When AI agents join a process, a tempting first design is this: let the agents do the work, and have a person approve every step. It sounds responsible. Nothing happens without someone signing it off, so nothing can go badly wrong.
In practice it goes wrong in a quieter way. Ask a capable person to approve fifty routine items a day, and the fifty-first gets a glance, not a decision. The approval stays on the process map, but the judgement has left it. Meanwhile the work waits in a queue for that one person, and the time the agents were meant to give back is spent clicking "approve".
Approval everywhere is approval nowhere. The answer is not fewer people in the process. It is people in the right places.
Put a person where the business commits
A step needs a person when the business is about to commit to something it cannot easily take back. That comes down to four moments.
- Money moves or a price is set. Paying a supplier, issuing a credit, quoting a customer. An agent can prepare all of it. A person releases it.
- Someone outside sees it first. A customer, a supplier or a regulator reads something for the first time. An agent can draft it. A person approves it until the record shows that kind of message is right, time after time.
- A commitment changes. A delivery date, a contract term, the scope of a job, or anything that affects a person's role or pay. These are judgement calls with consequences, and they stay with people.
- The case doesn't fit the rules. An invoice doesn't match its order, or a request falls outside policy. The agent stops and asks. It does not guess. The exception is where your experienced people earn their keep.
Everything else runs without a sign-off: reading, sorting, matching, drafting, chasing and filing. It is checked a different way. Each agent keeps a record of what it did, the owner of the work reviews that record, and a sample is checked often enough to catch drift early.
The line moves, with evidence
Where the line sits on day one is not where it stays. Picture an agent that drafts routine supplier confirmations. After a long run of drafts, each one checked and each one correct, it has earned the right to send the next one without waiting. This is earned autonomy: an agent gains room only when its record shows it can carry it, and the owner of the work decides when that happens.
Some decisions never move. Releasing payments, changing a customer's terms and anything that affects a person's job stay with people, however good the record looks. Write those down at the start. It saves an argument later.
Writing the rules is the creative work
This is where the COO's role changes. When agents take on the routine, the most valuable thing a manager produces is the set of rules the agents work under: what they may do alone, what they must ask about, and what they never touch.
Those rules hold the business's experience. Which customers need a call rather than an email. Which supplier's paperwork needs a second look. Which exceptions are early warnings in disguise.
Little of that is written in any system. Most of it lives in the heads of the people who have done the work for years.
Writing it down is judgement made explicit, and nobody can do it better than those people. That is what "human creativity is the catalyst" means in practice. The people who know the business decide how its agents behave, and then spend their days on the cases that need them.
A test for your next process
Take one process you plan to hand to AI agents.
- Mark every step where a person signs off today.
- For each one, ask: is the business committing to something here, or is someone only checking?
- Keep the commitments. Replace the checks with a record and a sample.
- Name the steps that will never move, whatever the record shows.
If most of your approvals turn out to be checks, your people have been spending their judgement in the wrong places. Moving it to where the business commits is the first gain the agents give you, before they have saved a single hour.
At Futuvara, we build, implement and run AI platforms where AI agents do defined work under rules your business sets. If you are deciding where people should sit in a process your agents will run, Start an Agentic conversation →.
