"How can we take advantage of AI?" That's what an SME asked me last year. I started where I always start: a diagnosis. The business, the processes, the real pain points before any technology discussion. Verdict: two high-impact projects, zero AI. Both were prioritised with the teams, then delivered. They are now part of their daily operations.
AI takes up most of my engagements and I am the CTO of a SaaS platform. My conviction hasn't changed: AI is one technological lever among others. The rare skill in 2026 isn't knowing how to put it everywhere. It's spotting early that a need presented as "an AI need" is first a process or data problem.
The grid: four criteria, in order
So every idea goes through the same grid. Four criteria, in order. A single one that fails and the project doesn't go ahead.
- Is the problem already expensive today, without AI? A sign: someone handles it by hand every week and it keeps breaking.
- Would the problem survive a well-organised, well-tooled process? In other words: is there a deterministic rule that covers most cases? If so, AI is a luxury.
- Does the gain hold up once everything is counted? Integration, change management and maintenance included.
- How much deviation can the deliverable accept? A draft reviewed internally: a lot. An official document sent to a client: none.
The case: lab results that gate the job sites
This SME takes samples on its clients' job sites and sends them to analysis laboratories. Results come back overnight by email and determine when work can resume, often before 7am. One hour of delay and an entire crew can't work.
Criterion 1, comfortably met. The whole business went through a single address: several hundred emails a day, client requests, sales exchanges, quotes, final reports and, buried in there, the overnight results. Everyone was copied on everything. Early each morning, one person opened the results one by one, identified the client, checked they matched the right sample, then forwarded them. If that person was away or late, everyone downstream waited. The clients had the numbers of the sales reps and the managers. They called them directly to chase their results.
Criterion 2: it fails. Each result maps to a sample and each sample to a client. The rule is strict, with no exception. The AI project stops there. Criteria 3 and 4 never had to be asked: one that fails is enough.
What was delivered instead
- First the reorganisation: shared mailboxes, created and structured, with migration and team training. Every flow now has its place and its owner.
- Sending results to clients could then be automated: each analysis is cross-checked against client data and samples, then sent to the right people, every day, with no intervention. The tedious morning sorting is gone, and so are the chasing calls.
The symptom, the cause, the stake
No AI consultant sells a shared mailbox. The symptom was an email, the cause was organisational, the stake was commercial. Addressing only one of those layers would have solved nothing. That's the ordinary situation of an SME: no one inside has the mandate to look at the problem end to end.
Do you have AI project ideas in mind but don't know how to get them started? They can go through this grid. Write to me and we'll talk it through.
