An AI agent team is a group of AI workers that runs a defined part of your business under one human's direction. The human sets the strategy and approves anything customer-facing; the agents do the work in between. One common shape — the one we run ourselves — is a commercial team: agents that research leads, keep the CRM accurate, draft content, look after customers and report back. The same model fits whichever part of your business generates the recurring work, and the rest of this post shows how it runs in practice — which is far more boring, and far more reliable, than the hype suggests.
We run this model on our own business, and we run a complete demonstration team (eight agents and a human lead at a fictional food marketplace) that we use with clients. Everything below is how those teams work today, including the parts that went wrong.
What is an AI agent, in this model?
An agent, in this model, is a job description plus a set of permissions — readable files rather than a mysterious black box. Ours are defined in plain files anyone can read:
- A job description — what this agent owns, what it must never do, and what good work looks like. Ours are a page long. When an agent misbehaves, you edit the file — the same way you'd correct a process document. No support ticket, no retraining.
- Allowed skills — the playbooks it may run. A sales agent might hold "find prospects", "draft outreach" and "handle inbound enquiry". Each skill is itself a written procedure with a quality bar.
- A schedule — ours run a morning sweep on weekdays: check the queue, do the most useful unit of work, write up what happened.
- A seat in the task tracker — the agent is assignable and @mentionable in the same tool your team already uses. Tag it on a ticket and it picks the work up, usually within seconds.
- A budget cap — a hard weekly spending stop per agent. Our whole eight-agent demo team runs on less per day than a round of coffees.
One human lead sets the direction
The agents take direction from one human — we call them the lead — and from two standing sources. The first is the company's own strategy, written down: goals, quarterly objectives, and each agent's targets live in files the agents read before every piece of work. When the strategy changes, you change the file, and the whole team's behaviour follows. The second is ad hoc: anyone tags an agent on a ticket the way they'd tag a colleague, and the agent responds with the work.
The lead's real job is judgement rather than supervision. In our teams the lead reads one short digest a day — what the agents did, what they're waiting on, what needs a decision — and spends their attention on approvals and direction.
The approval gate between agents and your customers
Nothing customer-facing leaves the building without human approval — a hard rule the system enforces rather than relying on politeness. Agents in our teams hold no sending credentials at all. When a sales agent finishes an outreach draft, it stages the draft with an approval label and cannot touch it again until a human clears it. The email account, the publish button, the money — all of it stays with people.
There is a second gate: everything an agent produces is written down where you can check it. Every run ends with an audit entry — what triggered it, what the agent did, what it cost, what it's waiting on — committed to a repository your whole team can read. When one of our agents once reported work as newly completed that was actually left over from an earlier run, the audit trail exposed it in about a minute: the record showed no new files, so the claim couldn't survive. We then tightened the reporting rule so it can't recur. You don't have to trust an agent's account of itself; you can check it.
A working day with an agent team, hour by hour
Below is a weekday from the commercial agent team we run. It manages a real sales pipeline, a real CRM and real reporting, so the rhythm below is lived, not imagined.
06:00–07:00 — morning sweeps. Each agent works through its patch before anyone is at a desk. The CRM agent checks every record against its sources and stages corrections. The sales agent works the queue: new leads researched and scored, follow-ups prepared, outreach drafted and parked behind the approval gate. The market agent sweeps competitor news and pricing pages.
07:10 — market intelligence lands. The market agent posts what it found to a dedicated channel — say, a competitor raising its prices — and hands the signal to the sales agent, which starts drafting pitches to the businesses that news affects. Nobody asked it to; the handoff is part of both job descriptions.
09:00 — the human joins in. You tag an agent on a ticket the way you'd tag a colleague — "can you pull together a shortlist of prospects in this area?" — and it acknowledges within seconds, then comes back minutes later with the work attached. Colleagues do the same all day; the agents sit in the same tools everyone else uses.
15:45 — afternoon tidy. A shorter sweep: close the loops the morning opened, finish queued follow-ups, no new heavy work.
17:00 — the digest. The report agent posts a short summary to the team channel: what every agent did, what it cost, and — most usefully — a list of items waiting on you, each linked to the ticket where one click approves or declines it. That list is the team's entire ask of your attention.
Friday and Monday. On Fridays the report agent writes the weekly report — activity against the company's goals. On Mondays it reviews spend against each agent's budget cap and proposes reallocations, which you approve like anything else.
The human time in all of that is roughly ten minutes a day, spent on decisions rather than chasing.
What goes wrong, and how it's caught
Like looking after a real team, when your agent team is set up and managed poorly there's potential for things to go wrong. That's why the guardrails exist. Agents drift off their remit if the job description is loose — early on, one of ours answered a task using another company's material that happened to be within reach. We fixed it by rewriting its job description to name exactly what it may read, and by changing the setup so that material never reaches it at all. Agents occasionally over-claim their own work, which the audit trail catches. And an agent given a task too large for its allotted thinking budget will fail mid-job — so tasks need sizing, and failures need to be loud (ours now report failure on the ticket rather than going quiet).
These are the kinds of mistakes you'd expect from a keen junior hire, and they respond to the same management: clear responsibilities, reviewed output, and a paper trail. Agents make that management easier in one respect — every piece of work leaves a written record, and a fix is usually an edit to a file rather than another difficult conversation.
If you're wondering what a team like this would look like in your business — which jobs it would take on first, what it would plug into — that's exactly the conversation we like having. Book a call and we'll show you a working team live, or start with what a single agent can do.