Every software vendor now has an AI feature, and every one of them will tell you that small businesses who fail to adopt it will be left behind. That pressure produces a predictable outcome: owners buy three or four AI tools, use them enthusiastically for a fortnight, and quietly let the subscriptions renew for a year without anyone logging in.
The problem is almost never the technology. It is that the adoption starts from the wrong end — from the tool rather than from the task. This article works in the opposite direction.
Start by auditing your week, not the market
Before looking at a single product page, spend a week tracking where time actually goes. A simple spreadsheet with three columns is enough: what the task was, roughly how long it took, and whether the output was largely the same as last time you did it.
That third column is the important one. AI systems are pattern machines. They perform well on work that is repetitive in structure but variable in detail — the tenth version of a quote letter, the forty-third product description, the weekly summary of the same five data sources. They perform badly on work that is genuinely novel, requires judgement about your specific relationships, or carries consequences if it is subtly wrong.
After a week you will usually find that between 20% and 40% of working hours fall into that repetitive-but-variable bucket. That is your realistic target, and it is a large number. You do not need to automate everything to see a return.
The four categories worth examining first
In small businesses across very different sectors, the same four areas tend to come up:
Drafting first versions of routine written work. Proposals, follow-up emails, job adverts, product copy, standard operating procedures. The AI produces a serviceable draft in thirty seconds; you spend five minutes making it correct and specific rather than forty minutes staring at a blank document.
Summarising and extracting. Pulling the action points out of a long email thread, turning a two-hour recorded meeting into a page of decisions, extracting structured data from a pile of invoices or CVs.
Translating between formats. Turning a bulleted outline into a formatted document, a spreadsheet into a client-facing summary, a set of customer reviews into a themed report.
First-line customer questions. Not complex complaints — the twenty questions that make up most of your inbox. Opening hours, delivery timescales, returns process, whether you service a particular postcode.
Notice what these have in common. In each case a human still reviews the output, the cost of a mistake is low, and the alternative is a task nobody enjoys doing.
What to leave alone
It is equally useful to name the work that should stay human, at least for now.
Anything where being confidently wrong is expensive belongs on this list. Regulated advice, tax and legal specifics, medical or safety information, and precise numerical claims about your own business all fall here. Language models generate plausible text, and plausible is not the same as correct. They will produce a figure that looks exactly like a real figure.
Relationship-critical communication is the second category. The apology to a client whose order you lost, the difficult conversation with a supplier, the message to your team about a redundancy. Recipients can usually tell, and the perceived insult of an automated apology outweighs any time saved.
Third, avoid automating a process you have not yet worked out manually. Automation fixes the cost of a good process and permanently entrenches a bad one. If your quoting process is inconsistent, speeding it up produces inconsistent quotes faster.
Choosing tools: general-purpose first
There are two broad options. A general-purpose assistant — the chat-style tools most people have already tried — or a specialist product that embeds AI into one workflow, such as a bookkeeping tool that categorises transactions or a helpdesk that drafts ticket replies.
For most small businesses the sensible order is to become genuinely fluent with one general-purpose tool before buying any specialist ones. The reasons are practical. A general assistant costs roughly the price of one lunch per user per month, covers a very wide range of the tasks identified above, and requires no integration work. Crucially, six months of using it teaches you where AI actually helps in your business — which means that when you do buy a specialist tool, you are buying it for a problem you have measured rather than a problem the vendor described.
Specialist tools earn their place when a task is both high-volume and tightly integrated with your existing data. If you process four hundred invoices a month, a dedicated tool that sits inside your accounting software will beat copying and pasting into a chat window. Below that volume, the integration overhead rarely pays back.
The skill that matters more than the tool
The gap between people who find AI transformative and people who find it useless is mostly a gap in how they ask.
The common failure is a request that is short, vague, and missing context: “write a marketing email for my business.” The output is generic because the input was generic. A far better request supplies four things — the role, the audience, the constraints, and an example of what good looks like.
In practice that means: “You are writing for a family-run plumbing company in Leeds that serves domestic customers. Write a follow-up email to a customer who had a boiler serviced three months ago, reminding them about our annual service plan. Friendly and direct, no jargon, under 150 words, one clear call to action. Here is an email we sent before that worked well: [paste].”
Three things matter here. Context that the model has no way of knowing, explicit constraints on length and tone, and — most valuable of all — an example of your existing work. Pasting in two or three samples of your own writing produces output that sounds like you rather than like a press release.
The second skill is iteration. The first output is a starting point, not a deliverable. Telling the tool what is wrong with the draft — “too formal, cut the second paragraph, and mention the twelve-month guarantee” — is faster than rewriting it yourself and produces a better result than trying to engineer a perfect first request.
Costs, hidden and obvious
Budget for three things, not one.
The subscription is the visible cost and usually the smallest. Expect somewhere between £15 and £30 per person per month for a capable general-purpose tool at business tier.
Training time is the real expense. Realistically it takes a person four to six hours of deliberate practice before AI saves them more time than it costs. Most abandoned rollouts fail because that investment was never budgeted — the tool was announced in a meeting, nobody was given time to learn it, and everyone reverted to their old process within a fortnight.
The third cost is review. Every AI output needs a human check, and if your process assumes otherwise, the errors will eventually surface in front of a customer. Build the review step into the workflow explicitly, and assign it to someone by name.
Data: the question to ask before you sign up
Small businesses handle customer data, and feeding it into a third-party service has real implications under UK GDPR and equivalent regimes.
Ask three questions of any vendor. Is your input used to train their models, and can you turn that off? Business and enterprise tiers generally exclude training by default, consumer tiers often do not. Where is the data processed and stored geographically? And do they offer a data processing agreement? If a vendor cannot answer these clearly in writing, that is information about the vendor.
A sensible internal rule, written down and shared with staff: no customer names, contact details, financial information, or health information goes into an AI tool unless that specific tool has been approved. Redact and describe instead — “a customer who bought a mid-range boiler last spring” works just as well as a real name for drafting purposes.
A realistic first month
Week one, track your time and identify three candidate tasks. Week two, buy one general-purpose tool and spend two focused hours learning it against a single real task. Week three, run that task through AI for everything that comes in, keeping a note of time taken versus your old baseline. Week four, review honestly and either extend to the second task or drop it.
That last option matters. Some businesses genuinely will not see a return, usually because their work is bespoke enough that every job is different. Finding that out for £25 and a month of attention is a good outcome, not a failure.
The businesses that get real value from AI are rarely the ones with the most sophisticated stack. They are the ones that picked two or three tasks, got properly good at delegating them, and left everything else alone.