John O'SullivanCloud & AI AutomationLet's talk

12 October 2026 · 4 min read

Someone in Your Business Is Already Building AI Agents. Do You Know What They're Doing?

A short read for business owners who suspect the "techy one" in the office has been busy.

It usually starts quietly. Your office manager connects ChatGPT to the shared inbox with a free automation tool. Your bookkeeper finds an app that pulls invoices out of emails on its own. Someone sets up a bot that answers website enquiries at the weekend. Nobody asked permission, because nobody thought they needed to. They were just trying to get rid of a boring job.

That's an AI agent: software that carries out a multi-step task by itself, rather than waiting for you to prompt it line by line. And the people building them are increasingly ordinary staff, not developers.

The numbers back this up. In August 2026, Howdy.com surveyed just over 1,000 employees who use AI at work. Two-thirds had gone beyond chatting with a bot and were using agents, 16% had built their own from scratch, and 40% had to train their agent themselves, on top of the day job. That survey was of US workers, but Simply Business found in September that nearly half of UK small businesses now use AI tools, double the year before. If your team uses AI at all, the odds are someone has already gone further than you realise.

This is mostly good news

The instinct to clamp down is understandable, but mostly wrong. The people building these things are the ones who know where the time actually goes. They've spotted a problem that was never on anyone's to-do list and fixed it themselves. In the Howdy survey, 96% of people using agents said they save time, and 84% said the savings outweigh what the agents cost to run.

This isn't new, either. Every business has a spreadsheet that only one person understands, or a macro someone wrote in 2015 that the whole invoicing process secretly depends on. Home-made agents are the latest version of that. The difference is what they can reach.

Where it goes wrong

A quarter of the people in the survey had switched an agent off for good. The reasons were mundane rather than dramatic: 58% because it was inconsistent, 22% because it never worked, and 7% because it "went rogue." In practice, rogue almost always means the agent did exactly what it was told, and what it was told was wrong. That's on the instructions, not the software.

For a small business, the real risks look like this:

  • Customer data wandering off. An agent that reads your inbox or CRM may be sending customer details to a third-party tool nobody has checked. Under UK GDPR, that's your responsibility, whether or not you knew it was happening.
  • Too much access. A tool someone built for their own convenience is now connected to your accounts, your customer records or your website, usually with their personal login.
  • Failing silently. The agent that stopped chasing invoices three weeks ago, and nobody noticed because nobody was watching it. Failing is fine. Failing quietly is not.
  • One person holding the keys. When they leave or go on holiday, the agent lives on in an account nobody else can get into.
  • Surprise bills. 14% of people in the survey had been hit with an AI bill far bigger than expected.
  • Nobody owns it. Only 28% said their workplace was clear about who is responsible for what AI produces. When an agent sends a customer the wrong price, whose problem is that?

These are the same reasons Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027: unclear value and weak controls. Big companies just have bigger versions of the problem.

What to do instead

You don't need a policy document or an IT department. You need to know what's running, draw a few lines, and keep the good stuff.

  1. Ask, don't audit. Get the team together and ask, with genuine curiosity, "what have you automated?" Make it clear nobody is in trouble. If people think they'll be told off, the agents just go further underground.
  2. Write it down. For each one: what it does, what it connects to, whose login it uses, what it can change or send, and what happens if it stops. Half an hour, one page.
  3. Draw the lines. Decide which kinds of data can go into which tools, and which can't. Anything that can email a customer, change a price or move money gets a person's approval before it acts. Tools get set up under business accounts, not personal ones.
  4. Give people a safe place to play. Rather than banning things, pick a couple of approved tools with company accounts, so you can see what's being used and what it costs. Curiosity is an asset. Point it somewhere safe.
  5. Promote the winners. The agents that work deserve to be rebuilt properly: with logs, limits, someone watching, and a named owner. The same rules that catch the ones going wrong are what let the good ones spread to the rest of the team.

That last step is the one most businesses skip. A rule book that only stops bad things is half a rule book. The other half is a path for a good idea to go from one person's clever shortcut to something the whole business can rely on.

Let's talk

If you've read this and thought of a specific person and a specific spreadsheet, you're not alone. We offer a short "agent amnesty" session: a half day finding out what's already running in your business, writing the one-page rules, and working out which home-made agents are worth rebuilding properly. Get in touch below and tell us what you suspect is running.

Sources

← All articles