Why doesn't a successful AI pilot prove an AI strategy works?

Getting an AI tool to work in a small test is one thing. Getting it to work reliably across an entire company is another, and many businesses are discovering the gap the hard way. That is the central argument from Zacharie Lahmi, founder of Ellis Partners, who says a successful pilot is not proof that a company’s AI strategy is working.

The numbers suggest the problem is widespread. A global 2026 survey by McKinsey found that only 37% of respondents said AI had improved their company’s earnings, unchanged from the previous year. Yet eight in ten said AI had helped them become more productive at work.

A tool that performs well during a limited trial often has a simple task and clean data to work with. Rolling it out across a company is a different job. The system has to connect to existing software, meet data access and security requirements, and include a clear process for handling failures. Someone also has to answer for how it performs. Doing all of that means coordinating IT, operations, finance and risk teams, which is where many projects stall.

Set the cost before the pilot

Money is another sticking point. Companies are accustomed to budgeting for software and estimating staff time. AI costs are harder to forecast. They depend on how much information the system processes, which model it runs on and how often employees use it. Without a clear estimate, finance teams may find it difficult to decide whether a project is worth the spending.

What should companies define before an AI pilot starts?

Lahmi, who has ten years of experience in technology, including work helping launch monday.com in France, recommends evaluating AI opportunities one business process at a time. His approach starts with estimating the cost and resources each use case will need before anything is tested.

“Companies need to understand the cost of a use case and the resources it will require,” he said. “The return-on-investment discussion should happen before the pilot, not after.”

Every test also needs clear goals and a plan for when things go wrong. Lahmi suggests setting measurable targets for usage, maintenance costs and errors. He says the success criteria should be agreed on in advance.

“Useful pilots are the ones we really want to see in production,” he said. “We should agree on success criteria beforehand, so there is no debate after the pilot about whether it worked.”

Look past sales and marketing

Lahmi also cautions companies against focusing only on sales and marketing when picking AI projects. Those functions tend to offer benefits that are easier to tie directly to revenue. He argues that everyday operational processes may offer bigger opportunities.

“The most boring operational tasks, the core processes of your industry, are the ones you should be thinking about with AI,” he said. “Ask yourself how you can fundamentally change the way your industry operates, rather than simply maximizing short-term gains.”

Some enterprise AI projects stay in testing for months, Lahmi said. Often, no one has defined what a project must achieve to justify a wider rollout. He recommends judging each project by the business problem it is meant to solve, then weighing the expected benefits against the effort required and the time it would take to see results.

“A smaller project that delivers results quickly may be a better starting point than a larger one,” he said.

He also suggests looking closely at routine business operations. Improving a process can change how a company works, not just help employees finish tasks faster.

Company size changes the pace. Lahmi notes that smaller businesses often have fewer layers of approval, which lets them make decisions and change direction more quickly. Large enterprises may face more obstacles because a change can affect several teams and existing systems.

How can companies tell whether an AI system is delivering results?

Regardless of size, companies need to check whether an AI system is delivering what they expected. They can track how much time it saves, whether it reduces mistakes and what it costs to maintain. Those measures can help leaders decide whether to keep a project, change it or drop it. A successful demonstration alone may not be enough to justify broader adoption.

From experiment to dependable tool

For executives, Lahmi says the real test of enterprise AI is not doing the same work faster. It is changing how value is created. Launching more pilots does not necessarily mean a company is making progress. Each project needs a clear goal, a realistic budget and limits on the risks the business is willing to accept.

Companies can also review performance after launch and make sure someone has the authority to pause and audit a system if problems emerge. Those steps can help turn AI experiments into tools a business can depend on.

Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.