AI for Small Businesses: What Actually Works (and What's Hype)

I build AI features for a living, which is exactly why I talk clients out of about half of the ones they ask for.

The technology is genuinely good now. The problem is that “add AI” has become a goal in itself, disconnected from any particular problem. So let me be specific about where it earns its keep and where it quietly costs you money.

What consistently works

Reading documents so people don’t have to

This is the strongest use case and it isn’t close.

Invoices, contracts, delivery notes, application forms, insurance schedules, CVs — anything where a human currently reads a document to extract a handful of fields. Modern models handle messy layouts, bad scans, and inconsistent formats in a way that older OCR never could.

A client of mine was processing supplier invoices by hand: open the PDF, find the supplier, the date, the line items, the total, type them into the system. About 90 seconds each, several hundred a month. Now they’re extracted automatically and a person reviews a pre-filled form in about ten seconds. Same accuracy, a fraction of the time.

Why it works: the task is well-defined, the output is checkable, and a mistake is caught immediately by the person reviewing it.

First-line customer support on your own content

Not a generic chatbot. A system that answers questions using your documentation, policies, and past support tickets, and that says “I’ll get someone to help with that” when it doesn’t know.

Done properly this handles the 60–70% of questions that are genuinely repetitive — opening hours, order status, how to reset something, what your returns policy says — and hands the rest to a human with the conversation history attached.

Why it works: the questions repeat, the answers exist in writing, and there’s a safe fallback when confidence is low.

Drafting the first version of routine writing

Proposal sections, job descriptions, product copy, follow-up emails, meeting summaries. Not final output — first drafts that a human sharpens.

The value isn’t that AI writes better than your team. It’s that editing something mediocre is dramatically faster than starting at a blank page, and it removes the procrastination tax on tasks nobody wants to begin.

Why it works: a human reviews everything before it leaves the building, so the failure mode is “slightly worse draft,” not “wrong information sent to a customer.”

Classifying and routing the inbox

Sorting incoming messages by intent, urgency, language, or topic. Deciding what’s a sales enquiry, what’s a complaint, what’s a supplier, what’s spam.

Unglamorous, cheap to build, and it removes a genuine daily bottleneck.

What usually disappoints

“An AI assistant that knows everything about our business”

This is the most commonly requested project and the most commonly abandoned one.

The problem isn’t the model. It’s that the knowledge is scattered across four systems, three of which are out of date, and much of what matters is in people’s heads and never written down. An assistant built on inconsistent information gives confident, inconsistent answers — and one confident wrong answer destroys trust in the whole thing.

What to do instead: pick one well-documented domain — your product documentation, your HR policies — and make the assistant excellent at that. Expand only where the underlying information is actually reliable.

AI making decisions with real consequences

Approving refunds, pricing quotes, screening candidates, flagging fraud. These get pitched as automation wins and they’re where things go badly wrong, both practically and legally.

The failure is rarely dramatic. It’s a slow drift — the model is subtly wrong in a consistent direction and nobody notices for months, because there was no human in the loop to notice.

What to do instead: let AI prepare the decision and a person make it. Surface the recommendation, the reasoning, and the evidence. Keep the click human.

Predicting the future from thin data

“Can AI predict which customers will churn?” With five years of clean, labelled data and a real pattern in it, sometimes. With eighteen months of messy CRM records, no — you’ll get an expensive random number generator that reflects your data’s biases back at you with unearned confidence.

Content generated at volume for its own sake

Publishing fifty AI-written articles a month doesn’t build authority; it builds a liability. Search engines have gotten good at spotting it, and more importantly, readers have. Your credibility is the asset. Don’t trade it for volume.

The four questions I ask before building anything AI

Would a knowledgeable human do this task well? If a smart new employee couldn’t do it reliably with the same information, neither will a model. AI doesn’t manufacture missing context.

Can we tell when it’s wrong? If a mistake is invisible or only surfaces months later, you don’t have an automation — you have an unmonitored risk.

Does it fail safely? Good AI features degrade gracefully: they escalate, ask, or defer. Bad ones guess confidently.

Is this the actual bottleneck? Often the real problem is that data is in three places, or nobody owns a decision. AI on top of a broken process just makes the breakage faster.

What I’d do with a modest budget

If you had a small budget and wanted the highest chance of it paying back, here’s the order I’d go in:

1. Automate one document-reading task. Highest certainty of return, easiest to measure, quickest to build.
2. Add a support assistant over your existing documentation. Only if your documentation is actually good — if it isn’t, fixing the docs is the project.
3. Give your team good AI tools and half a day of training on them. Genuinely underrated. Most of the value in the first year comes from people using these tools well, not from custom software.
4. Then consider something custom, informed by what you learned in steps 1–3.

Notice that only one of those is a build project. That’s deliberate.

The honest summary

AI is very good at reading, summarising, classifying, and drafting. It’s unreliable at deciding, remembering, and knowing things you never wrote down.

Build for the first list. Keep humans firmly in charge of the second. That’s most of the skill.

---

Wondering whether an AI project is worth doing at your size? Book a free call. I’ll tell you honestly if the answer is no — it often is, and that’s a useful answer too.