From pilot to scale: why so many AI projects stall

The pilot worked, and nothing happens

In many companies of a few hundred to a few thousand people, the same scene repeats. An AI pilot, run over a few weeks, produces encouraging results. The demonstration convinces the executive committee. Six months later, the tool is used by a handful of people, the budget has not been renewed, and nobody is sure whether to stop or continue.

We do not think this is primarily a technology problem. It is a problem of the question being asked. A pilot answers "does it work?". Scaling requires answering a different question: "is it worth its cost in our organisation, and who is accountable for it?". Until that second question is asked, the pilot remains a demonstration, and a demonstration does not generalise.

This follows from what we wrote in Generative AI: moving from experimentation to value creation. Experimentation is necessary, but it only becomes useful when it is designed as the first step of a decision.

Why the pilot does not prove what we think it proves

A pilot almost always takes place under favourable conditions. Participants are motivated volunteers. Data has been prepared by hand. The project team is close by and fixes defects as they appear. Atypical cases, which often make up most of the real workload, are set aside. The measured gain is therefore an upper bound, not an expected average.

Five breaking points

1. Nobody owns the outcome. The pilot is carried by IT, an innovation team or an enthusiastic junior colleague. Scaling requires a business leader to accept changes to a process, to targets and sometimes to staffing. Without that leader, the project has no recipient.

2. Integration is underestimated. A tool used alongside daily work, in a separate window, does not last. Connecting it to the ERP, the CRM or document management, with coherent access rights, often takes more effort than the pilot itself.

3. The economics change scale. Licences, usage, support, human supervision and data maintenance are running costs that do not appear in a pilot funded from an innovation budget. They appear in the income statement.

4. The risk becomes real. With ten users, a mishandled piece of personal data is an anecdote. With two hundred, it is a compliance question. The revised Swiss Federal Act on Data Protection (FADP), in force since 1 September 2023, requires transparency, a record of processing activities and risk assessment. For companies serving customers in the European Union, the EU AI Act, in force since 1 August 2024, will see most of its obligations apply from August 2026. These frameworks do not prohibit deployment, but they demand a level of governance that the pilot has rarely built.

5. No stop decision exists. A project with no abandonment criterion is never abandoned. It goes dormant, consumes attention and occupies the place of another.

A worked example, with its assumptions

Consider a services company of 400 people testing a drafting assistant for its customer service. The assumptions are deliberately simple and illustrative: the pilot involves 5 people who save an average of 3 hours per week; the rollout covers 60 people, 44 weeks per year; the fully loaded hourly cost is CHF 85.

The theoretical gain is 60 × 3 × 44, or about 7,900 hours, in the order of CHF 670,000 per year. This is the figure that circulates in the executive committee.

But that gain only becomes a result if the freed time is actually redeployed: more volume handled, shorter lead times, a hire avoided. Suppose only one third is redeployed in the first year, around CHF 225,000. On the cost side, assume CHF 40,000 for licences and usage, CHF 120,000 for integration and data work, and CHF 60,000 for training and support. The total is CHF 220,000.

The outcome: year one is close to break-even, whereas the pilot suggested a substantial gain. Value appears afterwards, if the conversion rate improves and integration costs do not recur. The project is not a bad one. But presented with the pilot figure, it will disappoint and be judged a failure.

What leaders can decide now

Define the decision before the pilot

Before starting, write down three things: the criterion that will trigger scale-up, the criterion that will trigger a stop, and the date by which the decision will be made. A 90-day pilot with a firm deadline is better than open-ended exploration.

Name a business owner

The sponsor should be the person whose indicators move if the project succeeds: the head of customer service, of operations, of finance. IT is a partner, not the owner of the result.

Pilot under representative conditions

Include non-volunteer users, raw data and difficult cases. The gain will be lower, but it will be credible, and it is the one the CFO can commit to.

Budget for operation, not only for build

The investment case should show recurring costs and the share of the gain that will convert into results. Commercial use cases, such as lead qualification or pricing support within a growth and commercial performance effort, often lend themselves well to this measurement because the link to revenue is direct.

Treat pilots as a portfolio

Ten pilots running in parallel in a mid-sized company is nine too many for its capacity to absorb change. They need to be prioritised like any other investment, which is the subject of the case described in project portfolio rebalancing. And those choices need to be revisited regularly, as we explain in The three-year strategic plan that survives its first quarter.

AI in service of strategy, not the reverse

The point we made in 2025, the year AI has to show results still holds at the start of 2026: the question is no longer whether the technology works, but where it serves the strategy. Sometimes the answer is better organisation, a simpler process or a governance decision, with no AI at all. Our role, within our Strategy & value creation expertise, is to connect these choices: set the direction, deploy, then steer the results.

If you have pilots in limbo and are unsure whether to scale or stop, our diagnostic provides a structured way to ask these questions, with your teams and your numbers.