Implementing AI: from pilot to daily use

Implementing AI works when you pick one process, check your data first, build the tool into the workflow your team already uses and make someone from the business responsible for the result. It sounds simple, and yet most AI projects stall exactly there: between a good-looking demo and something that runs every day. Below is the route that does lead to daily use.

Why do so many AI projects stall?

Because the demo is the easy part. Projects rarely die of the technology. They die of a model that runs alongside the real systems, that nobody owns and that gets judged on whether it works instead of on hours won. The figures draw the same pattern.

According to McKinsey, 88% of organisations now use AI somewhere, while almost two thirds still scale it company-wide nowhere. Deloitte calculated that barely a quarter of organisations get at least 40% of their AI pilots into production.

And then there is the notorious "95% of AI pilots fail". That figure comes from a preliminary MIT report that took heavy criticism on its method, and it measured something more specific: how many organisations saw no measurable return on their GenAI pilots. Take it as a signal rather than a law of nature.

Which process do you pick for your first AI project?

One process that is repetitive, runs by rules and happens often enough to free up real hours. How you make that choice we worked out earlier in our article on AI automation for SMEs. For the implementation there is one extra requirement: agree up front what you will measure success against.

Hours a week, the lead time of a file, the error rate. With a goal like "gaining experience with AI" you cannot say afterwards whether it worked.

Is your data ready for AI?

Often not yet, and that is the blind spot that costs the most. Gartner predicts that through 2026 some 60% of AI projects that do not rest on AI-ready data will be abandoned, and found that 63% of organisations lack the matching data practices or are unsure they have them.

The practical checklist for an SME:

• Where does the data live? In one system, or spread over mailboxes, spreadsheets and an accounting package?
• Is it consistent? Is the same customer called the same thing everywhere, or does it exist three times in three spellings?
• Who owns it? If nobody is responsible for the quality of the customer data, the AI will not be either.

Bad data does not have to block the project. Cleaning it up is often just the first, unglamorous step of the AI project itself.

How do you get from demo to daily use?

By building inside the real work from day one instead of next to it. An AI that reads invoices but whose output someone has to copy over by hand stays half a solution. Fitting it into the systems you already run, your AI orchestration, is usually the biggest chunk of the work. The approach that works for us:

1. Build the smallest version that covers the task from start to finish, inside the tools your team already uses.
2. Let old and new run side by side for a while and compare the results. That way you build trust with figures instead of promises.
3. Appoint an owner in the business. Whoever knows the work guards the result, and the builder guards the technology.
4. Invest in the people. BCG sums the ratio up as 10-20-70: 10% of the work sits in the algorithms, 20% in technology and data, 70% in people and processes. Most companies budget it exactly the other way round.

That same BCG study of a thousand companies found that only 26% take the step from proof of concept to tangible value. The difference rarely sits in which model they picked.

Who checks what the AI does?

Someone, always, and how strictly depends on the risk. The more money or customer contact hangs on it, the closer a human sits to the output. A usable ladder:

• Human approves: the AI proposes, a human confirms before anything leaves. For everything involving money, contracts or customers.
• Human watches: the AI acts on its own, a human checks samples and can roll back. For reversible, low-risk work.
• AI works within fixed limits: only for high volume and low risk, such as archiving or labelling documents.

Skip that check and you shift the work around instead of saving it. Researchers from BetterUp and Stanford described in HBR how 40% of the office workers surveyed received AI output that looks solid but does not move the work forward. Repairing it cost nearly two hours per case on average.

How much time and budget does a first AI implementation take?

For a well-defined process at an SME you are talking weeks, with the integration into your existing systems as the biggest chunk of work. We answered the budget question in our article on AI automation, and the rule of thumb we follow there is this: buy what is standard and only have built what is woven into the way you work. AI has made that custom work cheaper than it was a few years ago, which we wrote about separately. The MIT report too, caveats and all, saw bought and externally built solutions succeed roughly twice as often as fully in-house builds.

How do you measure whether it works?

With the yardstick you set before the start: hours, lead time and errors over the parallel period. If the project hits its numbers, expand to the next workflow. If it misses them, stop and take the lesson with you. A pilot stopped after a few weeks is a cheap insight. A zombie pilot that keeps simmering for months is not.

---

Stuck with an AI pilot that will not leave the demo phase, or do you want to get it right from the start? Send us a short description of what you want to automate. You will hear which route we would take and where the pitfalls sit for your case. No sales pitch, just an honest assessment.