AI strategy in mid-sized companies: why the tool question is the wrong first question

What is an AI strategy, and what is it not?

An AI strategy answers where AI should make a difference in your company, who owns it, and how you'll know it's working. That sounds narrow, and it's considerably more than what's in most papers on the subject.

It is not a list of tools. Which model, which vendor, which license: these commitments have a half-life of months, and they're reversible. Start there and you run the discussion at the point where a wrong answer costs the least. And where a right one delivers the least.

The real question has shifted. It's no longer whether an organization uses AI, but how it decides when systems deliver contradictory, ambivalent, or uncomfortable results. That makes the use of AI in strategy work a non-technical question. It touches your understanding of leadership, your logic of accountability, and how deeply a decision is anchored in the organization.

Why do AI initiatives in strategy work almost never fail because of the technology?

Because the technology is the easiest part. Connecting a model that summarizes proposals or condenses market reports is routine work today. What's missing is the interplay: without a clear architecture of strategic intelligence, AI remains an isolated analysis tool. Powerful and ineffective.

The pattern behind it is older than any language model. In conversations with management teams and family businesses, three symptoms show up that existed long before AI and that AI amplifies when no one addresses them:

• Data overload instead of clarity. Every additional analysis opens up perspectives that raise new questions. "Let's run one more scenario" is the sentence that gives it away. More information leads to less clarity, and an AI system delivers information faster than any team can process it.
• Strategy drowns in day-to-day business. Meetings, projects, and escalations push strategic thinking out of everyday work. Reflection happens at the offsite, and then not again for a year.
• The bottleneck is no longer external. Strategy work used to be seen as a one-off intervention where outside specialists delivered a quality that wasn't permanently available in-house. Today the bottleneck is the ability to make good decisions consistently, even without outside support.

Start an AI initiative without looking at these three points and you automate the symptoms. The result is a pilot that works technically and decides nothing.

Which four levels do you need for AI to have strategic impact?

Working with leadership teams, four levels have emerged that need to work together. Each one delivers something on its own. The strategic impact comes from the interplay.

Human judgment runs across all four levels. Strategic decisions aren't optimization problems with one objectively correct solution. They depend on values, on risk appetite, and on ambition, on factors that can't be calculated. The more powerful the systems become, the higher the demand on leadership. AI can analyze and simulate. It cannot integrate.

Which decisions should a system prepare, and which shouldn't it?

This distinction is the most practical part of the whole topic, and it's almost never made. Some of your decisions are recurring, data-rich, and reversible: priorities in the project portfolio, adjustments to target systems, resource shifts within clear guardrails. Here a system creates real quality, because it makes implicit assumptions visible and lays out options so they can be compared.

Other decisions are strategic because they happen once, depend heavily on context, and can't be undone. Entering a market, repositioning, an acquisition, a deliberate exit. Decisions like these can't be optimized. This is where the distinction between calculable risk and true uncertainty applies, the one that goes back to Frank Knight: probabilities can't be calculated because even the structure of the problem is unclear.

That leads to a simple house rule. Define which types of decisions fall into which category before the first system goes live. For the first group, describe what the system should deliver and who sets the frame. For the second, describe what it may contribute, and put on record that the decision is made at the table, not in the tool.

What happens when the preparation gets better than the decision?

Then an effect kicks in that hardly anyone is prepared for: the better the preparation, the more visible the undecidability. A good system doesn't deliver one clear answer. It sharpens the trade-offs that used to sit in the vague.

At first that feels like a step backward. Before, there was a presentation with a recommendation; now three viable options are on the table, each with its own price tag. That's exactly the progress. The trade-offs were there before too, they were just invisible and, in case of doubt, settled by rank.

Leadership here means deciding where calculability ends. And that's the moment you need to prepare for by clarifying who decides and who bears the consequences. If that question only comes up once the system is already running, postponement usually wins.

Where do you start if you want to begin this quarter?

Not with a tool comparison. The first step costs nothing and determines everything that follows: for the next strategic decision on your table, write down which assumptions have to be true for it to work.

In conventional strategy processes, these assumptions stay unspoken. They sit in the heads of the leadership team, in the culture, in experience from the good years. That holds up as long as they're accurate. The moment they no longer apply, nobody can say what the decision was actually based on.

Behind this is an insight that is uncomfortable and reassuring at the same time: better decisions don't come from smarter people. Even experienced people are subject to bias, time pressure and group dynamics. Decision quality is a question of structure, process and discipline. That's something you can work on.

• Write down your assumptions, for every major decision, with a date and a name.
• Pick one type of decision that is recurring and reversible, and put the first system to work there. Portfolio prioritization works in almost every case.
• Define who decides when the preparation produces two equally strong options. That person needs a name, not a role.
• Set a fixed date to test your assumptions against reality. Quarterly is enough, monthly is better, never is the norm.

If after two quarters you can name which of your assumptions turned out to be wrong, the system is working. If after two quarters you have a tool in use and make the same decisions as before, it isn't.