The question we get most often isn’t “which model should we use”, but “where do we start”. That’s the right question: in most of the companies we work with the bottleneck isn’t the technology, it’s figuring out which piece of work is worth automating first.
Start from the process, not the model
A process is a good candidate when it has three properties:
- Volume: it repeats often enough to make the saving measurable.
- Implicit rules: the people running it follow criteria they can explain, even if those criteria are written down nowhere.
- Verifiable output: there is a way to tell whether the result is correct, without waiting months.
If the third one is missing, the project isn’t ready: with no way to verify, you can’t measure quality — and what you can’t measure, you can’t improve.
What changes in practice
Value doesn’t come from the largest model, it comes from the most precise insertion point.
An assistant that answers everything produces brilliant demos and little else. A system that covers one specific step — classifying an incoming request, extracting data from a document, preparing a draft — reaches production, gets used every day, and can be measured.
The three phases we follow
- Assessment: we map the candidate processes and estimate the impact.
- Pilot: one process only, in a real environment, with metrics defined up front.
- Rollout: we extend only what the pilot has proven.
Where governance fits
The AI Act isn’t paperwork to be postponed to the end. Risk classification and technical documentation are far easier to build while the system is being designed, rather than reconstructed after the fact.
If you want to work out which process to start from, the assessment is the entry point: two weeks, a clear scope and an impact estimate before any development.