KaptiMethod
How we keep up with AI without chasing every announcement
Every week brings a new model, a new tool, a new rule. Reading everything is impossible, ignoring everything is dangerous. Our answer comes down to four steps: sort, test, share, decide. Here’s how it works in practice within the collective.
Being an “AI expert” in 2026 doesn’t mean knowing everything. It means knowing, at any given moment, what has changed for the products we build, and what hasn’t. To do that, we have a simple routine, shared across several people, and fed by our client work rather than by social media.
1. Filter: few sources, but the right ones
We start with primary sources: release notes and documentation from model providers, the research papers that accompany a release, and official texts for regulation (the Official Journal of the European Union for the AI Act, the Commission’s guidelines, CNIL publications).
Everything else, we read with one question in mind: does this change anything for a screen, a cost or a risk for our clients? A spectacular demo that answers none of the three takes a back seat.
2. Test: on a real case, not a demo
A new development that looks useful is tested on a concrete case, often drawn from our internal projects : summarizing a customer history, reading an invoice, classifying messages. We compare it with what we used before, on the same data, and note three things:
- quality : is the result better, and on which cases does it get things wrong;
- cost and speed : what it would mean at the scale of a product in production;
- constraints : where the data goes, what the provider keeps, what it implies for the AI Act.
These trials are short, usually a few hours. Their value lies in being comparable from one to the next.
3. Share: what each of us sees on engagements
The most useful insights don’t come from announcements, but from the field. Each member of the core team works on different products: what a developer discovers while integrating a model for a client, or what a designer observes during a user test, is shared with the others. We keep a written record of the lessons learned, so the collective never has to learn the same thing twice.
The collective’s get-togethers (the KaptiEvent, the hackathon, the Tokyo workshop) serve the same purpose: stepping back together on what has changed.
4. Decide: what makes it into our methods
One successful test isn’t enough. A practice makes it into our engagements once it has proven itself on several cases, and once we can explain its limits to a client. That’s how our interface rules for AI features came about, detailed in another article : show uncertainty where it exists, link every suggestion to its source, leave the final say to the human.
Conversely, we also retire practices. When a new tool makes a step unnecessary, we remove it from our services rather than keep billing for it out of habit.
Regulation, tracked like a feature
The AI Act and the European Accessibility Act are treated as product topics, not just legal ones. Every change to the timeline or guidelines is reviewed from the screen’s point of view: what needs to be displayed, when, and how to make it accessible. Our readings are published here, dated, so it’s always clear which version of the text they refer to.
What this means for you. When we recommend a model, a tool or an interface rule, it’s because we’ve already tried it on a case close to yours, and we can tell you where it reaches its limits.
Frequently asked questions
How do you choose an AI model for a client?
We compare several models on a sample of the client’s data, against three criteria: answer quality, cost at scale, and confidentiality and hosting constraints.
Are your recommendations up to date with the AI Act?
We follow the official texts and the Commission’s guidelines, and our articles are always dated. For a specific case, we also recommend seeking legal advice.