AI strategy: from ambition to a first decision

Anchor the strategy in business goals

An AI strategy is a set of choices about where AI should change how work is done, and where it should not. Goals such as efficiency or better service become strategic only when they are tied to concrete outcomes: shorter lead times for quotes, fewer open queries in procurement, faster preparation of management reporting. Each of these outcomes points to accountable owners and to a measure that already exists or can be introduced.

This view also makes deliberate non-decisions visible. Some processes benefit more from conventional automation, better data maintenance or a changed approval rule. Recording that openly protects the budget for workflows where AI can make a distinct contribution, for example in handling variation in documents, language or incomplete information.

Choose workflows, not slogans

A strategy becomes testable at the level of the workflow. “AI in customer service” is a theme; “prepare an evidence-backed reply to a warranty claim” is a workflow with inputs, a result and a recipient. Three questions help for each candidate: does the effort recur, can a specialist judge the result, and what happens if an output is wrong?

Viewed this way, the result is usually a short list rather than a single favourite. Some candidates promise high value but have uncertain data access; others are easy to test but of limited importance. A portfolio of two or three candidates, ordered by evidence and risk, is more robust than a large programme that depends on a single assumption.

Clarify data, accountability and rules early

The NIST AI Risk Management Framework organises its core into four functions: Govern, Map, Measure and Manage. Mapping begins with the context in which a system will be used. For a strategy, that means recording for each workflow which data sources are involved, who may access them, which decisions remain with people and who answers for the result.

Regulation belongs in the same step. The EU AI Act sets risk-based rules for specific uses of AI; systems that can seriously affect health, safety or fundamental rights are classified as high-risk. Whether a planned workflow touches such a category, which contractual or data-protection requirements apply and whether data must stay on your own infrastructure all shape which operating models are realistic: local, private or hybrid.

Sources: NIST · AI RMF CoreEuropean Commission · AI Act

Let the first pilot test the strategy

A strategy remains a hypothesis until it meets real work. A bounded first deployment, such as an Agent Pilot for one workflow, tests more than a model: it shows whether data access works, whether reviewers can judge results efficiently and whether the benefit survives rework and handovers. Before the start, define which findings would justify expansion, which suggest a narrower scope and which mean stopping.

After the pilot, review the strategy, not just the prototype. Did the assumptions about data, effort and acceptance hold? Which prerequisites, such as permissions, interfaces or review capacity, need investment before the next workflow follows? Updated in this way, the strategy becomes a sequence of informed decisions rather than a document that is outdated on completion.