“Success Tells You Something. Failure Tells You Too Many Things”. Vincent Fraitot on AI Transformation

By some estimates, more than 80% of AI projects fail to deliver the value expected of them. For Vincent Fraitot, those failures explain very little: it is the successes that are readable, and what they reveal goes well beyond technology. The HEC Paris professor on what it takes to scale an AI transformation.  

Vincent Fraitot is Associate Professor at HEC Paris and Scientific Director of the MSc Data Science and AI for Business (a joint degree with École Polytechnique), and academic director of the Executive Certificate AI for Business Transformation. He spent more than thirty years in technology and data before teaching it, as a CIO and Europe organization director at Pernod Ricard. We asked him where AI transformations break first, whether shadow AI is worth fighting, and which profiles a company actually needs to run one.

Where does an AI transformation typically break down first?

Vincent Fraitot: It really depends on how mature the company is.

Take a company with a solid technical base. There, the first thing to give way is almost always the organization. People push back, they resist, and the structures and the ways of working between them turn out to be far too rigid to move at the speed the project needs.

Now, a company that has not built that base yet will crack on the technical side first, and it generally starts with the data. There is not enough of it, or it is messy, or the machines simply cannot crunch it fast enough.

But a failure tells you too many things at once, because it can come from a hundred different places: technical, organizational, human…  And you rarely manage to pin down which one actually mattered.

Can you read it the other way round, as a test of the organization that hosts it?

V.F.: No, I don’t think it works that way. It is so multidimensional that you cannot read it like that. But when there is a success in AI transformation, that says a great deal about the company: that it has a reliable technological base, and that it has the capacity to make its processes and its organizations evolve. For me that is a criterion of adaptability or agility, of the capacity to transform. So yes, if it goes wellwell, it is a signal. If it goes badly, there can be so many reasons that for me it is not necessarily a signal.

Why do so few companies get past the pilot stage?

V.F.: Because there are two very different hurdles, and people confuse them.

The first is building a proof of concept. That alone is not given to everyone. The second is taking that pilot and turning it into something that runs across the whole company. The gap between the two is enormous.

Scaling drags in a whole set of problems the pilot never had. You suddenly need the muscle to process real volumes of information, the computing power to keep the models alive, and the ability to deploy to a very large number of users.

So who clears it? Almost always the same profile. Big IT departments, serious data science teams, and several years of projects already behind them. Banking is the example I always reach for: a bank's information system is its lifeblood, and banks have been disciplined about data since the day they opened their doors. Decades of it. So no, it is no accident that these projects scale better in banks and insurers. They laid the groundwork long ago.

Can smaller companies still catch the train?

V.F.: It is very hard, and not only for small companies. Mid-size firms run into the same wall.

The problem is that the base is usually missing. No data scientists in-house, no AI engineers. So if one of these companies genuinely wants to do AI, it ends up outsourcing, and outsourcing carries its own baggage: providers who barely know your business, systems that live outside your walls, and a grip on the whole thing that keeps slipping.

That is why, once you look past the chatbots, real AI use in these firms stays very low. 

Agentic AI, meaning systems that act on the business itself, is still out of reach for small and mid-size companies today.

What is shadow AI, and should companies fight it?

Definition. Shadow AI: the use by employees of AI tools their employer has not approved.

V.F.: The rules exist for a reason, and they should be respected where the company has bothered to write them. Which, by the way, plenty have not!

The risk is real. I was talking with someone in petrochemicals recently, and what keeps them up at night is losing proprietary know-how, and with it a slice of the market. When people can't find what they need in the approved tools, they grab a free version of something else, and company data slips out with it. It’s a genuine confidentiality and security issue, and that’s exactly why the first rule is to follow the company's own code of conduct.

Now, most of what gets labeled shadow AI is completely harmless. Someone who is hopeless at slides asks a chatbot to make them look good. Where is the harm? As long as it is out in the open, the company comes out ahead. Same with developers coding alongside AI. That ship has sailed, it is the norm now, and no employer can seriously scold a coder for it.

AI job titles are multiplying. Where do non-technical leaders fit?

V.F.: There is room! Far more of it than people fear.

I see three families of roles. First, the technical profiles, your AI engineers and ML researchers, who have trained anywhere from six months to two years and more. Then, the AI managers, who come up from the business side. They are not developers, but they have learned the vocabulary, the stakes, the concepts, well enough to turn AI into something that actually sells. Finally, the governance leads and the chief AI officers, who steer both camps at once and can come from either one.

There is a fourth profile companies keep describing without a settled name for it: we can call it “the translator.” This person sits between the engineers and the executives and carries meaning back and forth. How do you find them? Most companies grow them from the inside, often through executive education. It works well for that: the climb is gradual, spread over time, and the range of themes is wide enough that you can aim at wherever your gaps are. You can also go outside and hire someone who has already done the job elsewhere. They are usually in large structures and in tech companies, where they have spent several years connecting business and technology.

Can a CFO, a CMO or a CHRO really lead the AI transformation?

V.F.: Yes, absolutely.

What I have watched happen inside companies is that it is never really about the profile of the person carrying the project. It is about their appetite for it, and their sheer determination to get things moving. 

A finance director can lead this. So can an HR director, so can a marketing director. The one condition is that they are engaged, truly engaged, because the road is full of potholes, technical, business and organizational alike, and it can be painful.

Concrete cases are still rare, though. In the big groups, this increasingly lands on a chief AI officer, a role created precisely for the job. But in mid-size companies that cannot justify such a post, it is often a motivated department head who steps forward, whatever it says on the business card. It is more a question of the person than of the job description.

Does a young engineer who builds agents end up with more credibility than an executive committee member who does not?

V.F.: More credibility on that subject, yes, and rightly so.

But it stops there. Agents do not set the company's strategy, they do not make sure it gets implemented, and they do not steer investment decisions. That is what executive committees do. So no, there is no risk of the developer who builds agents ending up running the company. It will not happen that way.

What can happen, and this is a real possibility, is that he becomes chief AI officer and joins the executive committee himself. That takes nothing away from anyone. 

A company will always need a CFO and a marketing director, and the chief AI officer is not there to replace them. He or she is there to help them grow the business with AI.

A transformation only works if every level and every function of the company has a grip on these concepts.

What has to happen before technical and business people can even discuss AI in the same room?

V.F.: Ah, that is the real challenge, and it has been for years! Highly heterogeneous cohorts are simply the norm.

It all starts with vocabulary. Everyone in the room has to stand on the same base of concepts before anything else can happen. Once they do, we teach through business cases, real situations where AI moves the business forward, and those cases always have two faces at once, a technical one and a business one.

What will still distinguish great leaders in 2050?

V.F.: Adaptability. Full stop.

When an organization cannot adapt, it dies. For AI, the story is exactly the same as the arrival of the internet almost thirty years ago.

Creativity, on the other hand, stays with us humans for now. By design, generative AI hands you the average of everything that is already known and said. A good average, sure, but an average all the same. 

The creative leader is the one who makes the connection nobody else saw coming, and that is precisely where today's models are at their weakest. 

They are brilliant at automation, much less so at coming up with the idea in the first place.

Will that last? Honestly, I do not know. And does the rise of these systems scare me? People take it very differently. Some freeze, some fear for their jobs, and those fears are legitimate. Me? It pushes me to go further, and to do better than the machine.

The Executive Certificate AI for Business Transformation runs over twelve days, with participants from both the technical and the business sides of the same problem, under faculty from HEC Paris and the Hi! PARIS center.

The next cohort begins October 19, 2026, in English, on the Jouy-en-Josas campus.

Download the brochure or book an appointment with our Program Advisor.