A board's guide to agentic AI — why most strategies never reach production, and the few questions and moves that change that.
By Jens Paul Berndt — operator; former CEO & CTO behind some of Europe's largest digital marketplaces.Every board now has AI on the agenda. Most have a strategy deck. Few have results.
The gap is rarely the model. It is the operating model: the data, the decisions, the ownership and the willingness to put something into production. That is leadership work, not a vendor selection.
This guide is written from the operator's seat. It will not make you an AI expert. It gives you the few questions and moves that separate companies that talk about AI from those that quietly begin to run on it.
Generative AI produces content — text, code, images — when prompted. Agentic AI goes a step further: it acts.
An agent can plan, call tools and systems, take multiple steps and complete a task toward a goal — not just answer a question. In a business that means software that doesn't only suggest, but does: it triages a ticket and resolves it; drafts and books; reconciles and flags; researches and decides within set limits.
The opportunity is leverage. The risk is that an agent acts at scale — including when it is wrong. Both are why agentic AI is a board topic, not only an IT topic.
1 · Boiling the ocean. A grand programme instead of one shipped use case. Ambition outruns evidence.
2 · Slideware over software. A strategy with no single agent in production — so nothing is learned.
3 · No owner. A steering committee instead of one person accountable for the outcome.
4 · The data alibi. Waiting for perfect data instead of starting where it is already good enough.
5 · No metric. Success measured in pilots run, not in a P&L number moved.
1 · Start narrow. Pick one process with a clear owner and a measurable outcome.
2 · Ship to production. A live agent teaches more in a month than a strategy does in a quarter.
3 · Tie it to the P&L. If you can't name the number it moves, don't start it.
4 · Give it an owner. One accountable person — not a committee — from day one.
5 · Scale what works. Industrialise the winners; retire the rest without ceremony.
1 · Where first? Which single process will AI change first — and who owns the result?
2 · Good enough? What is our data actually good enough for today?
3 · When it's wrong? What happens — to customers and to liability — when an agent gets it wrong?
4 · How measured? Will we measure value in pilots, or in a P&L number?
5 · Whose judgement? Do we have operator-level judgement in the room, not only vendors?
Days 0–30 · Choose. Pick one use case tied to a P&L metric, name an owner, confirm the data is good enough.
Days 30–60 · Ship. Put one agent into production behind guardrails and measure it against the metric.
Days 60–90 · Decide. Scale, adjust or stop — and set light governance for the next two use cases.
The point is not speed for its own sake. It is learning in contact with reality — cheaply — before betting big.
How BERNDT helps: BERNDT Build takes companies down exactly this path, and BERNDT Board brings these questions into the boardroom.
Discuss what this means for your company — via the contact form or at office@berndt-gmbh.com.