How to find the work in your business that AI should actually do

Most owner-managers I speak to don't have an AI problem. They have a "where on earth do I start" problem. The tools are cheap, the demos look impressive, and everyone from your accountant to your teenager has an opinion. But when you're running a business turning over somewhere between £1m and £10m, you can't afford to spend a fortnight automating something that saves you twenty minutes a week while the genuinely painful job carries on eating your Tuesdays.

The stakes here are higher than they look. A recent survey of small firms using AI found savings ranging from 11 to 40 hours a month, and the difference between the low end and the high end is almost never the cleverness of the tool. It's whether the tasks they pointed it at were the right ones. Aim it at the wrong jobs and you'll land near the bottom, or below. Aim it well and you'll feel the difference by the second week. So before you buy anything or brief anyone, it's worth an hour with a notebook and an honest look at where your week actually goes.

Think in tasks, not jobs

The first mistake is thinking in whole functions, "we should use AI in marketing" or "in customer service." That's too coarse to be useful. A job like running your accounts or handling a client is a tangle of dozens of little jobs-within-the-job: chasing, drafting, checking, deciding, reassuring, remembering. AI can do a surprising number of the individual threads well. It can't hold the whole knot. So the question worth asking isn't "what job could I automate away," it's "which specific tasks are eating my week and don't really need me."

The three-box test

A task is a good candidate when it ticks three boxes. It's repetitive, so small savings compound. It's rules-based, meaning you could explain it to a sharp new starter in a morning. And it's checkable, so you can eyeball the output in seconds and know it's fine.

That last box trips people up. Plenty of tasks are repetitive and rules-based but genuinely hard to check, anything where "good" is a matter of taste or where a subtle error costs you a client. Those aren't off-limits forever, but they're not where you start.

Chasing an unpaid invoice is a lovely example. It happens constantly, the rules are simple, and you can check it instantly: did the right customer get the right message about the right invoice? Compare that to writing a proposal for a major new client, which is repetitive-ish but where the stakes and the judgement make it a poor first pick.

Do a boring audit of one ordinary week

Keep a rough log for one normal week, not the big projects, the small recurring stuff. Sorting enquiries. Copying figures from one system into another. Writing the same sort of email for the fortieth time. Pulling together a weekly report nobody reads closely. Booking, confirming, reminding, chasing. Most owners are surprised how much of the week is this connective tissue. It rarely shows up in anyone's job description, but it's where the hours go.

Where the wider market has landed tells you what tends to work. In the July 2026 industry figures, the jobs with real traction were customer-service handling (around 64% adoption), coordinating supply and scheduling (58%), and keeping an eye on systems for things going wrong (53%). The exotic use cases, legal review and the like, sit right at the bottom at roughly 18%, precisely because the stakes are high and the judgement is hard. The pattern is always the same: high volume, clear-ish rules, a human watching the edge cases.

Break big jobs into slices

A lot of jobs look un-automatable as a whole but come apart nicely into steps. Onboarding a new customer needs a human, because it's relationship work. But inside it sit half a dozen sub-tasks that don't: setting up the account record, sending the welcome pack, scheduling the kick-off call, chasing the signed paperwork. An agent can handle the plumbing while your person does the part that actually needs a person. Don't ask "can AI do my job?" Ask "which slices of this job are dull, rules-based and checkable?" Then hand those over and keep the judgement for yourself.

Rank by frequency times tedium, then pick one

Once you've got ten or fifteen candidates, the rule for choosing is unglamorous: multiply how tedious a task is by how often it happens, and start at the top. A quarterly job that takes a full day is worth less to automate than a ten-minute job you do six times a day, even though the quarterly one feels heavier. The daily grind is where the hours hide, and it gives you a clean read on whether this is working, because you'll test it again tomorrow. If you'd rather score your candidates more formally than multiply two rough numbers, the five-factor test for choosing a first AI use case weighs value, data readiness and risk alongside the effort.

Two warnings. Be suspicious of anything you want to hand off purely because you dislike it, because disliking a task isn't the same as it being safe to hand off. And resist the hero jobs, the one enormous process you'd love to make disappear, because that's exactly where projects come unstuck. Gartner reckons more than 40% of agentic AI projects will be scrapped by the end of 2027, and the leading reason, blamed in around 41% of failures, is unclear return. Those failures almost always start with someone picking too big a first target.

There's a data angle to the ranking too. When projects stall, poor data is a usual culprit, cited in roughly 39% of failures. Tasks that sit on tidy, self-contained information, an email thread, a single document, a clean list, sidestep that trap. Tasks that need the agent to reach across five disconnected systems don't. Start with the tidy ones.

Don't automate a broken process

One last warning, because it catches people out. If a task is a mess when a human does it, handing it to an agent won't fix the mess. It'll just make the mess faster. Before you automate an invoice chase, make sure your invoices are going out on time and your records agree with reality. Tidy the process first, then automate the tidy version.

What good looks like

Point an agent at a well-chosen task and you should see three things within a fortnight. The first draft should be good enough that editing it is genuinely faster than starting from scratch. You should stop dreading the job. And you should be able to point at the hours it gave back and spend them on something that actually moves the business. If you're not seeing that, the task was probably a poor fit or scoped too broadly. Shrink it, sharpen the brief, and try again. The owners banking 40 hours a month didn't find a magic tool. They just got good at choosing what to hand over. Getting a straight answer on whether it worked means writing the "before" down first, which is where measuring whether AI is actually paying off picks up.

That choosing is the part worth getting right, and it's genuinely hard to do from the inside, because it's tough to see your own week clearly when you're living it. If you'd find it useful to have someone sit down with your task list and help you sort the keep from the hand-over, that's a good deal of what we do at Vision Labs AI. Even an hour of honest sorting on your own will put you ahead of most.