The method

The tool doesn't decide whether AI works. The approach does.

Buying a tool is easy. An AI you can trust with real work comes down to how you go about it. Below is that approach, and what you can hold any system to, including someone else's.

Book an intro call See the offer

Four fields, one standard.

Education and learning materialsConsultanciesPublishers and editorial teamsResearch and think tanksTeachFoundryBookaflowValoraEducation and learning materialsConsultanciesPublishers and editorial teamsResearch and think tanksTeachFoundryBookaflowValora

Analyze first, build second

Most AI projects don't fail on the model. They fail on the step that gets skipped before it: understanding where the work actually gets stuck.

So we start by mapping how your knowledge work runs. Where the thinking sits, where the busywork sits, and what that busywork costs in hours and euros. Only once the problem is sharp does the choice of approach follow. Sometimes the outcome is that an existing tool will do, and sometimes that AI adds nothing here. That's an honest answer too.

The dividing line: verifiable or not

The danger of AI in knowledge work isn't that it makes mistakes. It's that a wrong answer looks convincing, and so slips past the check.

So the line between AI that helps and AI that harms doesn't run between creative and analytical work, and not between simple and complex either. It runs between verifiable and unverifiable. The only question that counts: can every claim be traced back to where it came from? Where it can, AI is a powerful tool. Where it can't, a convincing answer is more dangerous than a weak one.

Everything we build or advise sits on that side of the line: checkable, not taken on faith.

TRACES, the six layers

A trustworthy system is more than a clever model with an input box. It has six layers, and you can hold any supplier to every one of them. Where a layer is missing, there's a risk, and nearly every AI incident you've heard of traces back to a missing layer.

TRACES

These six questions are the yardstick for everything we make. And if you're judging a supplier yourself, take them with you: a system that can't answer one of them is out.

Take the supplier scan

Trail

Wat is vastgelegd: bron, bewering, wie keurde goed, wanneer. En het spoor ontstaat als bijproduct van het werk, niet als administratie die iemand moet bijhouden.

Retrieval

Het systeem antwoordt alleen uit een afgebakende bronverzameling, met verwijzing, en zegt dat het antwoord ontbreekt in plaats van iets te verzinnen. En het systeem kan tonen welke beweringen géén bron hebben, want zoeken alleen kan afwezigheid nooit aantonen.

Approval

Een mens bevestigt inhoudelijk, gedifferentieerd naar het soort wijziging, en de bevestiging wordt vastgelegd.

Constraints

Gevoelige handelingen liggen technisch buiten het model, in vaste regels. Dat het systeem iets níét kan, is sterker dan dat iemand het achteraf nakijkt.

Expertise

Ontwerpkeuze: het oordeel blijft bij de vakinhoudelijke mens, en het systeem maakt dat oordeel lichter in plaats van overbodig.

Storage

Waar staat de data, wat gebeurt ermee, en staat dat contractueel vast.

The human stays in control

The AI does the routine work: searching, sorting, preparing, drafting first versions. Anything that goes out, or that matters, passes a human first.

It's also a design choice with long-term consequences. A system can take over the routine work and keep judgement sharp, or quietly take over the thinking. We deliberately build for the first, so that in five years your team is more capable, not more dependent.

Measuring instead of believing

Whether AI helps isn't a matter of feeling. In practice, AI feels faster than the clock says it is, and that time then leaks away into checking and repairing.

So we measure before and after. Honestly timing a task for a week, with and without AI, says more than any impression can. That's how you know whether a project really delivers, and not just whether it feels that way.

Buy, build, or neither

The honest order is always the same: first the analysis, then the choice. For routine work that looks the same everywhere, a good existing tool usually wins. Where source obligations, sector rules and confidential material meet, off-the-shelf tools fall short, and that's where we build the custom layer: the search across your own sources, the citation, the sign-off.

And sometimes the answer is neither: a simpler process, or nothing at all. Hearing after a thorough analysis that AI adds nothing is just as valuable an answer as having something built.

Start small, go far.

Want to look for yourself first? The supplier scan and the Chain check are free and need no conversation.

Free supplier scan

Read on

Want to hold a system or supplier up to this straight away? The free supplier scan does exactly that.

Take the supplier scan

Here's what that looks like in practice.

Three different worlds, one principle: the AI does the work, the human checks and approves.

Read the case