Ask ten UK small business owners if they use AI and about five will say yes. The British Chambers of Commerce and Atos found 54% of SMEs actively using it, up from 35% the year before. Ask what they actually do with it, though, and the answer is usually “I paste things into ChatGPT.” Fair enough. Handy, even. But it isn’t an agent.
An agent is different. You hand it a goal, something like “chase every invoice that’s more than 30 days late,” and it works out the steps, uses your tools, and tells you when it’s stuck. That’s the idea, anyway. An AI Development Company can help businesses build that kind of system, but Getting there is where most people trip.
Start With the Boring Job
Everybody wants a clever agent. Start with the boring one.
Chasing invoices is a good example. Most owners hate it. It feels pushy, the customer is usually one you’d like to keep, so the reminder gets put off another week. And meanwhile cash sits in somebody else’s account.
Here’s a thing most small firms don’t know. A 1998 law lets you charge interest on late B2B invoices, 8% on top of the Bank of England base rate, plus a flat fee of £40, £70 or £100 depending on the invoice size. Almost nobody claims it. Too awkward, I suppose. (Your contract can change the terms, so check before you start waving the Act around.)
An agent doesn’t find any of this awkward. It spots the overdue invoices, writes a polite nudge, and hands the tricky ones to you. The same trick works for routine customer questions and new-lead replies, and for juggling bookings. What it shouldn’t touch is an angry customer. Those need a person, every time.
Picture a Tuesday morning at a twelve-person builders’ merchant. Nine invoices are past 30 days. The agent has already sent polite reminders to six, spotted that two of those customers replied “paid last week,” and flagged one big account to the owner because the reply sounded annoyed. The owner spends ten minutes on that one call instead of two hours on all nine. That’s the pitch, and it’s a modest one. Which is exactly why it works as a first project.
What an AI Development Company Does First
Not code. Questions.
How many invoices are late today? How many hours do you lose to chasing every month? Where does the “who owes what” list actually live, the accounts package or a spreadsheet only one person understands? And who on your side is going to look after this thing once it’s running?
A decent AI development company asks all of that before it writes a line. Boring, yes. But it’s the difference between “the agent saved us a day a week” and “we’re not sure it did anything.”
Test Small Before You Commit
A proof of concept is a short test. Does this idea work at all? Two or three weeks, one job, a clear yes or no. A pilot is the real thing, with real customers and real mess, and it comes later.
Mix them up and you pay pilot prices for a science experiment. A demo that handles ten sample emails tells you almost nothing about the next ten thousand.
Don’t Skip the Paperwork
Yes, there’s paperwork. An agent that reads customer emails is handling personal data, so UK GDPR applies, and the ICO says you’ll need a data protection impact assessment in most cases where AI touches personal data. Don’t let that scare you off. It’s a document, not a court case.
Check the ICO’s current pages, though, because things have moved. The Data (Use and Access) Act 2025 rewrote the rules on fully automated decisions and they’re looser now. People still have to be told, and they still get to challenge a decision.
Where First Agents Go Wrong
Mostly it’s not the technology. Somebody picks a tool before picking a problem. Nobody writes down what “done” looks like. The data turns out to live in five places. Then the whole thing lands on someone who already has a full-time job, and it quietly dies. Sensible AI ML development plans for every one of those before anything gets built.
Know When It Hands Over
Every agent needs a way out. Somewhere there’s a rule for when it stops and a human steps in. A complaint, say, or a refund over a certain amount, or anything that smells legal. The human gets the whole thread, so the customer isn’t asked to explain it all again.
Getting that line right takes a few weeks of reading real conversations. It’s not glamorous, and honestly it’s where good AI ML development earns its fee. The model is the easy part. Deciding what it mustn’t do is harder.
Oh, and when the project’s done, make sure you own what got built. The instructions the agent follows, the test emails, the notes on how it plugs into your systems. Plus an off switch. Some firms hang onto all of that. It’s not illegal, just annoying when you want to leave.
Picking a Partner
For a first agent, you don’t need a grand AI ML development roadmap. You need somebody who’ll scope one job with you before quoting. Then measure the result. Then hand the thing over. Firms that skip the scoping quote on guesses, and those quotes have a way of growing.
BiztechCS, which delivers AI/ML Development and GenAI Solutions for mid-market businesses, tends to start from the workflow and the data and leave the tool choice for last. With a small budget, that order matters, because there’s no spare money to rebuild.
So the best first question isn’t “can you build us an agent?” It’s “which one job should it do first?” Pick the boring one.
Feature image by Solen Feyissa on Unsplash
