The two-person AI team
A common objection from smaller businesses is that agents are for companies with engineering teams. The reasoning goes: agents need building, monitoring, and fixing, and we do not have people for that. So we will wait.
We think that gets it backwards. Small businesses are often the best fit for agents, because their workflows are simpler, their tools are fewer, and the person who owns the process is usually two desks away. What they need is not an AI team. It is two roles, and a plan for the parts they should not do themselves.
Role one: the process owner
The first role is the person who owns the workflow the agent runs. Not a technologist. The office manager who chases invoices, the coordinator who books appointments, the person who answers the same twelve questions by email every day.
This person decides what the agent does and whether it did it well. They write, in plain language, what a good outcome looks like. They review the outputs in the first weeks. They are the human at the human gate. And they are the one who notices, before any dashboard, that the outputs have gotten worse.
Nobody knows the workflow better. Trying to replace this judgment with a technical role is how agents end up doing the wrong thing very efficiently.
Role two: the operator
The second role is the person who keeps the agent running. They do not need to be an engineer, but they need to be comfortable with the tools the business already uses and willing to read a dashboard.
The operator checks the four numbers every week: how often the agent succeeds, how often a person overrides it, what it costs, and whether anything looks unusual. They handle the small things: reconnecting a tool after a password change, adding a new customer to a list, answering the agent's questions when it is unsure. And they know when a problem is beyond them and needs the care plan.
In many small businesses the process owner and the operator are the same person. That works, as long as the operator part gets a fixed slot in the week rather than being done when there is time.
What the two roles should not do
Building the agents in the first place. Choosing the workflows where agents pay, wiring them to the real tools, writing the guardrails and gates as code, and setting up the audit trail and dashboard is skilled work, and it is a few weeks of it, not an ongoing job. Have it done properly once.
Keeping up with model and tool changes. When the model vendor ships an update or a connected tool changes its interface, the agent's behaviour can shift. Evaluating that and adjusting is exactly the kind of work a small business should not be spending its attention on.
Fixing the agent when it drifts. The operator notices. Someone else should diagnose and repair.
This is what an agent care plan is: the standing arrangement where those three things are someone else's job, and the two internal roles are left with the parts only they can do.
What it looks like in practice
Month one: the agents go live, the process owner reviews nearly everything, the operator learns the dashboard. Overrides are high and that is expected. The care plan tunes the agents weekly against what the owner is correcting.
Month three: the owner reviews a sample rather than everything. The gates on the low-consequence actions have been loosened because the override data justified it. The operator's weekly check takes twenty minutes.
Month six: the agents are part of how the business runs. The owner has proposed two more workflows because they can see what the first ones saved. The care plan has handled one model update and one tool change that the business never noticed.
That is the whole team: two people, a fixed slot in the week, and a plan behind them. It is not an AI department, and it does not need to be.