AI for a small business is a clerk, not a strategy

If you hear “artificial intelligence” and picture a second company that runs itself, you have been sold a slide. What you can actually buy today is a clerk: something that drafts, tags, extracts, and summarises. It is useful. It is not a plan.

Small operators do not lose because they lack a model. They lose because the same eight questions still live in Instagram, the same rows still live in a sheet, and the site still cannot take a booking. A subscription does not fix that. Naming the job does.

The cheap part is the model. The expensive part is being wrong.

Tools got cheap. That part of the old article is true. You do not need a lab or a data-science hire to draft an email. You do need a person who will read the draft before a customer sees it, and a place the truth already lives — FAQ, policy, catalogue, hours — so the model is not inventing the business.

Start with one job. Not a stack. Not “AI transformation.” If you cannot say what got faster after two weeks, you bought a demo.

• Worth trying: first drafts, labels on mail, notes from a call you still attend, pulling a row from a PDF you keep.
• Not a toy: refunds, diagnosis, legal advice, unsupervised send, anything you cannot undo by Friday.
• Illegal-feeling: pasting customer lists, cards, or health records into a consumer chatbot. If it sees personal data, it belongs in software you control, with the processor named in the privacy policy.

It is not only for large companies. That was never the interesting sentence.

Ten years ago you had to build this. Now you can rent a box that writes. That did not make you a larger company. It made it easier to ship confident nonsense at the speed of a monthly invoice.

Larger teams already have someone who owns the mess when the bot is wrong. You probably are that someone. So the rule is stricter, not looser: draft, then a human sends. If there is no human, there is no launch.

What actually gets better

Efficiency is real when the work was already repetitive and textual. Inbox labels. A first pass on a product description from a spec sheet you wrote. A meeting transcript you search later. That is hours back. It is not “scale.”

Decisions get better only if you already have numbers. Most small businesses do not have “vast data.” They have Stripe, a sheet, and last year’s busy weeks. Export that. Look at it. A model can summarise a CSV you trust. It cannot forecast a quarter you have never measured.

Customers get a better experience when the site already answers hours, price range, and how to book. A chatbot on top of a dead FAQ is a slower phone. Personalisation that is just “people who bought X also bought Y” is a merchandising rule, not intelligence.

Do the site first. Then, maybe, an assistant.

Publish the questions you already type every night. Put booking or checkout on a page that works on a phone. Match the name, address and phone on the site to the listing people actually search. That is the work. Models are optional.

If volume is still real after that, a site assistant may answer only from your published FAQ and services. It refuses money, medicine, and legal questions. It never books a slot the calendar does not have. “24/7 support” that invents Saturday hours will cost you more than the hours it saved.

• Saved replies for the ten tickets you already know beat a bot on the head of the distribution.
• Route before you generate: billing, how-to, shipping, other. Tagging is the ROI. Chat is later.
• Measure reopen rate, not how many conversations the widget “closed.”

Skip the café story

You do not need a parable about a retailer, a café, a jewellery shop and an agency that all “revolutionised” themselves with the same paragraph. If inventory is a mess, connect the till to a count you believe. If the phone never stops, put hours and booking on the site. If average order value is low, the catalogue and the policy page are the product — a recommendation widget is a garnish.

We will not invent a case study to close a call. If you want this built into a site, a store, a portal or a queue, that is a statement of work with one job, a kill switch, and a price in USD. If the idea is a toy, we will say no.

Tools are not the plan

A writing model is fine for a first draft you will rewrite. A grammar checker is fine. A transcript is fine if you were in the room. None of those is an operating system. Google Docs with a prompt is still a document. The system of record stays yours: the site, the sheet you are trying to escape, the inbox, Stripe.

Do not bolt five subscriptions onto a process you have not written down. Write the process. Then pick the smallest tool that drafts inside it. Then measure. Then, if the hours moved, we can put the job in software you own.

How I judge whether the idea is real

I ask three questions on the first call. What job did a person do last week that you want a machine to draft? Who checks the output before a customer sees it? What happens on Friday if we turn the feature off? If you cannot answer those, we are not ready to write a statement of work. We are ready to write the FAQ or the SOP first — and that is often the actual project.

I would rather decline a fashionable chatbot than ship something that lies about hours, stock, refunds or a diagnosis. The invoice has my company’s name on it. That is why the no’s are part of the work.

What you should have in writing before you pay anyone

The job in one sentence. The data the model may see. The human who confirms. The metric you will look at in 30 days. The kill switch. The legal name of the vendor. USD price and when the card is charged. Who owns the repo. If a vendor cannot put those on one page, you are buying a demo.

• Nothing is billed by us before you approve a written scope.
• Typical project billing is 50% to start and 50% on delivery, through Stripe.
• Code, design and prompts we write for you transfer on final payment.

The steps — do these in order

1. 01

   List last month’s repetitive work

   One page. Every message, ticket, or form you handled more than three times. If it is not on the page, it is not a use case yet.

2. 02

   Circle only what is frequent, textual, and cheap to get wrong

   Drafts, tags, search, extraction. Cross out money, medicine, law, passwords, and anything you cannot undo.

3. 03

   Write the source of truth in your own words

   FAQ, policy, hours, catalogue facts. A model with no source will sound confident and still be wrong.

4. 04

   Put that text on a page people can find

   Your site — not a Google Doc nobody opens. Booking or checkout lives there too.

5. 05

   Draft with a model, send as a human

   Replies, rows, descriptions. A person clicks send or confirm. Measure hours saved for two weeks.

6. 06

   Skip the chatbot until the page is doing the job

   If DMs and calls did not drop, the FAQ is incomplete. Fix the site before you rent a widget.

7. 07

   Name the processor and the data

   Privacy policy, no consumer chatbot for customer files, no cards in prompts.

8. 08

   Scope the build or stop

   One job, USD price, weekly increments, you own the repo. Or you keep the discipline and no new software. Both are legitimate.

Keep reading

• Booking

  A booking website for a local business — not another scheduling app

  Square, Calendly and Setmore rank for “booking website.” Operators need a site that takes the appointment, matches the shop, and still works if you cancel the widget.

• Web app

  When a spreadsheet should become a custom web app

  No-code vendors rank for “turn Excel into an app.” Operators need a scoped tool they own — roles, a confirm step, and a kill switch — not another sheet with a skin.

• Website cost

  How much a small business website costs — and what $2,500 actually buys

  A real budget for a small business website in 2026: template rent, Origin from $2,500, custom builds, and what you own when the last invoice is paid.