AI automation for small business
By Kole OgundipeWhat AI automation does in a small business
The same thing it means anywhere: software doing a job that used to need somebody's attention. What changes at this size is which jobs are worth it and who pays for getting it wrong. A large company automates a process that runs ten thousand times a month and can afford a year of discovery to do it. A small business has one person doing six jobs, and the automation that helps is the one that takes a whole job off them, properly, this quarter.
That is why the advice written for enterprises reads oddly to anybody running a firm of eleven. It starts with strategy, data readiness and a pilot programme. The honest starting point here is one question: which job eats somebody's week and could be described to a stranger in ten minutes?
The UK businesses we build agents for, and the ones we do not
Small UK businesses: trades, practices, agencies, recruiters, clinics, small manufacturers. The person who would live with the agent is usually the person reading this.
It is not written for someone looking for a tool to try at the weekend. If what you want is to build something yourself, the tool market is large and you do not need an agency; this page is about handing a job over and having somebody own it.
How an AI agent differs from the automation already in your software
Your software almost certainly automates something already: an invoice reminder that goes out on day thirty, a form that creates a record, a text that confirms an appointment. That is rule-based automation, and it is excellent. It does exactly what it was told, in the same order, forever, and it stops dead the moment reality arrives in a shape nobody anticipated.
An agent is the part that can read the shape. It takes the message, the call or the document as it actually arrives, works out what is being asked, and decides which of the things it is allowed to do applies. That is why it can hold a whole job rather than a step: the caller who rings about two different things, the invoice laid out differently by every supplier, the email that is really a complaint wearing a question.
The practical difference for you is what happens at the edge. A rule breaks loudly and obviously. An agent handles the edge case badly but plausibly, which is why every agent worth running has a line it does not cross on its own and a person on the other side of that line.
How to choose which job to automate first
Four questions, in this order. A job that answers yes to all four is the one to start with, and the order matters: the first two decide whether an agent can do it at all, and the last two decide whether you will still be glad in six months. The last column is the useful one, because it names the kind of job that fails each question.
| Question | Why it decides | A job that fails it |
|---|---|---|
| Does it repeat in roughly the same shape? | An agent learns a job from its pattern. A job that is different every time has no pattern to hand over, only judgement. | Writing the quote for a one-off bespoke project |
| Can you write the rules down in an afternoon? | If you can tell a new starter how to do it by Friday, an agent can hold it. If the rules live only in one person's head, the first job is getting them out, and that is a conversation rather than a build. | Deciding which client to put first when both are angry |
| Is a mistake recoverable? | Everything new gets some things wrong early. The jobs to start with are the ones where a wrong answer is caught and corrected, not the ones where it is sent to your biggest customer. | Paying a supplier invoice without anyone seeing it |
| Would anyone notice the morning it stopped? | A job nobody misses was not worth the handover. It also means nobody will spot the day the agent quietly breaks, which is the failure that costs you. | A weekly summary report nobody opens |
Five jobs in a small business worth giving to an AI agent first
Five jobs, described by the job rather than the technology, with the split between agent and person drawn where we would draw it. Nothing here needs you to change the software you already run.
| Job | What the agent does on its own | What still reaches a person |
|---|---|---|
| Answering the phone | Takes the call, answers what it knows, books into the diary, writes down who rang and why | A complaint, a price negotiation, anybody who asks for a human |
| The shared inbox | Reads every email, replies to the ones it has answers for, files and labels the rest | Anything it has not seen before, and anything that commits you to something |
| Invoices and bookkeeping | Pulls the figures off the document, checks them against the order, puts them where your accounting software wants them | A mismatch, a duplicate, and the payment itself |
| Qualifying enquiries | Asks the questions you would ask, scores the enquiry, books the good ones straight into a diary | The sales conversation, always |
| Chasing and following up | Sends the second and third message nobody ever gets round to, at the interval you set | The reply, the moment it becomes a conversation |
The pattern across all five: the agent takes the volume and the person keeps the judgement. We build whichever of these fits the work rather than selling one of them — the full list of what we build is on the agents we build for businesses page.
Which agent to build first, by the kind of business you run
A rough map, from the shape of the work rather than the sector. Find the line that describes where your time goes.
| If your business | The hours are usually in | Start with |
|---|---|---|
| Takes bookings over the phone | Calls missed while everyone is with a customer, and the ones that come after five | The phone |
| Runs on a shared inbox | Reading and sorting before anybody can answer anything | The inbox |
| Handles a lot of paper in | Typing numbers from documents into software that already has a field for them | Invoice and document entry |
| Buys enquiries, or gets a lot of them | Working out which enquiries are worth a call, usually too late | Qualifying and booking |
| Sells over several conversations | The follow-ups nobody gets round to | Chasing |
Where two lines describe you, start with the one that is already costing you work you would have won. If that is the phone or the enquiry box, start there: a missed call and an unanswered enquiry both go to whoever answered first.
What you need in place before your first AI agent can start
Three things, and none of them is technical.
- One job, named. Not "use AI in the business". One job, with a person who currently does it and can say what good looks like.
- Logins, at the narrowest permission the job needs. This is the part that actually delays projects at this size, because it needs whoever administers your systems to sit down for an hour.
- Somewhere for the awkward cases to go. One named person who gets what the agent cannot settle. An agent with no escalation route is how a small problem becomes a lost customer.
What you do not need: clean data, a strategy document, or anybody technical on your staff.
What running an AI agent asks of you day to day
Quieter than people expect, and the daily commitment is one thing: somebody reads a plain list of what the agent handled and what it passed on. That is how you widen what it does alone, and it is also how you find out on the right morning that something has changed.
Because something will. Your software gets updated, an invoice layout changes, a field gets renamed, and an agent can stop working while everything still looks healthy. Who repairs that, and whether it costs you again, is the question worth asking anybody quoting you for automation. For us it is the service rather than an extra: the running and the repairs are what the monthly fee is for.
Four ways a first AI automation goes wrong
None of these is a technical failure, which is why a technical answer does not prevent them.
- Automating everything at once. Three agents arriving together means nobody learns to trust any of them, and the first bad case lands on all three.
- Nowhere for the hard cases to go. The agent handles the ordinary cases and the awkward ones sit in a folder nobody owns. The awkward ones are the customers who were about to leave.
- Running the old way alongside it. Keeping the manual version going as insurance doubles the work and guarantees the agent is judged on being an extra step.
- Nobody reading the list. It works, so the daily glance stops, and the morning it breaks is found out by a customer instead.
How to tell an AI agent is failing before a customer does
Agents rarely fall over. They degrade, and the three tells are all visible in that daily list without anybody technical reading it.
- Escalations climbing. The share it hands to a person was settling at some level and is now rising. Something upstream changed: a new product, a new question, a supplier's new layout.
- Escalations falling to nothing. Worse, not better. Either it has stopped receiving work at all, or it has started answering things it should be passing on.
- The same case twice. The same question or the same document arriving again a day later means something it did was not saved where it should have been.
The fourth tell is not in the list and is the one that catches people: silence. No calls handled, no emails answered, no errors. An agent connected to a system that quietly revoked its access looks identical to a quiet week, which is why somebody glances at the list daily rather than monthly.
How to measure the work an AI agent gave back
Decide the measure before it goes live, because afterwards everybody argues about the baseline. Two honest ones, and they take minutes.
- Count the work, not the hours. For two weeks before, count the thing itself: calls taken, emails answered, invoices entered, enquiries followed up. Count the same thing after. The difference is real and nobody has to estimate anything.
- Count what used to be lost. Missed calls, enquiries never followed up, invoices entered late. This is usually the bigger number in a small business, and it is the one an hours-saved calculation misses entirely, because the work was not being done at all.
What we would not do is multiply a guess at minutes saved by a guess at an hourly rate. That number can be made to say anything, and you will not believe it in six months.
What UK data protection asks of you when an agent touches customer records
An agent with access to your inbox, your diary, your CRM or your accounting software is handling personal data, so UK GDPR and the Data Protection Act 2018 apply to it exactly as they apply to the systems it reads. The government's data protection guidance sets out what that obliges you to do, and the Information Commissioner's Office is the regulator to read on anything specific.
In practice, at this size, it comes down to three decisions made in the open rather than discovered later: which systems the agent can reach, what it is allowed to do inside them, and what it must hand to a person. They get written down when the job is scoped, and they are worth re-reading once a quarter, because what the agent does tends to widen and the written version tends not to.
What drives the cost of an AI automation build
We publish no figure, because the figure depends on the business rather than the agent, and we would rather give it to you on a call than have you work out whether you are the exception. What moves it is worth knowing before any conversation, with us or anybody:
- How many systems the job touches. One inbox is straightforward. An inbox, a diary and an accounting package that must agree with each other is a different job.
- How much judgement sits in the middle. Work with a written answer is cheaper to hand over than work where the right answer depends on the customer.
- Whether the rules exist yet. If the job lives in one person's head, getting it out is part of the work.
- What has to stay in front of a person. Anything that moves money or goes to a customer unseen changes the design.
What an AI agent costs sets out what UK agencies charge to build an agent and what the off-the-shelf tools charge monthly, so you have the shape of the market before you ask anybody for a number.
Tools, a freelancer or an agency: how a small business gets this built
All three work at this size, and they fail differently, which is the honest way to choose.
- Doing it yourself with off-the-shelf tools is the cheapest in money and the most expensive in your attention. It suits a business with somebody who enjoys this and has the hours. It struggles when the person who built it leaves or stops having time, because the thing still needs maintaining.
- A freelancer gets you something working quickly. The risk is what happens later, when they have other clients and your supplier has changed their invoice.
- An agency costs more than either and should be judged on one thing: whether the agents are still doing the job in six months, and whether fixing them is included. If a build is priced per project, the second agent is a new negotiation, and most businesses only find the second and third jobs after watching the first one work.
How Agent Place works with a business this size
One job at a time, starting with the one that eats somebody's week, and we build whichever agent fits that job rather than the one we happen to demonstrate. The call is free and specific: we will tell you which part of the work an agent can hold, which part stays with a person, what access it would need from you, and which part of it worries us.
Unlimited agents. One flat monthly fee. Nothing extra per call, minute or message. Stop any month. We give you the price on the call.
Related reading: what an AI agent costs sets out what UK agencies charge and what the tools cost, AI automation services covers how the work is run, and the AI automation agency UK page is the plain description of the firm.
Frequently asked questions
- What is AI automation for a small business?
Giving one repeating job — the phone, the shared inbox, invoice entry, chasing enquiries — to software that does it on its own and hands the rest to a person. At this size it is one job at a time rather than a programme, and the test is whether a whole job comes off somebody's plate, not whether a tool has been bought.
- Which job should we automate first?
The one that repeats in the same shape, whose rules you could write down in an afternoon, where a mistake is recoverable, and which somebody would notice if it stopped. The four questions are set out in the table above, together with the kind of job that fails each one.
- Is our business too small for this?
No, but a business can be too early. If no single job repeats enough to describe to a stranger, there is nothing to hand over yet and anything built will be a toy. Two people with a phone that rings all day is a perfectly good starting point; two people doing something different every day is not.
- Do we need someone technical on staff?
No. You need somebody who can get us logins and somebody who reads the daily list of what the agent did. Neither is a technical job.
- Will it replace our staff?
Not in the businesses we work with, because at this size nobody has spare people. What it replaces is the part of the week that was being done badly because there was no time to do it properly: the second follow-up, the inbox at ten at night, the call that went to voicemail. The judgement stays with your people, who now have the hours for it.
- Do we have to change our software?
No. The agent works inside the systems you already run, and an automation project that opens by suggesting you replace them is solving a different problem. The one real exception is a system with no way in at all, where the honest answer is a different design rather than a workaround that will break.
- What happens when it gets something wrong?
Early on, nothing goes out unseen: the agent does the real job on real cases while a person reads every output, so its mistakes are found before a customer meets one. After it goes live it takes the ordinary cases and escalates the rest, and what it may do alone widens as it earns it rather than on a date.
- What does it cost?
One flat monthly fee for unlimited agents, month to month. There is no charge per call, minute or message, and no usage bill. The figure comes on the call, because it is set per business rather than published. What an AI agent costs sets out what UK agencies charge to build an agent and what the off-the-shelf tools cost per month, so you have something to compare it with.
- How many agents do we get?
As many as the business needs. That matters more at this size than it sounds: the second and third jobs are usually only obvious once the first agent is working, and a fee that covers unlimited agents means finding them is a conversation rather than another quote.
Bring us the job
Tell us the work, or the part of somebody's week you would like back. On the call we will say which of it an agent can hold, which stays with a person, what access it would need from you, which part of it worries us, and the price.
Unlimited agents. One flat monthly fee. Fully functioning and delivering value in 30 days guaranteed, or you pay nothing more until your agents are.
Written by Kole Ogundipe for Agent Place, an AI automation agency and consultancy based in the UK which builds bespoke agents for operations, sales and marketing.