Agents work best when the person with the problem is one desk away

Everyone is talking about AI agents. Most of the advice is written for companies with ten thousand employees and a head of AI. A mid-size company does not have that org chart, and it should not pretend to. It may be the better place for an agent right now, for a reason that has nothing to do with being ahead of a trend.

You do not have fourteen layers of approval. The person who owns the problem is often one desk away from the person who can fix it. That proximity is an unfair advantage. An agent is only as good as the description of the work, and the description is better when the owner can lean over and say that a step is wrong, that the customer calls on Thursdays, that the number in the second system is the one finance will believe. A large program can buy a model. It cannot easily buy that correction loop. A company of a few hundred people, or a business unit inside a larger one, already has it, if it chooses to use it.

Where agents earn their keep is narrower than the demos suggest.

Start with the work nobody owns. Chasing invoices. Reconciling a report that three systems already contain. Updating the same fact in the CRM, the sheet, and the finance tool. This work is boring, which is why it survives. It is also specified, which is why an agent can attempt it. If a careful person can do it from a checklist, an agent can draft it. If the checklist only exists as “ask Priya, she knows,” you do not have a candidate yet. You have a dependency.

Next, the questions that take a day to answer. “How did that region do last quarter?” should not need a ticket, a queue, and a hero who knows which extract is current. An agent that can read the certified numbers and show its work is worth more than an agent that can book travel. The value is the hour you did not spend hunting, and the fact that the answer came from a source the company has already agreed to trust. An answer that cannot show its source is a faster rumor.

Then the handoffs. Sales to operations. Operations to finance. The crack is rarely a missing system. It is a fact that lived in one person’s head until they were in a meeting. An agent that carries the agreed fields across that boundary, and stops when a field is missing, removes a class of failure that headcount does not fix. People still own the relationship. The agent owns the remembering. Stopping on a missing field matters more than a fluent summary.

There are places agents do not belong yet, and they are easy to name.

Anything you cannot explain to a new hire. If the process lives as folklore, an agent will improvise, and improvisation at machine speed is not a feature. Write the explanation first. The agent can wait.

Processes that are broken today. An agent just breaks them faster, and it does it in every region at once. Fix the handoff, the source, or the definition of done before you automate the current mess.

Decisions that need judgment you have not written down. Pricing exceptions, a customer you are unwilling to lose, a hire you are unwilling to rush. Those stay with a person until the judgment is explicit enough to audit. “Use your judgment” is a fine instruction for a colleague you trust. It is not an instruction you can review after an agent has applied it a thousand times.

The playbook I would actually use is short.

Pick one painful, boring, repeatable process. One. The temptation is a portfolio. A portfolio is how you learn nothing about any of them. Choose the process people complain about in the same words every week, the one a new hire could shadow in a morning.

Write down how a great employee does it. Not the policy. The steps, the source they open first, the check they do before they send, the case they escalate. Sit with them and read the notes back. The corrections in that conversation are the real specification, and they are the reason the desk has to be close.

Let the agent do the first draft, and have a human approve it. The draft is the product at this stage. Approval is not a lack of ambition. It is how you find out which steps were wrong before a customer, a supplier, or a ledger sees them. Keep the approval with the person who owns the outcome, not with a distant committee. The owner can see a bad draft and say why.

Measure hours saved, not “AI adoption.” Adoption is a slide. Hours are a schedule. If the person who used to do the work cannot point to the afternoon they got back, stop and fix the process or retire the agent.

Only then, give it more rope. A second process. Fewer approvals on the steps that have been right for a month. A wider set of people allowed to invoke it. Rope is earned by a boring track record, not by a launch. Widening access before the track record exists is how a useful draft becomes a system nobody will turn off.

The companies that win with agents will not be the biggest. They will be the ones that know their own business well enough to say, in writing, how the work is done, what done looks like, and who is still accountable when the draft is wrong. A mid-size company can know that. So can a team inside a larger one, if they are allowed to work at the distance of a desk instead of the distance of a program.

The desk is close enough. Use the proximity on one process, keep a person on the approval, and count the hours. That is the whole program until the hours say otherwise.