5 min read
What agentic AI can do today, reliably and in production: recognise and file incoming documents, watch expiry dates and deadlines, assemble dossiers, and draft follow-up mails or tasks — each time with a human who approves. What it cannot do: replace the general meeting, take legal decisions, or work well in an office where the documents themselves are still chaos.
Anything that carries judgement, mandate or liability: deciding on works, interpreting votes, settling disputes. And — the most underestimated one — AI in an unstructured office. An agent that has to scrape its context together from mailboxes and loose folders is guessing. Hence the sober figures: only 22% of Belgian companies feel ready for agentic AI and 4% have rolled it out (AWS/Strand Partners, 2026), and Gartner expects 40% of agentic AI projects to be cancelled before the end of 2027.
Pilots stall for almost always the same reason: they start with the AI instead of with the memory. The model gets no reliable context, and its output is rightly distrusted. The order that does work is dull and proven: day 1 the document foundation live, week 1 systems and people connected, month 1 the first workflow under human supervision. One workflow, measurable ROI within 30 days, and only then the next. That is how it runs today at 14 organisations, from 5 to 3,000 users, without losing a single client. You approve. They do the rest.