AI for Small Business

Start with the arithmetic, not the tool

Most small-business AI advice starts with a list of tools. That is backwards, and it is why so many of these projects quietly die after three weeks. The question that decides whether any of this is worth your time is arithmetic, and you can do it on the back of an envelope before you sign up for anything.

Take one task you do repeatedly. Write down how long it takes, how many times a week you do it, and what an hour of your time is worth — not what you charge a client, what it costs you to not be doing the next thing. Multiply. That is your annual budget for fixing it. A task that takes twenty minutes and happens twice a day is roughly 170 hours a year. A task that takes two hours and happens once a month is 24. The first one is worth automating badly; the second one is probably not worth automating at all.

This matters because the cost of an automation is not the subscription. It is the subscription plus the afternoon you spend building it, plus the hour a month you spend when it breaks, plus the cost of the one time it does something wrong to a customer and you have to apologise. Against 170 hours that is trivially worth it. Against 24 hours it is a hobby.

What AI actually changes for a business this size

A small business doesn't have a corporation's resources, and that is exactly where AI shifts the balance. Not because it is clever, but because it removes the minimum headcount required to do certain things at all. Answering enquiries at 11pm used to require either a night shift or an unhappy customer. Publishing consistently used to require someone whose job was publishing consistently. Those thresholds moved.

What has not moved: anything requiring judgement you would not delegate to a competent new hire on their first week. That is a useful test, and it is more reliable than any feature list. If you would not let a first-week employee send it without checking, do not let a model send it without checking either. New to this? Start with What is AI.

Choosing the first thing to automate

Rank candidates on two axes: how often it happens, and how much it hurts when it is late. High frequency and high pain is where you start. Low frequency and low pain is where enthusiastic people start, because those tasks are easy and the demo looks good, and then nothing changes about the business.

There is a third filter that people skip: is the task well-defined enough to describe in writing? If you cannot write down the rules for what a good output looks like, an AI cannot follow them and you cannot tell whether it did. "Reply to the customer appropriately" is not a specification. "Reply within two hours, confirm the date and the price, never quote a discount, and hand off to me if they mention cancelling" is one.

Customer service: what a bot is genuinely good at

A chatbot on your site or WhatsApp that answers frequently asked questions around the clock (e.g. Tidio) is the most common first project, and it works — for a narrower slice of conversations than the marketing suggests.

What it handles well: opening hours, location and parking, whether you carry a thing, roughly what something costs, where an order is, how to reschedule. These are questions with one correct answer that exists somewhere in writing. What it handles badly: anything where the customer is upset, anything where the answer is "it depends", and anything where being wrong costs you money — quoting a price on a custom job, confirming stock you have not checked, promising a delivery date.

The number that decides whether a bot was worth it is deflection rate: the share of conversations it finished without a human. Nobody will tell you what yours will be, because it depends entirely on how repetitive your enquiries are. A business with ten questions that make up most of the inbox will do well. A business where every enquiry is a bespoke project will not, and should spend the same effort on faster drafting instead.

The escalation rule is the whole design

Decide in advance what sends a conversation to a human, and make the bot do it eagerly rather than reluctantly. Mentions of cancellation, refunds, complaints, legal words, anything about a person's health or money, and any question the bot has already failed once. A bot that hands over quickly and cleanly is experienced as helpful. A bot that loops while the customer types "agent" three times is worse than no bot, because now the customer is angry and they have not been helped.

The inbox: draft, don't send

The highest-value, lowest-risk use of AI in a small business is drafting replies that you then approve. It removes the hardest part — the blank page and the fifteen seconds of context-switching — while leaving the judgement with you.

Set it up so drafts land in your drafts folder, not in the outbox. Generating and sending are two different features and only one of them can embarrass you. Most of the value arrives at the drafting step anyway; the marginal gain from full automation is small and the marginal risk is not. Tools like Claude work well here for drafting, summarising and translating.

Marketing and content, and the sameness problem

Content generation is where AI is most used and most visibly mediocre. The reason is structural: a model trained on the average of the internet produces the average of the internet, and your competitors are running the same prompt. If your posts read like everybody else's posts, the tool has cost you the one thing a small business actually has, which is sounding like a specific person who knows something.

The fix is not a better prompt template. It is feeding the model material that only you have: the actual question a customer asked last week, the job that went wrong and what you learned, the number from your own books. Use the model to structure and tighten that material, not to originate it. See AI for Marketing for the longer version.

• Works: turning one long thing you wrote into five short ones; tightening a draft; generating variations of a line you already like; writing the boring connective tissue.
• Doesn't work: "write me a post about my industry." You will get something true, generic, and invisible.
• Automate the schedule, not the voice. Consistency is the part that machines are good at.

Process automation: the worked example

This is where the hours actually come back. Connect the systems you already use and let the handoffs happen without you: leads, invoices, scheduling and reminders. The two platforms worth learning are n8n and Make — the first is open source and self-hostable, the second is hosted and gentler to start with.

Take the most common one, the lead pipeline. Someone fills in the form on your site. You want: the lead recorded in your CRM, a confirmation to them within a minute, a notification to you, and a follow-up if nobody has replied in three days. Four steps, each trivial, and together they are most of what a junior sales assistant does.

The version that works in a demo

1. Trigger on the form submission.
2. Create the contact in the CRM.
3. Send the confirmation email.
4. Post to your phone or your team chat.

Build that and it will work. It will also break within a month, in three specific ways that nobody warns you about.

The three things that break it

Duplicates. People double-click. Forms get resubmitted. The webhook gets retried by the sending service after a timeout that had nothing to do with you. Now the same person is in your CRM twice and has two confirmation emails. The fix is to make the workflow idempotent: before creating anything, search for the email address and update instead of insert. This is ten minutes of work and it is the difference between a toy and a system.

Silent failure. A step fails — an expired token, a renamed field, an API that is briefly down — and the workflow stops. Nothing happens. No error reaches you, because nobody is watching the platform's execution log, and the leads simply stop arriving. You notice a week later when the month looks quiet. Every workflow you build needs an error path that notifies you somewhere you actually look, and a weekly glance at whether the run count looks like the run count you expect.

The loop. A workflow that writes to a system that triggers the workflow. It fires hundreds of times in a minute, burns your operation quota, and sends a customer forty emails. Guard against it with a condition that checks whether the record was changed by the automation itself, and test new workflows with a filter that limits them to one test record before you let them near real data.

Documents, quotes and meetings

The back-office uses are less glamorous and often pay better than the customer-facing ones, because nobody is watching and a mistake is cheap to catch.

• Summarising documents and contracts — genuinely useful for getting oriented fast. Not a substitute for reading the clause you are about to be bound by, and not a substitute for a lawyer on anything that matters.
• Meeting transcription into tasks — works well, with one caveat: verify names, numbers and dates against the recording. Transcription is confident about exactly the things it gets wrong.
• Drafting quotes from a template — fill the structure automatically, but keep the price a field you type, not a field a model generates.

What this actually costs

You can start close to free: free tiers of the chat tools, and self-hosted n8n on a small server. As things grow, professional tooling for a small business generally lands somewhere in the tens to low hundreds of dollars a month — almost always less than the time it replaces, but only if you count honestly. A full list is in the best AI tools.

What matters more than the headline price is the shape of the pricing, because it decides how your bill behaves when the business grows:

• Per seat — predictable, scales with headcount. Fine for a team of three, punishing if you want to give everyone access.
• Per operation or per run — the automation platforms. Cheap until a workflow loops or you add a step inside a loop that runs for every row. Watch the multiplication.
• Per token — the model APIs. Genuinely cheap per unit and easy to misjudge in aggregate, because long documents and long conversation histories cost more every turn.

Prices in this market move often and in both directions, so check the current figure before you commit rather than trusting any number you read in an article, including this one.

Customer data: the part to get right early

The moment you paste a customer's details into a tool, you have made a decision about their data on their behalf. Some of this is regulated where you operate, and the specifics differ by country — that part is worth a conversation with someone qualified rather than a guess from a guide.

The practical habits, regardless of jurisdiction:

• Don't paste what you don't need to paste. Strip identifiers before summarising when the summary doesn't need them.
• Check whether the tool trains on your input, and whether the business tier turns that off. This is usually a setting or a plan tier, not a negotiation.
• Know where the data physically goes if your customers or your contracts care about that.
• Keep an access list. When someone leaves, their access to the automation platform matters more than their access to email, because a workflow can reach everything at once.

How these systems fail in practice

Two failure modes are worth internalising before you rely on any of this.

The first is confident wrongness. A model does not signal uncertainty the way a person does; the invented policy and the real policy are delivered in the same tone. This is why the escalation rules and the draft-don't-send rule exist. Anywhere a wrong answer has a cost, put a human between the output and the customer.

The second is drift. The workflow you built in March is still running in September, but the form has two new fields, the CRM has a renamed stage, and the price list changed. Nothing errored. It is just quietly doing the wrong thing. Put a recurring note in your calendar — quarterly is enough — to open each automation and ask whether it still describes how the business works.

When not to do this

A page with affiliate links on it owes you this section.

• Your volume is too low. Below a certain number of enquiries a week, answering them yourself is faster than maintaining the thing that answers them.
• The free tier is genuinely enough. For many one-person businesses, a good prompt in a free chat tool plus a calendar reminder replaces most of what a paid stack would do.
• The process is broken, not slow. Automating a bad process gives you a bad process that runs faster and is now harder to change. Fix the process on paper first.
• Nobody owns it. If no one in the business will be responsible for the automations, don't build them. Unowned systems fail silently and take a customer relationship with them.

Measure it, or you will never know

Before you build anything, write down the current number: minutes per enquiry, enquiries answered per day, days from lead to first reply, hours a week on invoicing. Whatever the project is supposed to improve, capture it now. Once the automation exists, nobody can remember what it was like before, and the project becomes impossible to evaluate — which is how businesses end up paying for four subscriptions that nobody can justify and nobody wants to be the one to cancel.

Re-measure after a month. If the number did not move, the honest conclusion is that this was the wrong first project, not that you need a better tool.

A realistic first month

1. Week 1 — measure. List the five repeated tasks with times and frequencies. Pick one. Write down its current number.
2. Week 2 — do it manually, with help. Use a chat tool by hand for that task, every time it comes up. You will learn what a good output looks like, and you will discover the edge cases before you encode them.
3. Week 3 — automate the narrow version. One trigger, one path, one notification. Make it idempotent. Add the error notification. Resist adding a second feature.
4. Week 4 — break it on purpose. Duplicate submissions, revoked credentials, a malformed entry. Fix what you find, then let it run on real traffic while you watch it.
5. Then re-measure, and only then pick the second process.

Build it yourself or pay someone

Building your own is the right call when the process is yours, changes often, and you want to be able to change it at 9pm without booking anyone. The platforms are visual and a determined non-programmer can get a real workflow running in an afternoon. The cost is that you now maintain it.

Paying someone makes sense when the integration is genuinely hard — legacy systems, anything touching invoicing or payments, anything with a compliance requirement — or when your hourly rate makes the arithmetic obvious. If you do hire, insist on two things: that you get access to the account rather than sitting behind an agency, and that they hand over documentation of what each workflow does. Automations you cannot see inside are a liability the day the relationship ends.