The wrong question
The usual question is "what can AI do for my business?" It gives you a list of forty possibilities and you act on none of them.
Ask a different one. Where did last week actually go?
Keep a rough tally for five working days. Not a time sheet — a note on your phone each time you switch task. Owners are usually surprised by two things: how much of the week is made of words (reading messages, writing replies, explaining the same thing to the eleventh person), and how little of it is the work they think they do.
Three filters
Put every item on that list through three questions.
Do you do it more than twice a week? Savings are per-instance, so frequency is the whole game. Ten minutes saved on something daily is about 40 hours a year. Two hours saved on something you do every December is two hours.
Is it made of words? These tools read and write text. They sort, summarise, draft, translate, extract, rephrase. They do not lift boxes, chase a supplier who is ignoring you, or know that Mrs Okafor always wants her order left with the neighbour.
If it came back wrong, would you notice cheaply? You will be checking the output. If checking takes as long as doing, you saved nothing. If a mistake is silent and expensive — a wrong tax figure, a wrong dimension on a fabrication order — the checking has to be so careful that the tool stops paying.
A task that passes all three is a real candidate. A task that fails one is not, however good the demo looked.
What this looks like
Three shapes of business:
- A tailor in Lagos gets 30–40 WhatsApp messages a day. Twenty-five of them are "how much for a two-piece?", "how long?", "do you deliver to Ikeja?" Frequent, text, obvious when wrong. Strong candidate.
- A physiotherapy clinic in Manila writes a short visit summary after each appointment. Frequent and text-shaped — but it is a clinical record, where a wrong detail is silent and serious. Candidate with a hard rule attached: the therapist writes the clinical content, the tool tidies the language.
- A three-person agency in Bogotá writes six proposals a month. Frequent enough, text-shaped, easy to check. Strong candidate — as long as the price and the dates never come out of the model. Lesson three covers that.
The arithmetic, before you spend anything
Do this on paper.
Hours the task takes now, per week. Hours you honestly think it will take with help — including the editing, which is where new users always undercount. Multiply the difference by 48 weeks. Then price your own hour: not what you charge a client, but what you would pay someone to take that hour off you.
A worked case. Answering enquiries takes 5 hours a week and drops to 3. That is 2 hours a week, 96 hours a year. If your hour is worth ₹600, that is roughly ₹57,000 of time against a tool costing perhaps ₹24,000 a year. Clear yes. Run the same sum on "writing my annual capability statement" and you get 4 hours saved, once. Clear no — and it was the more impressive demo.
What to leave alone
Deliberately ignore:
- Anything you do a few times a year, however painful.
- Anything a cheaper, dumber tool already fixes. A saved reply in WhatsApp Business costs nothing and answers "what time do you open" perfectly, forever. Lesson eight comes back to this.
- Anything where you cannot check the answer. If you do not know enough about your local tax rules to spot a wrong answer, a confident wrong answer is worse than no answer.
- Anything that is really a people problem. No tool fixes a supplier who does not call back.
Pick two
Two tasks. Three at the very most. One person, one tool, for the first month.
That is not caution for its own sake. It is that you only learn what these tools are like on *your* work by using one on your work, repeatedly, until you can predict what it will get wrong. That prediction is the asset. Eight half-tried tools buy you eight shallow impressions and no judgment at all.
Before you move on