What an AI agent could do in your business

Ask most people what AI could do for their business and they describe a chatbot. Something you type a question into, that types an answer back. Useful, but it still needs a person to think of the question, read the reply and do something about it.

The more interesting version does the whole job. It looks at your data, works out what matters, and takes the next step on its own. That’s what people mean by an agent, and it’s harder to picture because there isn’t a familiar consumer version of it yet.

So here are four, drawn from the kinds of jobs most businesses have somewhere. None of these are exotic. They’re the sort of thing today’s tools support once the data underneath is reachable.

Four examples

The Monday morning summary. Somebody spends the start of every week pulling figures out of two or three systems and writing them up for the leadership team. An agent can read the same sources, work out what moved and what’s worth mentioning, write the summary and post it to the right channel before anyone gets in.

The customer who went quiet. Regular customers drift away slowly, and it’s often a quarterly review that surfaces it. An agent can watch ordering patterns, notice when an account’s behaviour changes, and draft a message that refers to what they usually order, ready for the account manager to send.

The order that looks wrong. An unusual quantity, a price that doesn’t match the agreed terms, a delivery address that’s never been used. An agent can check new orders against what’s normal for that customer and product, and raise the odd ones before they ship instead of after.

The month-end chase. Finance loses days each period working out what’s missing and who has it. An agent can find the gaps, work out who owns each one, send the reminders and keep track of what’s come back.

What they have in common

None of these are clever in the AI sense. The model isn’t doing anything remarkable. What makes them work is that the agent can see across more than one system, and knows what normal looks like well enough to spot when something isn’t.

That second part is the bit that gets underestimated. An agent that flags an unusual order needs to know what usual means for that customer, which means history, modelled properly, in a place it can reach. The intelligence is available off the shelf. The context isn’t, and it’s the context that decides whether any of this works.

Which is the useful thing to take away: the work that makes agents possible is the same work that makes your reporting trustworthy. It’s one investment, not two.

How to spot one in your business

A rough test. Look for work that:

• happens on a schedule, or every time something specific occurs
• needs somebody to look at two or three different systems
• ends in a document, a message or a decision
• involves judgement, but mostly the pattern-matching kind

The more of those a task ticks, the better a candidate it is. Most businesses have several once they start looking, and they’re usually jobs nobody enjoys.

Starting carefully, then letting go

The sensible way in is to have the agent do the work and a person approve the output. You see what it produces, you correct it when it’s wrong, and you find out how often that happens.

Then you loosen it. Once an agent has drafted forty weekly summaries and the last thirty went out unchanged, the approval step is costing you more than it’s saving. Take it off. Somewhere else, on the customer-facing message or the payment run, you may decide to keep it forever. It’s a dial, not a switch, and where you set it should come from evidence.

It isn’t finished when it’s live

An agent’s understanding of your business goes stale. You add a product line, or change how a process runs, or start selling into a new region, and its picture of normal stops matching reality.

Somebody needs to keep that current. It isn’t a large job, and it isn’t a technical one most of the time, but it does need an owner. An agent nobody has looked at for a year is one making confident decisions about a business that has moved on.

Where this usually goes

The limiting factor is almost never the AI. It’s whether the data an agent would need is somewhere it can be reached, and whether anyone has written down what the numbers mean.

That’s worth knowing before you scope anything, because it changes the order of the work. If you want to talk through which jobs in your business might suit this, that’s a conversation we’re happy to have.