What AI can and cannot do for a small business

We sell automation systems, some of which use language models. That's exactly why this sort is worth writing: a client who buys AI where it does nothing becomes an unhappy client, and an unhappy client costs more than a declined contract.

Where it genuinely works

• Drafting a first version — a reply to a customer, a description, a summary. You correct it, you stay in control, and you skip the blank page.
• Sorting and routing — classifying an inbound request by urgency or type and sending it to the right person.
• Summarizing — condensing a long exchange or a call into three points before calling back.
• Extracting — pulling the useful data out of a badly structured document or email.

What those four have in common: a human sees the output before it matters, and a mistake is visible immediately.

Where it doesn't work

• Setting a price. A model will produce a plausible amount, which is exactly the danger: plausible isn't correct, and nobody will check.
• Committing on the company's behalf. A deadline, a warranty, an availability — a model doesn't know your real capacity next Thursday.
• Giving regulated advice. Tax, legal, medical. The risk isn't the error: it's your liability.
• Replacing missing data. If the information exists nowhere in your business, no model will invent it correctly — it will simply invent it.

Spotting a project that will fail

• Nobody can say what decision the output will change.
• There's no way to check whether the output is right.
• The information needed doesn't exist in the business yet.
• The word "AI" appears in the proposal more often than the word "problem".

Will AI replace my employees?

In a service SME, almost never — because the bottleneck isn't production, it's coordination. What disappears are tasks: retyping, searching, chasing. The businesses that gain most are the ones that give that time back to what the machine can't do, which is talking to customers.

Should we wait for the technology to settle?

For the uses listed above, no: they already work and don't depend on the latest model release. For everything else, waiting rarely costs much. The real risk isn't arriving late, it's paying today for a promise that hasn't materialized yet.