There’s a lot of noise about AI, and not much of it helps on a Tuesday afternoon. This article cuts through it: what an agent is, the work it’s good at, where it needs you, and what it costs to run.
What an agent is, in plain English
You’ve probably used a chatbot: you type a question and it types an answer. An agent goes a step further. It’s given one job, access to the systems that job needs, and rules about when to stop and ask. Then it gets on with it.
A rota agent, for example, might read a sick message, check who’s free and qualified for today’s jobs, offer them the work and tell you what it did, all before you’re out of bed. It isn’t a general assistant. It does one job, the same way, every time.
What agents are good at
- Reading messy information. Orders sent as photos of handwritten lists, emails, PDFs, voice notes from site.
- Typing it into the right place. Matching what they’ve read to your customers, products and jobs, and entering it.
- Drafting. Quotes, replies, reminders and letters, written in your tone for you to check.
- Chasing. Missing timesheets, overdue invoices, unanswered enquiries, patiently and on time.
- Spotting. Customers ordering less than usual, prices that have fallen below margin, certificates about to expire.
What they can’t do, or shouldn’t
- Know what isn’t written down. If the reason you never deliver to a customer on Fridays lives only in your head, the agent won’t know it until someone tells it.
- Make judgement calls. Whether to keep supplying a good customer who’s paying late is a business decision. An agent can bring it to you in time; it shouldn’t make it.
- Be right every time. AI models make mistakes, so everything we build expects them. Agents check their work against your records, and anything they aren’t sure about comes to a person.
- Replace a relationship. Customers notice when the human touch goes missing from the conversations that matter.
How it looks when it asks you
Good agents are clear about what they did and what they need. Here’s the kind of message the order agent at our example wholesaler sends just before cut-off.
- The Corner Café · 14 linesEntered
- Harbour Kitchen · 22 linesEntered
- “The usual oil” · two possible matchesNeeds you
Two orders are done. One needs a person, because “the usual oil” could mean two products. The agent doesn’t guess. It asks.
How we keep them on a short lead
- Each agent does one job, with access only to the systems that job needs.
- Anything that matters, like sending a quote, releasing an order or moving money, waits for your approval.
- Every agent keeps a record of what it did and why, so you can check its work.
- When something doesn’t match, it stops and asks rather than carrying on.
What they cost to run
Agents rely on AI models that are charged by how much they’re used, so running costs depend on how much reading and writing they do. With fifth day, that’s built into your monthly price, either as a generous allowance under a fair use policy or as a small usage-based part on top. Either way, it’s in writing before you commit.