Why so many AI strategies remain lists of pilots
An enterprise AI strategy often takes the form of a document: a vision, an overview of technologies, a list of use cases gathered in workshops. The document is approved, a few pilots start, then the momentum fades. The pilots are neither stopped nor rolled out, and leadership still cannot say what AI is bringing it.
Four mechanisms explain this outcome.
- Use cases are collected, not chosen. A workshop easily produces dozens of ideas. If none is set aside, resources are scattered and no project reaches the size that would make it useful.
- The gain has no owner. “Freed-up” time only has value if a business owner commits to what they will do with it: handle more cases, reduce a lead time, absorb an increase in activity with the same headcount.
- Prerequisites appear after launch. Inaccessible data, no legal basis for processing it, impossible integration with the target application: these obstacles are visible from the framing stage, provided you look for them.
- The operating cost is forgotten. A service based on a language model is most often paid for on usage, has to be monitored and must be re-evaluated with each change of version. A business case limited to the build cost is wrong.
The trade-offs of an enterprise AI strategy
A useful strategy settles a few questions that projects cannot resolve each on their own.
Where to look for value. Three horizons coexist: individual productivity (drafting, summarising, information search), the transformation of a process (case handling, customer relations, maintenance) and the creation of a new offering. They have neither the same level of risk nor the same mode of funding. Individual productivity spreads quickly but is hard to measure. The transformation of a process is easy to measure, but it requires changing an organisation.
Buy, configure or build. The answer is given use case by use case.
| Option | When it is suitable | Point of attention |
|---|---|---|
| Activate an AI feature in software already in place | Standard need, data already managed by the vendor | Data usage terms, cost per user, little differentiation |
| Configure an assistant on a market platform | A need specific to your documents and rules, without deep integration | Quality of the document base, reversibility |
| Build a custom solution | Specific process, strong integration with the information system, a differentiation stake | Skills to maintain, operating cost, evaluation set-up to be built |
How to organise. A central team brings consistency and pools scarce skills, at the risk of drifting away from the business. Teams spread across the business units move fast, at the risk of rebuilding the same foundation several times. A federated model combines the two: a shared foundation (access to models, security rules, evaluation method) and teams close to the processes.
What level of risk to accept. The classification of systems under the AI Act and the rules applicable to personal data guide the choice of the first use cases. This point falls within AI governance, which is built in parallel with the strategy and not after it.
Our framing method: from use case to business case
Our consultants work within your teams; decisions remain yours.
Start from the processes. We do not begin with an inventory of technologies. We examine with business leaders the processes where volume, lead time or quality pose a recognised problem. We then look at what AI would concretely change there: which task, for which user, with which input data.
Qualify each use case on four criteria. Value: which business indicator should change, and who commits to it. Feasibility: does the data exist, is it accessible, of what quality. Risk: classification level, personal data, consequence of an error. Adoption: does the process need to change, are users asking for it. Feasibility most often comes up against data, hence the value of involving those responsible for data governance early on.
Write a business case that can be verified. For each selected case, we formalise the gain assumptions, the full cost (build, operation, continuous evaluation, user support) and the conditions that would lead to stopping the project. A business case without a stopping criterion commits nobody.
Order the roadmap by dependencies. Several use cases often share the same foundation: a document base, secure access to models, an evaluation set-up. Launching them in the right order avoids funding the same foundations several times. The roadmap distinguishes what is committed, what is conditional on proof and what remains an option.
Embed arbitration over time. AI use cases join the organisation’s project portfolio, with the same prioritisation rules as other investments. We link the AI roadmap to your project portfolio management set-up where one exists. Failing that, we help put a light arbitration body in place.
The costliest framing mistakes
- Confusing AI strategy with choosing a supplier. A licence contract is not a strategy. It answers the question “with what”, not the question “to do what”.
- Launching pilots without a scale-up criterion. A pilot whose success threshold nobody has defined can neither fail nor succeed. It just goes on.
- Betting everything on a flagship use case. A showcase project, complex and risky, delays the simple gains that would have built trust and skills.
- Underestimating support. A use case that changes a job without the teams being prepared for it will be bypassed. Change management is part of the cost of the project.
- Freezing the roadmap. The capabilities of models and their prices change quickly. A strategy that is not reviewed regularly describes a market that no longer exists.
The signs of an AI strategy that works
A strategy is judged by the decisions it produces, not by the quality of the document. From the framing stage, we agree with you on a few observable signs:
- the share of use cases in the portfolio that have a named business owner and a tracked indicator;
- the number of pilots stopped by explicit decision (a portfolio where nothing is ever stopped is not being managed);
- the share of use cases that have moved from pilot to service in operation;
- the gap between expected gains and observed gains, reviewed at fixed intervals;
- the effective reuse of shared foundations from one project to the next;
- the time between the expression of a need and the decision to fund it or set it aside.
These signs also serve to review the strategy. When the gap between expected and observed gains widens across several use cases, it is the initial assumptions that need to be revisited, not just the projects.