The successes and failures of AI projects

Why some AI projects succeed and others fail

Integrating artificial intelligence into a business is a rich learning experience, through its successes as much as its failures. The technology is rarely the reason a project fails. What makes the difference is how the project is framed, the quality of the data it uses and the way people take ownership of the tool.

The lessons below apply to a wide range of projects: a chatbot for internal questions, automatic meeting minutes, document analysis, demand forecasting or the classification of customer requests. They will help you steer your own AI initiatives and avoid the most common pitfalls. For an overview of the possible uses, see our page on AI at the service of our organisations.

What makes an AI project a success

Understanding the specific need

Companies that succeed with AI start by understanding their own needs and the problem they are trying to solve. "Using AI" is not an objective. "Reducing the time spent writing meeting minutes" or "answering recurring HR questions faster" is. A well-defined need makes it possible to choose the right tool, to measure the result and to stop the project if it brings nothing.

Investing in data quality

High-quality data is essential. An AI model only produces results as good as the data it relies on: incomplete, outdated or contradictory data leads to wrong answers, delivered with confidence. Before launching, check where the data comes from, who maintains it and whether you have the right to use it.

A culture of innovation and experimentation

Fostering a culture of innovation and experimentation is a key success factor. In practice, this means accepting to test on a limited scope, measuring honestly, and stopping or correcting what does not work without looking for someone to blame. A pilot with a volunteer team is often more instructive than a large-scale launch.

Continuous training

Continuous training is crucial. Users need to know what the tool does well, where it can make mistakes and how to check its output. Training is not a one-off session at launch: it continues as uses evolve and as the tools themselves change.

Keeping people in the loop

Projects that work keep a person responsible for the final decision. AI proposes, people validate. This is particularly true for anything that commits the company: a decision, a message to a customer, a figure in a report.

The most common sources of failure

• No clear strategy. Failure is often the result of a project launched because "everyone is doing it", without a defined need, a sponsor or a success criterion.
• Underestimated costs and timescales. AI can be complex and expensive: data preparation, integration with existing tools, security, maintenance and user support all take time.
• Poorly managed expectations. Companies that have not communicated well about what AI can and cannot do have often disappointed their stakeholders. Promising too much at the start is a reliable way to lose user confidence.
• Ignored data quality issues. A model trained or fed with poor data gives poor results, and the project is blamed instead of the data.
• Overlooked compliance and security. Personal data, confidentiality of documents and the location of hosting must be addressed from the start, not at the end. In Europe, the GDPR and the European AI Act set the framework.
• A tool that sits outside daily work. If users have to leave their usual environment to use the AI, adoption drops quickly.

Steering an AI project like any other project

An AI project is first and foremost a project. The same rules apply: a sponsor, an owner, a defined scope, milestones, a budget and a follow-up of risks. The methods described in our 7 steps to effective project management apply here too.

Regular steering committee meetings help to decide on the next steps: extend the pilot, adjust the scope or stop. Each decision should leave the meeting as an action with an owner and a due date. Our example of an action plan to follow up your SteerCo meetings shows how to organise this follow-up.

Checklist before launching an AI project

1. Is the business need written down in one sentence, with a measurable expected result?
2. Is the data available, reliable and usable from a legal point of view?
3. Is there a sponsor and a project owner, with time to devote to it?
4. Have the future users been involved from the design stage?
5. Is a limited pilot planned before any large-scale rollout?
6. Are the costs of integration, training and maintenance included in the budget?
7. Have the limits of the tool been explained, so that expectations remain realistic?

By learning from successes and failures, you can steer your own AI projects towards success. Meetings are a good place to start: our article on how AI will revolutionise the way you run meetings describes the most concrete uses.