AI in an SMB: where it pays off and where it is wasted money

The pressure to "do something with AI" is real, and it leads many companies to start in the wrong place. The good news is that there is a simple rule for separating use cases that pay off from those that do not.

The rule: is the input structured or not?

If the data already sits in fields — quantities, dates, references, amounts — you do not need AI. You need an integration or a rule. It is cheaper, faster and vastly more reliable.

AI earns its keep when the input is messy: a scanned PDF, an email written by a person, a 40-page contract, a complaint in free text. That is where it does what rules cannot.

Three cases that usually pay off

Document extraction. Supplier invoices, delivery notes, forms. If you receive hundreds a month as PDFs and someone types them in by hand, the return is almost always clear.

Search over internal documentation. Procedures, manuals, contracts, technical history. An assistant that answers with the source cited saves hours of searching — and reduces dependence on the one person who knows where everything is.

Repetitive commercial content. Proposals and responses where 70% of the content is always the same. People review instead of writing from scratch.

Two that rarely pay off yet

Forecasting on thin data. Demand forecasting models need consistent history. Most SMBs do not have it at the required quality, and the result is misplaced confidence in a weak forecast.

Customer-facing chatbots with no solid base. An assistant built on outdated documentation creates more problems than it solves. Fix the knowledge base first.

How to evaluate a pilot honestly

Decide before you start: what accuracy rate makes this useful, how many hours it has to save per month to justify the cost, and who reviews the uncertain cases. If you cannot answer all three, you are not ready for a pilot — you are ready for a conversation about the process.