AI for SMEs: where to start

A Spanish SME does not need an “AI centre of excellence”. It needs three flows that save real hours this season, and a team that can run them on Monday. The rest — in-house models, unmoderated public chatbots, 18-month roadmaps — is expensive theatre.

At Kit Digital IA (Clutch Developer, Barcelona) we start with the unglamorous work: tickets, invoices, orders, internal policy. If you cannot point at the process on a whiteboard, there is no project. If you cannot measure hours or errors in 30 days, the scope is too big.

This is the start guide we send before talking about tools: what fails, which three cases pay, how to pick the first, what to leave out, and how it fits Kit Digital or Kit Consulting.

Why first AI projects stall

The pattern repeats: a tool is bought, a workshop is held, and a week later the team is back in Excel. There is no process owner, no before/after number, and too much scope. AI does not fix a broken process; it speeds it up, mistakes included.

Starting with the public website chatbot is the other classic error. Visitors notice vague answers immediately. Start inside the company, where a miss is a correction, not a review.

The third failure is treating AI as an IT project instead of a habit change. If nobody has two protected hours a week to run the flow, the pilot dies even if the model is excellent.

Three cases that pay back in weeks

• Internal policy and FAQ assistant. Manuals, hours, catalogue, agreements: answers that cite the source. People stop hunting through folders or pinging the same colleague.
• Support and email drafts. House-tone templates. A human checks and sends. Time per ticket drops; the voice does not dilute.
• Invoice or order extraction. PDF to ERP or sheet, with tax-ID, totals and VAT checks. Admin staff review instead of retyping.

All three share a design: a clear input, a structured output, a human confirmation, and a kill switch. Without that you have a demo, not a project.

A first-week walk-through

Monday: you pick “when a supplier PDF arrives, we extract lines into a draft”. You count PDFs and minutes per document. Tuesday: you lock fields (tax ID, date, net, VAT, total, number). Wednesday: you test twenty real documents, not the pretty demo invoice. Thursday: you name the reviewer and the rule when totals do not match. Friday: the flow runs in parallel with the old process. Nothing is switched off yet.

If you cannot write that sentence in a week, you are not ready to wire the ERP. You are ready for a sheet and a queue. That is still an honest start.

How to pick the first flow

Count repetition. If someone does the same task more than five times a day, it is a candidate. If a miss costs money or a customer (billing, lead times, stock), it rises. If the data is sensitive and you lack a legal basis, stop until the GDPR inventory exists.

A useful rule: the first pilot fits in one sentence. If you need a two-page diagram, cut. Do not pick the “smartest” flow; pick the one that already has an owner and volume.

What to postpone

• Training a custom model “just because”: slow, expensive, rarely the bottleneck.
• Public chatbots with no human queue and no cited sources.
• “Digital transformation” with no named owner and no concrete process.
• Wiring twelve tools at once. One measured flow, then the next.
• Seats for the whole company “to experiment”. That is noise, not a pilot.

A metric you can show in a meeting

Hours per week on the task, errors caught before the customer, response time, and the share of cases the human accepts without a rewrite. If there is no number at day 30, the pilot is badly scoped or unused. Both are fixable.

Skip vanity: workshop clicks, “tool satisfaction”, screenshots. A serious room asks what changed in last week’s work.

Fit with Spain’s SME digitalisation programmes

These starts map to advisory work (diagnose, prioritise, train) and to digitalisation solutions (site, operations, processes). The programme wants deliverables, training, and a habit that stays. We design the pilot so the team runs it without us by week three.

Do this week

• List three repetitive tasks with an owner and weekly volume.
• Pick one. Write the pilot sentence and the metric.
• Note which data it touches and whether you already have a legal basis.
• Ask for an assessment with sector, size and that task — not a wishlist.

Frequently asked questions

Do we need a data scientist on staff?

No. You need someone who owns the process and a supplier who leaves the system operable.

Should we just buy ChatGPT seats for everyone?

A licence is not a process. Without templates, sources and review you only speed up copy-paste.

How long is a first pilot?

If the data exists and the process fits in a sentence: weeks, not quarters. Indecision is what stretches it.

What if we only have Excel?

Land in a well-named sheet and a review queue. Excel does not block a start; it blocks pretending you are already a platform.

Ready for an assessment? Tell us about your SME — sector, size, and where time leaks. Related: prioritising use cases, the start checklist and the Kit Digital overview.