Where AI actually pays off in a small business (and where it doesn't)

The filter: four questions before any AI project

Run every idea through these. If any answer is no, do not start.

1. Is the job repetitive AND high-volume? AI pays back on frequency. A task you do twice a month is not worth the setup, however annoying it is.

2. Is a good-enough answer acceptable? If the output must be right 100% of the time with no human check, you are building a system where a human reviews everything: which is the old cost plus a new one.

3. Can you tell good output from bad in seconds? If judging the answer takes as long as doing the job, you have moved the work, not removed it.

4. Do you have the input already? AI on data you would have to go and collect is two projects, and the first one is the hard one.

What reliably pays back

Drafting where a human edits. First drafts of replies, descriptions, summaries. The economics work because editing is faster than writing, and the human stays in the loop where the judgement is.

Classification and routing. Which bucket does this belong in: support ticket, lead, expense. High volume, tolerant of an occasional miss, instantly checkable.

Extraction. Pulling structured fields out of unstructured input: invoices, emails, forms. This is the least glamorous and most consistently profitable one.

Search over your own material. Answering questions from documents you already have. Works because the source of truth exists and can be cited.

The pattern across all four: the human keeps the judgement, AI takes the typing.

What quietly does not

Anything where being wrong is expensive and hard to notice. Financial figures, legal wording, medical anything, promises to customers. The failure is silent, and silent failure at volume is how you find out in a complaint.

Strategy. It will produce a confident, plausible, average plan. Average is exactly what you cannot afford, and it reads well enough that you might ship it.

Anything where the input is your judgement. If the value is that YOU decided, automating the deciding removes the product.

Replacing a person you have not first tried to help. The pattern that works is a person plus AI doing three times the work. The pattern that fails is AI alone doing 60% of the work and someone finding out in month three.

The two-week pilot

Do not build. Do not integrate. Do not buy a platform.

Week 1: do the job by hand, with AI, in a chat window. You, pasting inputs, reading outputs, keeping score. 20-50 real examples from your actual work: not test cases you invented.

Keep two numbers: how often was the output usable with minor edits, and how long did the whole loop take versus doing it yourself.

Week 2: if week 1 cleared the bar, do it again with someone else on the team doing the pasting. This is where most pilots die: it worked for you because you knew what good looked like and unconsciously fixed the prompts.

The bar: usable output 70%+ of the time, and at least half the time saved. Below that, automating it makes the problem faster, not smaller.

The cost nobody budgets for

The API bill is the small number. The real costs:

The review layer. Someone checks the output. That is a permanent job, and it is the one that decides whether the project nets out positive.

The edge cases. The 20% it gets wrong are rarely random: they cluster on your most complex, most valuable cases, because those are the ones with the least precedent.

The drift. It worked in March. The model changed, or your inputs changed, and nobody noticed until a customer did. Whatever you build needs a check that fails loudly when quality drops: the same rule as any other automated job.

Budget the review layer from day one and the maths gets honest fast.

How to pick the first one

Not the most exciting. The most boring one that clears the filter.

List every repetitive job in the business. For each, write down: how many times a month, how long each time, and what happens if it is wrong. Multiply the first two: that is your annual hours. Sort by it.

Start at the top of that list, skip anything where wrong is expensive, and pilot the first survivor.

It will be something unglamorous like tagging inbound leads or extracting fields from supplier emails. That is the right answer. The unglamorous one pays for the ambitious one.