What AI can actually do for a small business (and what it can't)
Strip out the hype and there is a short list of things AI does reliably well today, a shorter list it does badly, and a middle ground where it works only with a human in the loop.
We get asked to "add AI" about once a week. Roughly a third of the time there is a genuinely good use case, a third of the time there is a better non-AI solution, and a third of the time the honest answer is that it will not work well enough to trust.
Here is how we sort them.
Where it reliably works today
Answering questions from documents you already have
If your team repeatedly answers the same questions from a policy manual, a product catalogue or a pile of PDFs, an assistant grounded in those documents works well. The important word is grounded: the model retrieves your actual text and answers from it, rather than recalling something vaguely similar from training. Done properly it can cite the source line, which is what makes it trustworthy.
Turning messy input into structured data
Invoices, delivery notes, CVs, forms photographed on a phone. Pulling fields out of unstructured documents is unglamorous and it is where most of the real return on investment sits. It is also easy to measure, because the output either matches the document or it does not.
Drafting the first version
Product descriptions, listing copy, reply templates, summaries of long threads. Not the final version. The first one, so a human starts from something rather than nothing.
Triage and routing
Deciding which of six queues a ticket belongs in, flagging the urgent ones, spotting the angry ones. A wrong guess here is cheap to correct, which is exactly what makes it a safe place to automate.
Where it needs a human in the loop
Anything where being confidently wrong is expensive. Pricing decisions, medical or legal wording, anything that goes to a customer under your name without review, anything that touches money.
These can still be automated — but as a recommendation a person accepts, not an action the system takes alone. Build the accept button. It costs a day and it is the difference between a tool people keep using and one they quietly turn off.
Where it currently does not work well enough
- Arithmetic and reconciliation. Use a database or a spreadsheet. This is a solved problem and language models are the wrong tool.
- Anything needing a guaranteed answer every time. If "usually correct" is not acceptable, you want deterministic code.
- Knowing about your business without being told. A model knows nothing about your stock levels or your customers unless you connect it to that data.
- Replacing a person entirely. Every deployment we have seen that worked made someone faster. The ones aimed at removing the someone got switched off.
A test before you spend anything
Before scoping an AI project, we ask a client to answer three things:
- What does this cost us today? In hours per week, or in errors per month. If nobody can say, there is no way to know whether the project worked.
- What does a wrong answer cost? This decides whether a human sits in the loop.
- Where does the data live? If the knowledge is in one person's head rather than a system, the first project is writing it down, not building software.
If those three questions have clear answers, an AI project is usually worth doing. If they do not, the project is not ready — and no amount of good engineering fixes that.
Start narrow
The successful deployments we have been part of all started with one workflow, one team, and a number to beat. The failures started with "an AI assistant for the whole company." Pick the thing your team complains about most and automate that.
