AI strategy in mid-sized companies: what has to be settled before the first pilot

What is an AI strategy, and what isn't it?

An AI strategy answers three questions: where should AI make a difference in this company, who owns it, and how will we know it's working. That's all. And it's considerably more than most papers on the subject contain.

It's not a list of tools. Which model, which vendor, which license: these are decisions with a shelf life of months, and they're reversible. Starting there means having the discussion where a wrong answer costs the least. And where a right answer gains you the least.

Nor is it a separate strategy alongside your corporate strategy. AI is a means, not an end. If your corporate strategy says you want to make money in the service business five years from now, then the AI question is: what in service can we do better with it. An AI initiative that can't be tied to a strategic intention is a technology project with an unclear sponsor. And those are exactly the ones that don't survive the first staffing change.

Which question comes before the tool question?

The question about the work that should disappear. Instead of "what could AI do for us": "which step in the work will no longer be necessary once this is done."

The difference is practical, not linguistic. The first question produces a collection of ideas that ends with twenty options and nobody able to decide. The second leads to a specific place, a specific person, and a measurable number: quote preparation in inside sales, currently two hours per quote on average, owned by the head of sales.

• Which step goes away or gets faster? If the answer is "everything gets a little easier," the case isn't sharp enough for a first initiative.
• Who notices when it works? A person with a name who suffers from the effort today. Without that person, you lack the force that carries an initiative through the first bad weeks.
• What happens with the time freed up? Better to settle this question up front than after the fact, and it's the reason workforces either block such initiatives or carry them.

Where does value actually show up in mid-sized companies?

Almost always where people today read, sort, compare, or transfer text from one system into another. It's unspectacular, and it pays off, because in many companies it accounts for a double-digit percentage of all office work.

What stands out in this table: the requirements are almost never technical. They're about whether your documents exist in digital form, whether it's clear which version applies, and whether someone decides what happens in case of doubt. Most of the work in an AI initiative is work on order, not on technology. It's also the work that pays off regardless of which tool you end up using.

What do you need in terms of data, rights, and ownership?

Three things that have to be settled before you start, because otherwise they'll stop the initiative mid-flight.

• Data situation. Which documents are needed, where they sit, who maintains them. A repository with five versions of the same price sheet reliably produces wrong answers, with any tool you choose.
• Legal framework. Where the data is processed, what happens with personal information, what your customer contracts say about sharing documents. For companies in the EU, add the obligations under the AI Act, which vary in strictness depending on the application. For the typical office cases above, they usually amount to transparency and training requirements. That's an afternoon's work with your data protection officer, not a reason to wait.
• Ownership. One person who owns the initiative and one who implements it technically. If they're the same person and that person sits in IT, the initiative turns into an IT project that the business unit watches from the sidelines.

What you don't need: a dedicated department, your own trained model, or a complete data strategy up front. These three answers come up regularly from vendors, and they add quarters to the path toward the first demonstrable benefit.

How do you run a pilot you can actually learn from?

With a measurement defined beforehand and a time frame short enough to support a decision within the current year. Six to ten weeks is realistic for an office case.

Three numbers belong on paper before you start: how long the step takes today, how often it occurs, and how frequently the result is wrong today or has to be redone. Without these baselines, every statement afterward is a matter of taste. And in practice, the opinion of the person with the most influence wins, not the result.

The pilot has to run on real work, not on examples. A test run with hand-picked documents shows that the technology works. Only the run with the real inbox shows how many cases are ambiguous, and that's the number the business case hinges on.

And from day one, it needs a rule for the gray areas: who checks, what happens when there's doubt, how a mistake gets reported. That rule is what turns an experiment into an operation.

Why do so many pilot projects stay pilot projects?

Because nobody owns the move into daily operations. A pilot has a sponsor, a timeframe and attention. Operations have none of that, as long as nobody has decided that the old way ends.

Three patterns keep repeating:

• The old process keeps running. As long as both paths are open, people under pressure use the one they know. A pilot without a shutdown date for the old step is a parallel world.
• No budget for year two. Pilots get paid out of leftover funds; operations need a line in the plan. If you don't settle that at the start, you'll negotiate it at the worst possible moment: when the pilot works and everyone takes it for granted.
• Nobody owns quality. Answers get worse as documents go out of date. Without one person maintaining the foundation, quality drops slowly and quietly, until somebody says the whole thing isn't worth it.

Each of these three points costs a decision, not money. They belong on the same single page as the use case, the one that holds your AI strategy. And if they're missing there, the initiative is worth no more than the next vendor meeting.