Where AI actually pays back in a small business
Most AI spend in small firms goes on tools nobody measures. The payback sits in a handful of dull, repeated jobs. Here is how I find them, and where I tell clients not to bother.
By Gary Kellett6 min read
For owners and directors of UK firms with 10 to 250 staff who are being asked 'what are we doing about AI?'
Every director I speak to has been told AI will change their business. Very few have been told where, or by how much. The honest answer is that AI pays back in a small number of places, and those places are rarely the exciting ones. They are the jobs your team does every week, in the same way, that nobody enjoys and nobody has had time to fix.
I spend most of my time working out the right answer for a business before I build anything. This article is the short version of how I decide where AI is worth the money in an SME, and where it is not.
Why so much AI spend shows nothing
The pattern I see most often is a licence bought for everyone, a few enthusiastic users, and no change to how the work actually flows. People write emails a bit faster. Nobody can say what the business got back. That is not a failure of the technology. It is a failure to pick a job, measure it and change it.
95%
of organisations in the study saw no measurable return from generative AI
The firms that do see a return have something in common. They started from a business problem with a number attached, not from a tool looking for a use.
The four places AI earns its keep
In owner-led businesses, the payback almost always comes from one of four kinds of work. If you are looking for a first target, start here.
- Re-keying between systems: orders typed from an email into the ERP, supplier invoices copied into accounts, enquiry details moved from a web form into the CRM.
- Matching and checking: purchase orders against invoices, delivery notes against orders, timesheets against job sheets.
- Chasing: overdue invoices, missing documents, unanswered quotes. Rule-based, repetitive, and easy to forget when the week gets busy.
- Summarising and drafting: turning a long thread, a call transcript or a set of documents into a short brief a person checks and sends.
Notice what these have in common. The work happens often, it follows rules most of the time, and a person can check the output quickly. That combination is where a modest automation gives back hours every week.
A simple test before you spend anything
Before I recommend any AI work, I ask four questions about the job. If the answer to any of them is no, I either fix that first or look elsewhere.
Step 1: Does it happen at least weekly?
Monthly jobs rarely save enough time to justify the build and the running cost.
Step 2: Can you say how long it takes today?
A rough number is fine. 'About three hours every Monday for two people' is a baseline. 'It takes ages' is not.
Step 3: Can the system reach the data?
If the information lives in someone's head or on paper, the first project is getting it into a system, not AI.
Step 4: Is there a named owner?
Someone who does the work today and will spend a couple of hours a week helping shape the new version.
Worked example
A distributor where two people spend three hours each every Monday re-keying web orders into the ERP is giving up six hours a week. Across a working year that is a meaningful slice of two salaries, and it is the kind of job a well-built automation handles with a person checking exceptions. That is a better first target than any company-wide chatbot.
Where I tell clients not to bother
AI is not always the answer, and saying so is part of the job. I steer clients away from AI in a few situations.
- When the process itself is broken. Automating a bad process just makes the mess faster. Fix the steps first.
- When a standard feature already exists. Many ERP, CRM and accounts packages can already do the matching or chasing you want. Configure before you build.
- When the decision carries legal or significant effects on a person, such as hiring, credit or pricing for an individual. Here the governance cost is high and a person should stay in charge.
- In regulated work where validation and audit trails matter more than speed. In pharma, healthcare and financial services, the right answer is often a properly implemented workflow system with AI used in narrow, checked places inside it.
How to start without a big bet
Pick one job. Measure it for two weeks. Build the smallest thing that removes most of the manual effort, with a person checking the output. Measure again. If the number moved, keep it running and pick the next job. If it did not, you have lost very little and learned a lot.
That is exactly how I run my own work. I estimate the effort a conventional team would need, build it with an agentic delivery process, then compare the two. The habit of measuring before and after matters more than any particular tool.
If you want a quick read on where your business stands, the automation and AI readiness scorecard takes about five minutes and gives you your top three places to look first.
Questions directors ask
What is the best first AI project for a small business?
Usually a weekly admin job that moves information between two systems, such as entering web orders into the ERP or matching supplier invoices to purchase orders. It happens often, follows rules, and a person can check the result quickly.
How do I know if AI has paid back?
Measure how long the job takes before you change anything, then measure again after. If you do not have a before-number, you cannot prove a return, however good the tool feels.
Should every member of staff have an AI licence?
Not as a first step. Licences without a changed process tend to show little measurable return. Start with one job and one team, prove the gain, then widen it.