Skip to content
← Work
Self-initiated2026

Cotomate

An AI workflow platform for Dutch SMBs, starting from industry and role. A Dutch-language interview turns what people do into runnable, audited processes.

AIProductNext.jsPostgres

The Problem

Dutch SMB owners don't think in processes or agents. They think: I run a twelve-person accountancy, so what are my people doing that AI could help with? Every tool on offer asks them to start somewhere else. Map your process, and you must reason like an automation engineer. Build an agent, and you must already know what the agent should do.

Underneath sits a second problem. Regulated Dutch sectors carry real obligations: Wwft, the AVG, seven-year retention under the AWR, now the EU AI Act. Anything putting a model near client work has to be auditable, in Dutch, on EU infrastructure.

What It Does

The entry primitive is industry × role. A company sits in an SBI cluster such as accountancy or software, with its roles mapped to ESCO occupations. The user then talks to an AI interviewer, in Dutch, about what their day contains.

What comes out is structure: skills the role uses, processes those skills sit in, a report on what AI could take off the plate. Smaller interviews sit behind single steps and skills, so refinement is another conversation rather than a form. A finished process can then be run by hand or on a schedule, with approval points that pause it until a human signs off.

The hard part

Two firms in the same sector must never see a trace of each other. In Wwft and AVG territory that isn't a bug to fix later, it's the end of the company. The usual approach is to filter by customer in the application code, which works until the one query somebody forgets, and that is a bet on perfect discipline forever. I didn't want to make that bet, so isolation is enforced underneath the application, where forgetting isn't an option.

One firm's dataNever seen by anotherenforced beneath the appfilter in app coderely on reviewone forgotten query is enough
Fig. 1 Isolation by convention is a bet on perfect discipline, forever, by everyone. In Wwft and AVG territory that is not a bet worth making.

The other constraint is that anything putting a model near client work has to be explainable a year later, to someone who wasn't in the room. That rules out instructions buried in code and answers that vanish once they've been read. Every run leaves a record of what was asked, what came back and what it cost.

It runs on EU infrastructure, in Dutch and English, with the marketing site, the product and the admin tooling kept separate. Nothing the AI drafts reaches a customer without review.

Where It Stands

An interview prototypeApril 2026Foundations firstearly May 2026The product built outmid May 2026Quiet sinceJune 2026
Fig. 2 Nearly all of it was built inside a single month, which is both the strength and the problem.

The prototype was promoted and then deliberately held back while the foundations went in underneath it, before a single product screen existed. Ten screens followed in under a fortnight.

Pre-MVP, and I'd rather say that plainly. It deploys, in Dutch and English, and it has not launched. The interview is the piece that still has to earn its keep, and the only way to find out whether it does is to put it in front of Dutch owners rather than in front of me. That is the next milestone. The architecture is ahead of the validation, which is a comfortable position only if you say so out loud.