An AI agent in a business takes one defined task, reads what it needs, uses your tools over several steps and keeps going until the job is done. That might mean drafting a client reply, sorting incoming requests or answering an order question. At Pillet Grenié Bureau, we build and run agents for ourselves and for clients. What we've learned is that an agent is only as useful as the one-page agreement behind it.
What does an AI agent actually do in a business?
A chatbot answers once and stops. An agent works in a loop. Salesforce describes it as reading the context of a request, planning, calling the tools or data it needs, and adjusting until the goal is met. Their everyday examples are high-volume routine work: order status, returns, troubleshooting, onboarding, after-hours support.
That loop is the point, and it's also the risk. Druid AI notes that an agent can write records, trigger workflows, send messages and process transactions. How far it reaches depends on what you connect it to.
Most companies are now trying this. McKinsey's survey, cited by Prefactor, found 62% of organisations at least experimenting with agents, but only 23% scaling them in even one function. And Gartner expects more than 40% of agentic AI projects to be cancelled by 2027, pointing to weak governance. In our experience, weak governance usually means nobody wrote down the five decisions below.
The five things to agree before an AI agent goes live
One page. Plain language. Signed off by the person who owns the result.
1. The task
One task, described the way you'd brief a new hire. "Handle customer emails" is not a task. "Draft replies to delivery questions and file them for review" is.
2. What it reads
List the inboxes, folders and records it can open. Then list what it should ignore. This decides both quality and cost, as we'll see.
3. What it does alone
Reading, searching and drafting usually run unattended. Anything that leaves the building, spends money or changes a record gets a separate line.
4. When it hands over
Write down the moments the agent stops and calls a person. That includes when it finds nothing. We learned this on a review agent. When we forced it to always produce an answer, it started inventing problems on clean work, because we'd given it no honest way out. The fix was one sentence in its instructions:
Ending without an answer is a failure.
An empty list is a good answer when you found nothing.
Do not invent findings to fill the page.
In plain words: "I found nothing" has to be an allowed answer. Otherwise the agent fills the silence with confident fiction.
5. How it's checked
Decide what success looks like before launch. One go-live checklist recommends recording the baseline first: volume, time spent, cycle time, errors, exceptions. That way the launch gives you evidence, not opinions.
Also check the output itself, not just the logs. Our own agent once spent its whole working budget on research, wrote a complete document, then hit its step limit and returned a one-line "stopped" message. The work existed. Nobody saw it. Now, when an agent runs out of steps, it must hand back what it has.
What should an AI agent never do without a person?
The short list: send, pay, publish, delete. Single Grain advises a gate on anything that spends money or publishes, and Aona adds human review before high-stakes financial, legal or medical outputs. We agree.
The mistake we made was putting the gate in the wrong place. On our own Slack agent, we put an approval step on reading an email attachment, to be safe. The approval worked. The answer never came. The attachment was opened after approval, but the agent never got to read it and reply. Everything showed green, and the user got nothing. That's worse than a refusal, because nothing flags it as an error.
The rule we kept is simple. Approve outcomes, not inputs. A read feeds the answer, so it runs alone, with limits on what can be read. A send is the outcome, so a person confirms it.
Second rule: the approver must see exactly what will run. One risk we check for in every review is an approval screen that shows what the agent asked for, while the system quietly adjusts the request before running it. The person would be signing off on one thing while another happens. So before launch we check that the draft shown and the action that runs are the same thing.
Third rule: spending needs an explicit yes, for that specific charge. A general "go ahead" given the night before doesn't count.
When you don't need an AI agent at all
If you can draw the process as a flowchart with no judgment calls, you don't need an agent. A rules-based workflow in a tool like n8n is cheaper, faster and predictable. A form submission that creates a record and sends a confirmation needs no reasoning.
Agents earn their place when the input is messy: free-text emails, mixed documents, requests that need sorting before anything happens. They also need volume. If the task comes up twice a week, a person does it better.
And if nobody will own the agent's results, don't start. Alex Lieberman's guide lists a single accountable owner, approval thresholds and a kill switch as basics. Without an owner, there's nobody to read the logs.
How we set up an AI agent for a client
We start with the one-page agreement, and we fill it in together. The decision is always one task, one owner, one agent. We widen access only once the logs justify it.
The safeguards are written in before any build. Reads run alone. Sends, payments and record changes wait for a person, who sees the exact action. The agent hands over when it's unsure or finds nothing. Running costs get a daily ceiling, and we track spend in two columns: runs that produced something used, and runs that failed and burned money for nothing. A high share of failed runs means you fix the flow, not the budget.
We also estimate running costs before you commit. During one plan review, we caught a design that would have paid to analyse every image sitting in a chat history each time the agent was called, even for unrelated questions. Cost would have followed how much clutter was lying around, not what anyone asked for. We changed it so only the conversation in progress gets processed. Background items are noted but cost nothing.
The result, in business terms: you know what the agent does, what it costs per month, and who answers for it, before it touches a client. If that's the agent you want, see how we build custom AI agents for one defined task.
