"We should be doing something with AI" is now a standing item in most management meetings. It is also the wrong sentence. It names a technology instead of a problem, which is why so many SME AI projects start with a customer-facing chatbot, disappoint everyone, and poison the well for the projects that would have worked.
Why the chatbot is usually the wrong first move
A public chatbot is the most visible AI project, and the worst-shaped first one: it faces customers, so mistakes are public. It needs your knowledge base to be accurate and current, which it rarely is. And its value is hard to measure, so six months later nobody can say whether it worked.
Strong first projects have the opposite shape: internal, low-blast-radius, and measurable in hours saved or errors prevented.
The pattern behind good first projects
Across the businesses we audit, the AI work that pays back fastest sits in a handful of repeating patterns:
- Reading and routing. Inboxes, forms and documents that a person currently reads, classifies and forwards. Quotes requests, supplier invoices, CVs, support mail. This is the single most common win because the volume is daily and the rules are learnable.
- Drafting against a template. Quotes, proposals, job ads, follow-up emails, report sections. The person stays in the loop as editor instead of author. Time per document drops by half or more.
- Extraction. Pulling structured data out of unstructured documents: invoices into the accounting system, delivery notes into the stock system. Eliminates retyping and the errors that come with it.
- Summarising the backlog. Meeting recordings, long email threads, tender documents. Valuable wherever someone senior spends hours reading so they can make a five-minute decision.
Notice what these have in common: none of them are customer-facing, all of them replace a measurable number of hours, and all of them keep a human approving the output.
Rank by payback, not excitement
For each candidate process, you need only four estimates to rank it:
- Hours per month the process currently consumes, times the loaded cost of the people doing it.
- Error cost: what a mistake in this process costs when it happens, times how often it happens.
- Implementation cost: tooling plus setup time. For most of the patterns above this is hundreds of dollars and days, not thousands and months.
- Disruption risk: what breaks while you switch over, and who needs retraining.
Divide expected monthly saving into implementation cost and you get payback in months. In our audit data, the best SME first projects pay back in under three months. Anything over twelve should wait, however exciting it sounds.
The readiness questions that decide success
Two businesses can pick the same project and get opposite results. The difference is usually readiness, not the tool:
- Is the underlying data accessible? If your processes live in one person's head or a drawer of paper, digitise first. AI amplifies what exists.
- Does someone own it? A project without a named internal owner stalls at the first edge case.
- Is there a rule for exceptions? The win is automating the routine 80 percent and routing the odd 20 percent to a person, cleanly.
- Do you have a policy? If staff will feed business data into AI tools, decide which tools and what data before scaling up, not after the first incident.
Start smaller than feels impressive
The honest version of "doing something with AI" is one boring internal process, automated well, measured before and after, finished in a month. That win funds and justifies the second project, and by the third one you have something most competitors do not: an organisation that knows how to adopt this technology instead of one that bought a chatbot.