You’ve heard two things about AI, and they can’t both be true.
One says it’ll run your whole business by Christmas. The other says it’s a chatbot that makes things up and the whole thing is a bubble. If you run a company with staff and a P&L, neither is much help. You don’t need a prediction. You need to know what it can do for the work sitting in front of you on a Tuesday, and what it still can’t.
So this is the honest version. What AI genuinely does inside a small or medium business today, where it falls over, and a simple test for telling which side of the line any given task sits on. If you want the full picture first, the complete guide to AI for small business is the place to start, then come back here for the part everyone argues about.
Both camps are selling you something
The hype has a reason to exist. Somebody is raising money, selling a subscription, or building a personal brand on being early. “AI does everything” gets clicks and cheques.
The cynicism has a reason too. Plenty of people bought the hype, wired something up, got nothing, and now feel daft about it. “It’s all nonsense” is a comfortable place to stand once you’ve been burned.
The truth is duller than either. AI is genuinely good at a specific shape of task, useless at another, and the businesses getting real value are the ones who learned to tell them apart. The numbers back this up. The Office for National Statistics found 29% of UK businesses using at least one AI technology by June 2026, rising to 35% among those with 10 or more staff. But the average adopter runs about 1.6 AI technologies, barely up from 1.4 in late 2023. Lots of businesses have switched something on. Very few have it doing real work.
McKinsey’s State of AI survey puts a sharper edge on it: 88% of organisations use AI in at least one function, and roughly 6% get a meaningful bottom-line impact. That gap between using it and gaining from it is the whole subject of this post.
The test that settles it
Forget the use-case lists. There’s one test that tells you whether AI can do a task, and it’s three questions:
- Does it happen often? Once a quarter is a job for a person. Forty times a week is a job worth handing over.
- Does it follow a shape? It doesn’t have to be identical every time. The same rough steps, the same kind of input, the same kind of output is enough.
- Can you check the result? When the AI produces something, can a person tell at a glance whether it’s right, or does spotting a mistake take as long as doing the work?
Three yeses and AI can probably do it, or most of it. A no on any one of them and you’re either better off with a person or better off wiring existing tools together. That’s the filter. Everything below is just working through what passes it and what doesn’t.
What it can genuinely do today
Four things, all of which pass the test, all of which show up in almost every business we look at.
Read the messy stuff and sort it
An enquiry comes in as free-form English. A supplier sends an invoice as a photo. A customer emails a question that could go to three different people. Someone has to read it, work out what it is, and put it where it belongs.
AI does this well now. It reads the unstructured thing, pulls out what matters, cross-references what you already know about that person, and routes it. This is usually the highest-value place to start, because it’s high volume and it’s currently eating your team’s mornings. The person keeps the judgement calls; the machine does the reading and the filing.
Draft the reply that’s 80% the same every time
Quotes, proposals, follow-ups, chase emails, standard responses to standard questions. The bones are the same each time and only the details change. AI writes the draft with the right figures pulled from your systems, and a person reads it and sends it. You keep the sign-off. You lose the typing.
This is what most UK businesses already use AI for, and it’s why adoption looks wide and shallow. Drafting on its own speeds up one person. The value shows up when the drafting is one step inside a process that runs on its own, not a window someone opens by hand.
Pull data out of documents
Invoices, purchase orders, delivery notes, timesheets, contracts, CVs. Anything where a human reads a PDF and types the contents into a system. This is the single most common manual job in British small businesses, and it’s now genuinely solved for the common cases. The awkward ones still route to a person, which is exactly as it should be.
Watch for things nobody has time to watch
A job that’s gone quiet. A customer whose order pattern dropped off. A document that never came back. A margin that slipped on one product line. Your data already knows. Nobody’s looking, because looking is boring and constant and there’s always something more urgent. AI watches all of it and flags the handful that need a human to act.
Notice the pattern. None of these are your product. None are the thing your customers actually pay you for. They’re the tax you pay to run the business, and the Amex SME Barometer put that tax at 11 hours a week on admin and finance for the average UK small business owner. That’s the pile AI is good at.
What it can’t do (and won’t soon)
This half matters more, because getting it wrong is how you end up in the group that tried AI and got burned.
It can’t own the judgement when being wrong is expensive. Final pricing on a big contract. A hiring decision. A credit call. Anything with a legal consequence. AI can draft, summarise, and lay out the options, but a person decides. This isn’t a limitation we expect to disappear. It’s the design. AI produces plausible text, and plausible is not the same as correct, so anywhere a wrong answer costs you real money and is hard to spot, a human stays in the loop by design.
It can’t do your genuinely bespoke work. The thing your customers pay you for because you’re good at it. If your value is that you look at a situation and know what to do, that judgement is both the hardest thing to automate and the last thing you’d want to.
It can’t cope with a process that changes every month. If how you do something is still being argued about, anything built around it gets rebuilt at your expense. Wait until it settles.
It can’t fix a mess you haven’t fixed first. If your customer records live across three spreadsheets, two inboxes and one person’s memory, AI can’t read what isn’t readable. Tidying the data is the real project, and it’s usually cheaper than the AI work that follows it.
It can’t be the last human your customer speaks to. If the only moment a customer hears a real voice is the one thing you automate, you’ve saved a few pounds and lost the relationship. Automate the work around the contact, not the contact itself.
The trap: “can” is not “does”
Here’s the mistake that catches most people, and it’s the reason for that 6% figure.
Almost anything you demo will work. Ask “can AI do this?” in 2026 and the answer is nearly always yes. That’s a useless question. It tells you nothing about whether the thing runs next March, whether anyone trusts its output, or whether it saves a single hour once a person has checked the work.
A capability is not a system. The distance between “it worked in a demo” and “the team relies on it at 8am on a Monday” is error handling, someone who owns it, a place for the awkward cases to go, and a number that proves it saved something. That distance is most of the actual work, and it’s the part that quietly doesn’t get done. We wrote the full anatomy of that failure in why most small business AI pilots never ship, and Gartner has been putting a figure on it for two years, predicting at least 30% of generative AI projects would be abandoned after the proof of concept.
So when someone tells you AI can do a thing, believe them, and then ask the harder question: what has to be true for it to do that thing every week without anyone thinking about it? That’s where the value is, and it’s also where the work is.
How to tell what AI can do for your business this week
You don’t need a strategy. You need a list and the three questions.
Write down the repetitive jobs your team does. The ones people grumble about, the ones that happen daily, the ones that eat a morning. Then run each one through the test: does it happen often, does it follow a shape, can you check the result? The ones that score three yeses are your shortlist. The ones that don’t, leave alone or wire together with the tools you already have.
Then put a number next to each survivor, because a job worth automating has a cost you can measure. How many times a week, how long each time, and what the person doing it costs per hour once you count National Insurance, pension and the hours nobody actually works. Our cost of manual work post walks through that arithmetic properly, and the manual work cost calculator does the sums for you.
That shortlist with a pound figure against each item is worth more than any AI strategy anyone will sell you. To turn it into an order you can defend, which processes to automate first sets out the four questions we use, and the systemisation scorecard ranks your list in about 3 minutes.
What it looks like when the line is drawn right
Founderise took one expensive, repetitive process, the delivery a founder was personally running every step of, and put it through to production properly. 12 hours a week back, a 3.5x increase in margin, 9 weeks from start to a working product. The AI did the reading and the routine steps. The founder kept the judgement. That’s the line in the right place.
MidShift has served over 20,000 professionals through an AI guidance platform, with 92% faster progression through the process it replaced. That volume was never reachable by hand, which is the clearest case of AI removing work rather than speeding it up. Neither started with an AI strategy. Both started with one process that passed the test and a number that made it worth doing.
The short version
AI can read, sort, draft, extract and watch. It’s good at high-volume work that follows a shape and produces something you can check. It can’t own the judgement when being wrong is costly, it can’t do your bespoke work, it can’t run on a mess, and it can’t be the last human your customer speaks to.
And “can” is not “does”. A demo proves a capability. Only a shipped system, owned and measured, proves a return. Get that line right and AI is the cheapest hire you’ll ever make. Get it wrong and you join the majority who tried it and got nothing.
Related reads
- AI for small business: the complete guide
- Why most small business AI pilots never ship
- The cost of manual work: what repetitive admin really costs
- Which processes to automate first (and how to rank them)
- Case study: how Founderise automated delivery
Not sure which of your tasks sit on which side of the line?
That’s exactly what the AI audit settles. Two weeks, fixed price, every repeatable process in your business mapped and costed in hours and pounds, ranked by what to do first, plus one working proof of concept running on your own data rather than a tidy sample. If you go ahead with a build within 90 days, the audit fee comes off it in full. There’s more on how we work on the AI consultancy page.
Book a free call and we’ll tell you honestly what’s worth automating and what to leave well alone.