Where AI automation actually pays off for a small business.

Start from the work, not the technology

The best automation projects we see begin with a sentence like "two people spend every morning sorting the inbox" or "every quote takes forty minutes to write and they all look the same". The worst begin with "we should be doing something with AI". The first kind has a clear task, a clear owner and a clear way to measure improvement. The second usually ends as a demo.

So before any tools, write down the three most repetitive tasks in your business, roughly how long each takes per week, and what goes wrong when a person does them in a hurry. That list is your shortlist.

What today's AI models are good at

Language models are strong at work that involves reading and writing text, especially when the text is messy:

• Classifying and routing. Reading an incoming email or message and deciding whether it is a sales enquiry, a complaint, a booking change or spam, then sending it to the right person.
• Extracting structure. Pulling names, dates, amounts and items out of emails, PDFs and forms into fields in a system.
• Drafting. First drafts of replies, quotes, summaries and follow-ups in your house style, for a person to check and send.
• Summarising. Turning a long thread, a call transcript or a week of tickets into a few lines someone can act on.
• Answering from your own material. Responding to repeat questions using your approved policies, price lists and FAQs, rather than the model's general knowledge.

Where it needs a human in the loop

Models can be confidently wrong. That is manageable when a mistake is cheap and visible, and dangerous when it is expensive or hidden. A simple rule works well: the more a mistake would cost, the closer a person should sit to the decision.

• Low risk, fully automatic: tagging and routing messages, filling internal fields, producing internal summaries. If it is wrong, someone notices and fixes it in seconds.
• Medium risk, human approves: customer-facing replies, quotes and anything that commits the business to a price or a date. AI drafts, a person clicks send.
• High risk, human decides: refunds, legal or medical statements, anything involving money moving. AI can gather the information, but a person makes the call.

Designing these checkpoints in from the start is what separates a dependable automation from one the team stops trusting after the first embarrassing reply.

Five good first projects

1. Inbox triage. Every inbound message is labelled, prioritised and routed, with a suggested reply waiting for the person who opens it.
2. Document to record. Supplier invoices, application forms or booking requests become structured records in your system, with anything uncertain flagged for review.
3. Quote and proposal drafts. The model fills a template from the client's request and your price list, and a person adjusts and sends it.
4. Customer questions on messaging channels. Common questions on WhatsApp or the website are answered from your own content, with a smooth hand-over to a person for anything else.
5. Daily or weekly briefings. A short summary of what came in, what is overdue and what needs a decision, delivered where the team already reads.

How to measure whether it works

Pick the measure before building, and keep it boring:

• Time: minutes per task before and after, measured on real work for a couple of weeks.
• Accuracy: the share of AI outputs a person accepted without changes, and the share they had to fix.
• Speed to customer: how quickly an enquiry gets a first reply.
• Escalations: how often the automation handed over to a person, and whether it handed over the right cases.

If the acceptance rate is low, the fix is usually better instructions, better examples or a narrower task, not a bigger model.

Data, privacy and cost

Three practical points that are easy to skip:

• Know where your data goes. Use providers and settings that do not train on your business data, keep personal data to the minimum the task needs, and make sure your privacy policy says what you do.
• Keep the source of truth in your own system. The AI reads from and writes to your records; it should not become the only place information lives.
• Watch the running cost. Usage-based pricing is cheap for most small-business volumes, but it should be measured like any other cost, per task, not guessed.

The honest summary

For most small businesses, the first useful AI automation is not a chatbot on the homepage. It is a quiet helper that reads, sorts and drafts so people spend their time on the decisions and conversations that need them. Start narrow, keep a person where mistakes are costly, measure honestly, and expand only what works.

This is the work we describe under AI. If you have a repetitive task in mind, tell us about it.