes need to be trusted, need to be secured, ethics, fair, eco-sustainable, and of course transparent and explainable.
Here we have the explainability. That's what we have in Sanofi, what we call RAISE framework. It's basically responsible AI at Sanofi respecting the whole pillar adjustment, and yes, explainability is part of it. And what's the funny thing is that when we talk about Gen AI, there is a contradiction between explainability and Gen AI, right? It's inherently a black box, right?
Yeah, exactly. So what kind of tools or processes do you use to extract explainability from this Gen AI black box? So we built what we call the AI Foundry, and basically every AI product is using the ecosystem that we have in this AI Foundry. Three pillars, of course AWS, because if you don't mention AWS, I think both of us will be out today. So we have AWS, and we have Dataiku. We have you guys for everything around the data, the ingestion, the explainability of it, and the bias that we can have on the data, but also on the monitoring of the ML and the foundation models that we are using. And we have Snowflake. So basically three providers to build the whole pillar that I mentioned on RAISE framework.
So maybe split up those three providers. Explain what each one of them does. We have AWS for infrastructure, Snowflake for database, and then Dataiku. What do you use Dataiku for? So Dataiku is to make the link between the scientists, the data people, and the business users. So it's basically the front door for the whole scientist personas, manufacturing supply chain experts, commercial experts, R&D experts that don't care about if I'm using Gen AI or something else, but they need to have access to the data. They need to monitor the data, and they need to build an ML model and to use it without being experts on the domain.
Got it. Part of what I understand in your process is this desire to have AI-ready data. You've said that phrase to me a couple times in previous conversations. Tell us more about what you need, what that means. What is AI-ready data? So AI-ready data, first, the data is a shared accountability between the business and the IT or digital team. Then the data is known. We know where it is and we can find it. Also, it's something about data being shareable because we are moving in a new era where I am not building my product for my business, but the value of my product is cross-functional, so it needs to be shareable with other products to make a decision and to be at the end available to everyone.
And then we have something related to the data itself, which is the data is secured, is with high quality, without bias, and trusted and safe. That's what I mean with AI-ready data. Secure, high quality, without bias, trusted, safe makes a lot of sense. And to do that, you need, like what we're using, you guys, is to showcase that we have this lineage of the data end to end, and we can share outcomes between products and sharing the data of course with the control access control, but we can have this traceability and explainability of what we are doing and the why and how I got this outcome.
Governance as the Foundation: Managing Risk and Value in AI Implementation
Makes sense, being able to show your work basically. So what do you think some of the biggest risks are in pharma when you're implementing these new AI capabilities? What do you really need to watch out for? So we have one objective that I mentioned at the beginning. We want to reduce the timeline from discovery to therapy, and we have a lot of products that help us to do that from R&D to commercial to employee experience. And to do that, we need to have a clear governance. It's not because, again, it's not about hype. AI can be used, yes, for several purposes, but we have one purpose. If the product that we want to build, that we want to deploy, is not helping this statement, it's out of the governance.
So that's why we set a clear governance, and that's why we set one platform, which is our AI Foundry, around this ecosystem. And the AI, again, for us is opportunities, a technology that helps us to accelerate, to go faster with efficiency.
Got it. Yeah, I think we saw early on in this generative AI phase the inclination to put something out there because you needed to have some kind of generative AI, right? Early on, the classic thing was like the HR chatbot, and now we're moving on to maybe things with a clearer ROI, perhaps a more obvious way of moving your very clear needle. Shortening that time to value makes a ton of sense.
One thing that I hear often among clients is obviously around this governance concept. I guess a common governance idea is data governance, so basically who has access to what data, but then there's also a workflow governance, right? Who has approved of using this project, whether there's documentation, whether there's been a risk analysis. Do you see the need for both types of governance, or is there really one that's a lot more important?
We have basically two governance. One is top level with Excom members. Basically all our AI, generative AI initiatives, they go through this governance. We have what we call front door. It's the first entry point where every employee in Sanofi has an idea about AI that can change the world. We told them, okay, prepare your pilot ID card. What does it mean? That they need to have a business sponsorship, a high one at Excom level, digital sponsorship, because we need to ensure that the value will be there, concrete value, and also that's in terms of feasibility, technical feasibility, it can be done.
We ask them to have commitments, short-term commitments, three months. You need to showcase part of the value because somebody can tell you, look, we will have 40 million ROI. Cool, okay, so in a year? Yes, in a year, okay. I'm a little bit stupid. I will just do an easy calculation. In a year, let's come back to three months. I will give you a sandbox with the whole tools and show me the value in that three months. Of course we have accountability of Excom.
Then people, they will say, oh, okay, I will not burn myself because I'm putting myself in front of Excom, in front of a sponsor, a high senior leader in digital. At least in that, we moved from, I can give you an example on commercial, we received 56 use cases and at the end we landed at two. They removed all the use cases when we asked them. That's the prerequisite before pushing your use case. That's this governance. It's needed because if not, just remember in 2010 when all advisors, they said if you invest one dollar in AI you will get four dollars. A year after that they lost four dollars for every one dollar invested, so we can't go without this governance on the value.
Then you have the other governance which is related to data and AI, which is I would say more technical one and about all the concept of AI-ready data. We need to understand the data because there are some issues, there are some risks, we have a regulator and patient data, all these kind of things. We need to have this easy to digest data governance, what we call AI-ready data. It's basically the data governance.
That makes a ton of sense. When we're talking about those regulations, with the pharmaceutical industry, you really have different regulations and different rules in every country you're going into. How do you manage all of that? That just seems like it's such a heavy burden. Do you have different teams or different ways of targeting, let's say the US regulations versus EU regulations, or is there a central governance team that manages all this?
The governance is central, central governance. Basically it's two. The data team where I'm heading that, and we have the governance with the generative AI board with the Excom members, and yes it's at global level. Then you have sub teams for countries level because you need them. They are more experts on the domain. At the end, if you look to the whole regulation, there are a lot of commonalities and yes there are some specificities which you need to manage, but that's why we have this global central team for governance and you have local team but they're working with the global team. They are not reinventing the wheel, but they are specialized in governance in countries.
Beyond the Hype: Focusing on Shared Objectives and Patient-Centered Outcomes
Yeah, okay, that makes a lot of sense. I've been thinking a lot about how the names of workers or maybe the names of titles are going to change as this stuff evolves in organizations, and I'm starting to see a split where maybe 80% of workers are going to be agentic consumers or users, and then 15% are going to be designers of these agentic tools or components, and then you have 5% that's governance.
Do you see that sort of distribution reflected inside of pharma organizations?
It's not all pharma that are the same. I can speak at least for Sanofi. We have a purpose which is to put medicine at the hand of patients in the world. For us, the patients are not US patients or French patients or German patients or Chinese patients, they're patients in the world. So we don't care about the hype. We are not following, and if you look, not only Sanofi but all pharma, they were a little bit late to embrace the AI domain because of regulation, because of constraints, because of a lot of things. Because at the end, it's not only a quick win. You said that at the beginning, there is research of molecules long term, so you can't do it just because you are changing roles.
Yes, I see that there are some teams saying, okay, I will be an agentic AI specialist, I am a Gen AI specialist. But what does it mean, Gen AI specialist or agentic AI specialist? If you are an AI expert, you manage Gen AI, you manage AI agents, so there is no expert in the domain. Honestly, we don't care about the title. We don't care about the name. For us, it's a technology that serves an objective, and we see if it makes sense to put it for some objectives or not. That's why the governance, again, is really a key.
It's interesting that you keep coming back to having a single objective, having a goal or a set of decisions, a North Star, a target to lead towards. As we look at other industries, I won't ask you to be an expert on them, but do you think every industry or every company needs to have a single driving objective or should have a single driving objective when they're implementing AI?
I think it's key. If the company doesn't have a shared objective, that means we'll have more and more silos, more and more local objectives. But in mathematics, we know that the sum of local objectives is less than the global objective. So if we want to achieve as a company an objective, it needs to be a shared one. That's why we put this governance on AI and Gen AI with the executive committee, because we needed accountability for all divisions.
All companies have a starting point of their product and they have an end point. So at the end, all companies have a shared objective that they need to highlight. For us, it's the distance between treatment and discovery. A molecule starts in research, but the research themselves can't put it in the market. They need to develop it, they need to manufacture it, and then they need to commercialize it. So the workflow of the data and of the objective is the same, but with different ways of how I will contribute to that objective.
We're touching the entire value chain. That makes sense. We need to be injecting AI inside that value chain only in the places where it actually makes sense and pushes towards that end goal. Right, perfect. So we're right about at time here. Any last words you'd like to say about explainability, about getting your data ready, and about rolling out AI inside the pharmaceutical industry?
So there are two things. One, we don't need to follow hypes because every company has a context, has constraints, and has a history. We need to move forward with the whole heritage that we have. To go far, it's not about going fast, because alone we go fast, but together we go far. That's one thing. The other thing is biology in my domain. Not everything is explainable, but medicine needs to be explainable for patients. If I'm going to put it in my body, I want to know why it works.
Exactly. Makes a lot of sense. Thank you so much, and I hope everybody out here learned something. Thank you.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)