Ask where AI agents work best and the market will answer with job titles — sales agents, support agents, research agents — as though capability lived in the role. In our experience it lives somewhere plainer: an agent does its best work wherever the information a job runs on is within reach, and the way of working can be written down. The information might be sitting in your inbox, your CRM and your spreadsheets, or out on the web waiting to be gathered. Either way, if an agent can get at it and a page of instructions can describe the job, the agent can hold it — whatever the department, whatever the industry.
We run agents on sales and marketing work because that's the work our own business generates, and it makes for vivid examples. But nothing about the pattern is commercial. The same structure would run a property firm's compliance desk or a clinic's intake checks without changing anything underneath. So rather than shopping for a job title, the better way in is to look at your own week — at the recurring work that has to get done, gets done the same way every time, and eats hours nobody enjoys losing. That's where agents earn their keep.
The page test: how to tell a job suits an agent
There's a test we run on every job a client brings us: could you describe it to a capable new hire in a single page? Where to look, what to produce, what good looks like. If the page can be written, an agent can hold the job — the page becomes its job description, and the agent follows it with a patience no person brings to repetitive work.
Our CRM agent is a working example. Its page names the sources each record gets checked against, what a complete record contains, and what to do when something doesn't match. None of it calls for judgement without criteria; all of it calls for consistency, every day, without the standards slipping on a Friday afternoon.
The inputs don't have to be sitting in your systems already. Some of the best agent work starts from an empty file: our sales agent builds its prospect research from the open web — company sites, local news, public registers, the same places you'd send a junior researcher — and comes back with the findings organised the way its template demands. What matters is that you can say where to look and what to bring back. Whether the information lives in your CRM or three web searches away is a detail.
Jobs that pass the page test tend to share three properties. The inputs can be named. The output has a shape you can describe — a report, an updated record, a draft in your voice. And the work repeats, which is where agents pull ahead of one-off automation: every correction you make is a permanent edit to the page, so the hundredth run is better than the first. We've written more about this shape of work on our individual agents page.
The same agent pattern fits any industry
An agent, in the model we run, is a written job description, a set of allowed skills, a schedule, a seat in your task tracker and a budget cap. Point the same anatomy at different information and you get a different team.
A property business has compliance certificates with renewal dates in a spreadsheet, contractors in an inbox, and a standing routine for chasing both. A clinic has intake forms to check for completeness against a fixed list. An agency has time entries to reconcile against project budgets every Friday. Each of these is a page of instructions pointed at reachable information — which is the whole job.
This is why, when we talk to a new business, we'd rather hear about their recurring work than show off an agent roster. The roster is simply what the pattern produced when we pointed it at a commercial pipeline. Pointed at yours, it would produce a different set of names.
What to do when the conditions aren't there yet
Where agents fall short, the cause is usually a missing condition rather than a missing capability. Sometimes the data was never recorded — nobody has ever written down why customers leave, so there's nothing to aggregate (though an agent can often start collecting it). Sometimes the tools can't be reached, and the work stalls at the boundary. And often the process exists only in one person's head, which gives the agent nothing to read.
Two of those three gaps close with ordinary effort. Wiring up access and writing the process down is most of what a setup engagement actually is, and it is the main cost of the approach: the running costs are cheap, and the investment is the first few weeks, spent making your business legible enough for an agent to work in. Businesses that do it tend to find the written processes valuable on their own — a process that can brief an agent can also brief the next person you hire.
Why agent work still relies on people
Some jobs can't be captured in a page of instructions, and those stay with people. Judgement calls without written criteria — should we take this client, is this candidate right, does this feel off — have nothing to hand an agent except your instincts. The same goes for delicate relationships, genuinely novel situations, and any decision where the job is to form an opinion the company doesn't yet have. An agent can prepare those decisions — gather the record, lay out the options, draft the recommendation — and the deciding remains yours.
People keep the last word even in jobs agents run well. Anything customer-facing waits behind human approval however routine it is, because the cost of a wrong send lands on your reputation, and quality is a brand signal long before it's an efficiency question.
Where to start scoping out your first agent
The first job to give an agent announces itself with a sigh — the recurring work someone describes as "it just has to get done". Check it against the conditions: the information is reachable, in your systems or out in the world; the tools can be wired up; the process fits on a page. Most businesses have several candidates, and the best first one is internal, so an early mistake costs a correction rather than a customer.
When the first job runs well, the second is easier to choose, because by then you've started seeing your own operation the way this piece describes it: as a set of pages an agent could hold. If you'd like help finding yours, bring the work you'd most like off your desk and book a call — we'll tell you plainly whether an agent can hold it, with a live team on screen.