What belongs in an executive AI strategy, and why

An AI strategy is not a choice of model. The decisions that outlive any model: where AI pays, what it stands on, what the law expects, what changes for people.

Lasse Feldt-Rasmussen· 25 October 2026

The board has asked for the AI strategy. What you have is a set of licences, a few experiments, and a question you keep hearing: which AI is best? That question will not help you. The model you pick this year is not the one you will run next year. A strategy is the set of decisions that stay true when the model changes. Here they are.

Where it pays

AI pays where it makes you better at one of the basics of your business: serving the customers, running the processes well, deciding well, keeping the company relevant. That is the test for any AI idea. If it improves none of the four, skip it, however impressive the demo.

Choose the jobs by what they cost you today: the hours a team loses to retyping and reconciling, and the price of an error in a posting, a contract or a report. Not by what looks good on a slide.

A strategy says which of three roads each job takes:

The licences you already pay for (Copilot, Gemini or Claude), with the right permissions and guardrails.
The AI features in your ERP and your other systems.
Knowledge workers made for your own jobs that run on your own data.

Most companies use all three. The mistake is one road for every job.

Count what it costs to run, not what it costs to buy. A language model is priced by the token, a document service by the page, a licence by the seat. If you own what was built, you can maintain it without us. The invoice reader case shows the arithmetic on one job.

What it stands on

Every job needs three things: data in order, your company’s own dictionary of what things mean, and the methods the job follows. Each job reuses what the last one wrote down, so setting up the next job is easier, and running the same job again the next day is easier still. This was always true of good systems, and it matters more with AI.

Grounding, the data foundation, evals and guardrails matter more than how well anyone prompts. The model decides how well an agent reasons; these decide what it knows, what it may do and how you know it worked. Grounding means the agent answers from your records, not from memory. Evals are the tests an agent must pass before and after every change. Guardrails are the limits it cannot cross, enforced outside the model. The knowledge workers page walks through everything an agent needs.

Where the data lives and who controls access is a strategy decision. Keep the data in your own cloud, keep access under your own control, move nothing into a platform you do not own, and read the provider’s terms for what they may do with your data. The data foundation page shows what this looks like built.

What it must respect

Legislation belongs in the strategy from the start. Since February 2025, the EU AI Act has required companies that use AI to support their staff’s AI literacy. Its transparency rules have applied since August 2026, and its high-risk rules start in December 2027. The GDPR requires meaningful human involvement in a decision made solely by automation that significantly affects a person. The Danish bookkeeping act keeps management responsible for every entry, whoever or whatever made it. A strategy names which rules apply to your company and who owns each of them.

Behind the rules sits one fact: an AI cannot be held accountable. A person can be, and in your company a person always is. So the strategy includes the three zones (acts alone, asks first, the operator signs off) and a logbook of what the AI did. The How we work page has the sources.

What it changes for people

AI is a new family of tools, and it changes how people do their jobs. The bookkeeper stops typing entries and starts approving a batch and handling the exceptions. The controller stops assembling the report and starts reading it. Roles move from doing the work to instructing, inspecting and deciding. That is a management task, not an IT rollout. The strategy says who becomes an operator, the person at your company who checks and signs off the knowledge workers’ work. It also says what they do instead, and what you tell people who fear being replaced.

Everyone who operates the tools needs AI literacy, and so does every leader who signs off on the tools’ work. It is a legal duty, and it is plain sense: nobody can carry responsibility for a tool they do not understand.

Then the risk nobody puts on the slide. When AI does more, you depend more on your providers: on their uptime, on their prices, on the same model being offered next year. And your people lose, or never build, the capability to do the work when the AI is down. A strategy says what you keep the ability to do by hand, how you switch models, and what you own if a provider changes its terms.

What a strategy is not

A five-year plan. The technology will not look the same in five years; the useful question is what helps you today and prepares you for later.
A choice of model. See above.
A document. A strategy you cannot point to in a running job is a slide.
A pilot without a data foundation. It will impress once and not run twice.
An agenda set by a vendor. Every vendor sells the feature it has; none of them sells the data foundation you need.

What to do with the list

Take it to your next leadership meeting and grade your current answer against it, group by group. Start the strategy work where the honest grade is low. If you want to do that in one session with us, the Executive AI strategy page says what you leave with.

Sources

Not legal advice.