Where AI pays off first in a small business

Start with the hours, not the model

Most conversations about AI in a small business start in the wrong place. They start with a tool someone saw a demo of, and work backwards to a use. The ones that pay off start with a question an owner can answer in a minute: where do your people spend hours every week doing something a careful new hire could do from written instructions?

That is the test we use. AI pays off first where the work is repeated, the inputs are already in a system, and a wrong answer is cheap to catch. Here is where that tends to be true, in the order we would look.

1. Reading and typing that nobody wanted to do

Every business has a pile of documents that a person turns into fields. Supplier invoices matched to purchase orders. Order emails keyed into the system. Forms copied into a spreadsheet. This is the first place to look because the job is already defined, the examples already exist in last year's inbox, and the output is easy to check: either the invoice number matches or it does not.

A model that reads an invoice and fills the fields, with a person approving the mismatches, saves hours in the first month. It also gives you something to measure. If you cannot say how many documents a week and how long each takes, start by counting. The count alone sometimes changes the decision.

2. The question you answer forty times a day

Delivery status, opening hours, whether an item is in stock, how to return something, what the fee is. If a person types the same answer into WhatsApp or email many times a day, a chatbot built on your own documents can take the first pass and hand over when it is unsure. The condition is that the answers exist somewhere in writing. If they live only in one employee's head, write them down first. That document is worth having even if the chatbot never ships.

3. The forecast you make by feel

Owners already forecast. They order stock for next week, roster staff for Saturday and decide when to run a promotion, mostly by feel and last year's memory. If the sales record exists, a model can do the same job with more of the past in view. Ethos has recorded every bill in each outlet it serves since 2013, and that record is exactly the raw material such a forecast needs. The forecast is only ever as good as the record. If the till has not been counting stock by recipe, the model cannot invent what left the kitchen.

4. The draft, not the decision

Product descriptions, quotation covers, the first reply to an enquiry, a summary of a long thread. Language models are good at a first draft that a person then edits. They are poor at being the last word. Keep the person on the send button and the saving is real. Remove them, and the first embarrassing email cancels a year of gains.

Where it does not pay off first

Not in the thing your business is best at. If your edge is a buyer's eye for fabric or a chef's palate, do not start by automating it. Not where there is no data: a model cannot forecast a shop that has been open four months. Not where a mistake is expensive and hard to reverse, such as pricing, credit or anything regulatory, until a person is reliably in the loop. And not for the sake of being seen to use AI. We have written separately about when not to use it at all.

How to choose the first one

Four checks, in order.

• Count the hours. The job must repeat. The cost of building is recovered only when it does.
• Find the data. The documents, messages or bills must already be in a system you can reach.
• Price the mistake. A wrong output must be reversible, or must wait for a person to approve it.
• Gather thirty real examples. Thirty past cases are enough to test against before anything is built. If thirty cannot be found, the job is rarer than it felt.

Our own rule when we build this kind of thing is that it runs in shadow mode first, alongside the person doing the job, before it acts alone. The model is the last thing chosen, after the job, the limits and the hand-off points have been agreed on paper.

What a first engagement looks like

A short piece of consulting: a few days with the people who do the work, a written list of the candidate jobs ranked by hours and risk, and a recommendation for one of them, with the data check already done. Sometimes the honest recommendation is a rule, a report or a cleaner spreadsheet, and no AI at all. We would rather say that than sell a model.

If you want that ranked list for your business, our AI consulting page explains what it covers and how it starts.