Why you're failing to implement AI in your company. Again.

Right now, in conversations with several companies, I keep seeing the same scenario. Everyone wants to implement AI. Everyone expects something from it. Many sense that it will become increasingly prominent. In practice, few teams turn this desire into useful, maintainable projects that business teams use and that can go beyond a demo shown once before moving on to the next topic.

I've heard this tune before.

In 2016, companies had already experienced this first wave

When I graduated from engineering school in 2016, I saw many companies beginning to want to implement artificial intelligence. At the time, people tended to talk about machine learning, recommendation systems or forecasting. Very often, the impetus came from the top. The company wanted to implement AI, then looked for somewhere to put it.

This approach quickly creates a mismatch. Projects emerge even though the actual need sometimes calls for a conventional IT tool. Other familiar initiatives are reclassified as data science because the subject is fashionable and the company wants to show that it is moving forward.

Many projects therefore remained at the POC stage.

The demonstration could be polished, sometimes even impressive. Moving into real-world use brought everyone back to the company's constraints. Imperfect data, users with established habits, security requirements to meet, maintenance to provide. At that point, the issue went far beyond choosing the right algorithm. It was necessary to understand how the project would function within the organization.

That is where many teams stumbled. They had treated AI as a technology trend. The work already involved data, business operations and organization.

Generative AI repeats some past mistakes

With generative AI, I really feel that we are replaying part of the same scene. The tools are far more capable. Adoption is much faster. The interface is much simpler. The instincts, however, look very similar to those of the first wave.

Over the past two or three years, many companies have invested heavily in AI. They have launched working groups, tested internal assistants, discussed RAG as a general answer to internal knowledge problems, and then started talking about agents as though the next step were obvious.

Some projects obviously deliver very strong results. AI is already serving real use cases. In many other situations, gains remain below expectations because the starting point is still the technology rather than the problem to solve.

To give a sense of the scale, McKinsey reported in its 2025 survey that 88 % of the organizations surveyed regularly used AI in at least one business function, but only 39 % reported an impact on EBIT at the company level. The use is there, and so are the experiments. Value at scale remains much harder to achieve. Gartner, for its part, explained in early 2026 that at least 50 % of generative AI projects had been abandoned after the proof of concept, particularly because of data problems, poorly managed risks, rising costs or business value that was too unclear.

I think that is exactly where the obstacle lies. The models can work. Companies often want to move forward. But an AI demonstration and an AI project capable of lasting within a company require two different levels of work.

We saw it with the internal chatbot trend. Many companies wanted to build their own in-house ChatGPT, sometimes for good reasons, sometimes also because it was the topic of the moment. We then saw it with RAG. Some teams imagined that connecting a model to documentation would give them a reliable assistant. The difficulty was simpler and more troublesome. The documentation has to be reliable and maintained, and someone has to know which source is authoritative.

Now we are starting to see the same thing with agents. The idea is to automate entire areas of work, while many companies still operate with unclear processes, poorly shared decision rules and overly vague improvement goals.

The gap is widening between power users and ordinary users

There is also another issue that clearly deserves an article of its own. The gap between AI power users and more ordinary users is becoming enormous.

Within the same company, some people use AI every day and have already changed the way they work. Alongside them, others have opened ChatGPT twice, found it impressive or disappointing, and then kept the tool out of their daily routines.

This difference changes a great deal. We often talk about AI adoption as though everyone were starting from the same place. In practice, levels of use vary greatly. Some already know how to interact with a model, reframe a problem, challenge an answer and identify the right moment to use the tool. Others still face something abstract, sometimes intimidating, sometimes even a bit gimmicky.

This subject really deserves another article.

Five tips to avoid making the same mistakes again

When a company truly wants to implement AI, I think there are a few simple tips to keep in mind. They sound basic. Yet they prevent the same mistakes made during the first wave of artificial intelligence from being repeated.

1. Talk to people who really know how to build AI

The company needs to talk to people capable of building an AI system, not just people who are very comfortable with the latest tools.

The distinction matters. Knowing how to use ChatGPT, Claude, Gemini or a no-code automation tool is a real skill. Building, evaluating and maintaining an AI system in a business context requires something else. The project needs people who understand models, data, limitations, metrics and deployment to production. Above all, it needs people capable of anticipating what happens when the system leaves the comfortable setting of a demonstration. It is a bit like cars. You can be passionate about cars and know every model and every option. When you want to build a car, at some point you still talk to engineers.

2. Start with the business problem before the desire to implement AI

The most common mistake is to start by saying "we need AI", then look for a use case that fits that desire. The right approach starts with a real problem. The team needs to understand why it exists, look at how people work today, and then ask whether AI is a good way to address it.

The right use case may be less impressive in a demonstration. Above all, it removes real friction from everyday work.

3. Accept that a conventional solution may do the job better

In many cases, the company needs a clearer process, more reliable data or a better-integrated tool more than it needs a model. Saying so is not a failure.

It is often a sign that the problem has been understood. Conventional automation can create more value than an AI project laden with uncertainty. A well-defined business rule or a better interface can also produce a more stable result, with lower costs and less complexity.

4. Take data seriously from the outset

This was already true in 2016, and it is still true today. Many AI projects fail less because of the model than because of the data surrounding it.

Outdated documents, contradictory sources, reference information that cannot be found or unclear access rights always end up surfacing in the project. AI inherits these problems. Sometimes it even makes them more visible, because it provides a very smooth interface to a system that remains disorganized underneath.

5. Stop doing POCs for the sake of doing POCs

A POC can be very useful if it prepares the way for something real. It quickly becomes a waste of time when its only purpose is to prove that a demo works.

Before launching a project, the team should already know who will use it, in what context, with what level of trust and under whose responsibility once the initial enthusiasm has faded. The difficulty of AI in business lies less in producing an impressive demo over two weeks than in the ability to keep the system working over time.

The issue remains business transformation

Ultimately, many companies think they are launching technology projects when they are actually launching business transformation projects.

This is probably where the difference lies between companies that will carry out a few nice experiments with AI and those that will succeed in making it a driver of transformation.

Implementing AI requires more than choosing a model, launching a chatbot, organizing a prompting workshop or announcing work on agents. The company needs to take a clear-eyed look at how it actually works. It is less spectacular at the outset. It is probably how we avoid making, yet again, the same mistakes as during the first wave of artificial intelligence.