Framework
What to automate in a small business first (and what to keep human)
You answered the same postcode question four times last week, then retyped the name into a note you'll lose by Friday.
What to automate in a small business, before anything clever, is that repeat: the first reply, the logging, the reminder of who's waiting. Price, complaints and advice stay with you.
The better week is one where the common question answers itself, and your morning is three names rather than a hunt through chats.
Support the work, or do the repetitive bit
The government's AI Knowledge Hub page on using AI at work is a starting point, and it says it is not legal advice. It splits the help in two. Augmentation means the tool helps while you stay in control. Automation means the tool completes a common, repetitive task with clear rules, and you still check the output. Its examples are practical: pulling a fact out of a document, formatting to a pattern, or drafting from a template you already use.
The same page says AI is useful for summarising, finding themes, suggesting edits, restructuring something that exists, and offering initial ideas. It is weak when the task needs high accuracy, your specific context, professional judgement, or current facts you have not given it. It uses "jagged frontier" for a tool that does a hard-looking task and then fails a simple one.
Test it on examples you already know. AI cannot replace your professional judgement. For a small firm, the decision is the price, the date, the complaint and the advice.
The page also says AI can invent facts, statistics and references, and that you check before a customer relies on the output.
The demo that "runs the whole business" still can't tell you've put the van in for a service.
Ten minutes: split one job you did yesterday into steps. Mark each step "rule" or "judgement".
Start with one job you can describe
The Small Business Commissioner's advice on agentic commerce is aimed at how customers and AI tools find a small firm online. One of its steps is still the right pace for automation inside the business: use AI to improve how you work, try it for things like customer enquiries and admin, start small, test what works, and build up over time. The page also says you do not need a large budget or a major technical change to begin, and it tells firms to answer real customer questions in plain content, especially FAQs, and to keep prices and availability accurate.
That last point decides what you are allowed to automate. If the price on the site is out of date, a bot that repeats it will be confidently wrong. Fix the source, then let a tool repeat it.
Automation copies. It does not know that you put the van in for a service and cannot take Friday.
A fair first week looks like this:
- Pick one job. Write it in a sentence. "When someone asks if we cover their postcode, reply with the area we actually cover and ask for the job details."
- Write the correct answer yourself. Keep it short enough that you would send it.
- Run the tool on five old enquiries whose right answer you already know.
- Read all five. If one is wrong, the job is not ready.
- Switch it on for new enquiries, and read every reply for five working days.
- Only then ask what the second job should be.
What a super intelligent CRM does is this shape, not a promise to run the firm. The chatbot replies in your words. You're told who needs a reply.
Saving leads into the CRM comes with the Solo plan. You still decide.
Small business automation goes wrong when it starts at the interesting end and never reaches the dull job that was costing you enquiries.
Ten minutes: write the one sentence the tool is allowed to send, then test it on one old enquiry whose right answer you already know.
What to automate in a small business, on one page
Use the table when someone suggests automating "customer service" as if that were one task. Break it up. A job can be drafted by a tool, sent by a tool, or kept with a person. Those are different.
| The job | A tool can do this | A person keeps this |
|---|---|---|
| Common first question | Send the answer you wrote, at any hour, and ask for the missing detail | Any question that is not on the list you approved |
| Logging the enquiry | Put the name, the ask and the contact into one list as it arrives | The judgement of whether this is a customer you want |
| Reminder of who is waiting | Surface the people with no reply, or a quote that has sat still | The choice of who you ring this morning, and what you say |
| Chasing a quote once | Draft the nudge from a line you approved, on a date you set | The decision to stop, and any second chase that needs a lighter or firmer tone |
| Price | Repeat a fixed price you published, for a job you defined | Anything unusual, any discount, and any price you have not written down |
| Complaint | Acknowledge it, record it, and tell them a person will reply | The apology, the remedy and the decision |
| Advice | Collect the facts and book a time for a person to talk | The advice itself |
| Yes, no, or a date that commits you | Pass the request to you with the details attached | The answer the customer will hold you to |
Read down the middle column and you've a sequence, not a shopping list. The top three are the usual first month. The bottom four are why "automate the business" is a bad brief.
The Monday pipeline check is the human side of the reminder row. A tool can show you the quiet leads. Five minutes on Monday is you deciding what to do with them.
After a month of the top row, you stop retyping postcodes in the van. The odd jobs are still yours, and they're the ones that needed you.
Ten minutes: copy the eight rows and tick the one you'll switch on this month. Leave the rest blank.
Where time goes in a service firm
A message arrives while you are with a customer and sits until evening. Why slow replies lose work is the longer version. The useful automation is an honest opening: the answer to the common bit, the missing detail, and a person who will follow up.
Then the retyping. Automate repetitive admin means the name lands on the list as the chat ends. If you still retype it, you have not automated the job that was hurting you. Then the memory problem: you remember the loud customer and forget the polite one. A reminder that lists who has had nothing from you should be names, not an essay.
The government's page says to break the work into steps and notice which are rules and which are judgement. Filing the enquiry under the right name is a rule. Agreeing what a quote includes is judgement. Chatbot, live chat or AI receptionist will stop you buying a phone product when the leak is the website, or the other way round.
Ten minutes: time yesterday's delays. Write "first reply", "retyping" or "forgot to ring" next to each one.
Keep a human in the loop where the effect is real
Keep a human in the loop is not a slogan against automation. It is the condition that makes the first three rows safe. The loop can be small.
You read replies once a day. You approve the ten answers before they go live. You take every complaint yourself.
You are still in the loop if the bot handled twenty "do you cover this postcode?" chats and you handled the two that were odd.
You step out of the loop, in a way that can hurt people, when a system says yes or no and nobody can review it. The ICO's guidance on automated decision-making and profiling draws that line. For you: if no person sees the outcome, and the outcome matters, don't automate the yes or the no.
Collect the facts and decide yourself. The clause detail is in the FAQ.
A chatbot that collects a postcode and says "we'll ring you" isn't that kind of decision. A system that refuses someone, or sets a price they can't query, might be. Customers forgive a short wait on a quote more readily than a wrong price nobody will revisit.
Complaints belong in the same bucket even when the law isn't the point. An automated apology that invents a remedy is worse than a same-day line that says you'll reply. Acknowledgement can be instant. The outcome can't be guessed.
Ten minutes: read yesterday's automated replies, if you have any. Pull any line that commits you to a price or a date.
What to teach it, and what to leave out
The answers you automate are an asset. Write them as you would say them, then store them where the tool reads. The first chatbot briefing is a concrete list. For this decision, the rule is smaller: if you would be annoyed to see the sentence go out under your name, it is not ready to automate.
Leave out:
- Prices you calculate case by case.
- Dates that depend on the diary in your van or your head.
- Opinions about a customer's situation, especially where you advise them. A financial adviser is one example.
- Anything you are only "pretty sure" about. Pretty sure is how invented details start.
Include:
- The area you cover.
- The services you do and do not offer.
- What you need in order to quote.
- When a person will reply.
- How to reach a person if the bot is stuck.
The Small Business Commissioner page's advice to answer real customer questions directly, and to keep that content up to date, is the maintenance plan. When the price changes, change the sentence the bot is allowed to say. When you stop offering a service, take the sentence away.
A quarterly read of those notes is enough for most firms. A weekly read is wiser in the first month.
Ten minutes: delete one answer you're only "pretty sure" about. Replace it with the handoff line.
Sort one real day before you buy a second tool
Write yesterday's delays down and look for repeats. A question you answer every week, in three sentences or fewer, is a candidate for the approved first reply. A name you retyped is the logging job.
A person you meant to ring is the reminder: the tool keeps the name, and you still make the call. The point of small business automation here is fewer leaks, not a smaller role for you.
Leave the stressful tasks. Pricing a job you have not seen, answering a complaint, saying you are fully booked: a wrong sentence costs the most. Automate the admin around them.
If only one task repeats, automate that one. Unchecked automation is how mistakes go out at scale.
Ten minutes: list yesterday's repeats. Circle one. Ignore the stressful task you were tempted to hand over.
Signs it is ready, and signs it is not
The pilot that "just needs a quick look" has been in the quick-look pile since March.
Ready:
- You can point at the sentence the tool is allowed to send.
- You have tested it on old enquiries and you agreed with every reply.
- A person can see the conversation afterwards.
- The customer can reach you when the tool stops.
- The source notes have a date, so you know when you last checked them.
Not ready:
- The demo answered a question you have never written down, and you liked the confidence more than you checked the facts.
- The only record is an email notification you already ignore.
- You would be embarrassed to read the reply out loud under your own name.
- Turning it off is unclear. If you cannot pause it this afternoon, you do not control it.
The Knowledge Hub page tells you to test with known examples, because success on one task does not transfer. Use last month's real questions. If a reply fails, change the source or take that question off the list. Don't add a second tool to compensate.
Ten minutes: run five old questions through the tool. If one is wrong, it isn't ready.
A voluntary check, if you want one later
Once a tool is in use, some owners want a structured way to ask "are we being sloppy with this?" GOV.UK publishes guidance for the AI Management Essentials tool, from the Department for Science, Innovation and Technology. AIME is a self-assessment of the way an organisation manages AI. The guidance says it is not there to evaluate the AI product itself.
It is aimed at organisations of different sizes, and it is primarily intended for small and medium-sized firms and start-ups. The tool is not mandatory. It does not provide formal certification.
Working through it is meant to show strengths and weaknesses in your own processes, and to suggest practical actions.
You don't need it to answer a postcode question well. It's optional homework for the month after the first automation has settled, if you want a checklist wider than "did the bot behave this week?" Don't wait for it before you fix a slow reply. The enquiry you missed this afternoon won't be saved by a framework.
Ten minutes: if a tool is already on, write one sentence on what it's allowed to decide. If you can't, pause it.
A month that stays small on purpose
Week one: write the ten answers and the handoff line. Test them on old messages. Do not announce anything.
Week two: let the first reply run on the website, and log every enquiry to one list by hand. Read every automated reply.
Week three: turn on the reminder. Each morning, look at who has had no human follow-up. Make the calls. The follow-up rhythm tells you how to do that without becoming a nuisance.
Week four: change one thing that went wrong, and nothing else. If the month was calm, pick the next repetitive job. If it was not, stay on the first one.
That pace feels slow next to a product tour. It's how a firm of one to twenty people adds automation without waking up to a customer who was promised the wrong thing.
Picture Friday of week four. The common question has a written answer. New names are on one list. You spend the morning on the two conversations that needed a person.
Ten minutes: do the week-one task only. Write ten answers and the handoff line. Don't announce anything.
Common questions
What should a small service business automate first?
The first reply to common questions, the act of writing the enquiry into one list, and a reminder of who is waiting. Those three are repetitive, they follow rules you can write down, and they fail today because you are busy, not because they need judgement.
Should I let a tool send prices to customers?
Only if the price is a line you wrote, for a job you defined, and a person can still change it. A tool that invents a price from old messages will eventually be wrong. Quoting unusual work, discounts and anything you would normally think about stays with you.
What if the automated reply is wrong?
Treat that as a fault in the setup, not as a surprise. Read the replies for the first week. Correct the source note the bot used, or take that question off the automated list. The government's own guidance on using AI at work says you remain responsible for checking outputs.
Do I need a certificate before I use AI in the business?
No. The government's AI Management Essentials tool is a voluntary self-assessment for smaller firms. It does not certify you, and it is not a legal requirement. It is there if you want a structured look at how you manage AI, after you have a real use in place.
When does a human have to be able to step in?
Whenever the outcome matters and the customer can't reach a person. The ICO's guidance on automated decision-making says UK GDPR covers automated decisions and profiling generally. Article 22 adds further rules only when a decision is made solely by automated means, with no human involved, and it has a legal or similarly significant effect. That kind of decision is allowed only where it's necessary for a contract, where the law authorises it, or where the person gave explicit consent. If you're in that territory, tell people, give them a simple way to ask a human to review or challenge it, and check the system is doing what you intended. A chatbot that takes a postcode and says you'll ring isn't that decision. Refusing a service, or setting a price nobody can query, might be. Outside that line, a phone number or a clear handoff still belongs on every automated reply.