A chatbot answers questions. An agent does jobs. Give an agent an outcome — chase the unpaid invoices, summarise this inbox, keep our price list current — and it works through the steps itself: reading documents, using your other software, checking its results, and coming back with the work done rather than a paragraph about how you might do it.
That one difference — answers versus outcomes — is why we think agents will quietly become part of how small firms operate. Most of the repetitive work in a small business isn’t a question. It’s a procedure. And procedures are exactly what agents are built to run.
The difference in practice
Ask a chatbot “which of these invoices don’t match their purchase orders?” and you’ll get advice on how to check. Give an agent the same job and it opens the folder, reads each invoice, compares the figures, and hands you a list of the three that don’t match — with the discrepancies highlighted.
The work agents already do well shares a shape:
- Repeatable — the job runs the same way every week: invoice triage, inbox summaries, report assembly.
- Rule-shaped — you could brief it to a capable temp in a page: what to look at, what good looks like, what to flag.
- Checkable — a person can look at the output and quickly tell whether it’s right.
If a job needs judgement your firm would argue about — pricing a deal, handling a complaint, anything with a relationship attached — it isn’t agent work. That line will move over time, but it’s the honest line today.
Why we think this matters for business operations
Three reasons, none of them hype. First, the cost shape changed: capable agents now run on the same £15–20-a-month plans teams already buy, not on enterprise contracts. Second, the plumbing arrived: assistants can now connect to the tools a firm already uses — inboxes, drives, calendars, spreadsheets — which is the difference between an AI that talks about your work and one that touches it. Third, the work is there: every small firm carries hours a week of procedure-shaped admin that nobody was ever going to hire for.
Agents don’t replace the decision. They replace the forty browser tabs between decisions.
The firms that benefit first won’t be the ones with the biggest AI budgets. They’ll be the ones that picked one boring workflow, ran an agent on it supervised for a week, measured the time saved, and then did that again.
What every agent needs — and how to start
An agent is three things wired together: a model (the intelligence — usually a monthly plan), connections (the hands — what it’s allowed to reach), and a human gate (the judgement — a person approves the output before it counts). Nearly every AI horror story in business is a missing approval step; keep the final button human until an agent has earned otherwise, one task at a time.
Starting is smaller than it sounds. Pick the most repetitive, least contentious job in the building. Choose a plan — our five-person team comparison is that decision in one table. Run the agent supervised for a week, checking every output like you would a new starter’s. Then measure: minutes saved times hourly rate, against the subscription. Keep it if it clears the bar; cancel if it doesn’t.
Full disclosure, and the best evidence we can offer: this publication runs on exactly this model. An agent finds the stories, researches them and writes the drafts — and a human approves every article before it goes live. How we built it is the long version, including everything that broke.
Sources & quotes
Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. How we verify →
