Why we never let the agent decide alone

When we started building Metrivo, the obvious pitch was full automation: let the agent read a request, decide what to do and do it. It demos beautifully. It also breaks trust the first time it gets something wrong, and in operations, the first wrong decision is usually the one that matters most.

So we made a different promise. The agent does the work, and a person stays in charge of the decisions that need one. Here’s how we decide where that line sits.

Human-in-the-loop is a design decision, not a toggle

Many tools treat human review as a setting you switch on or off. That forces a bad choice between an agent that asks about everything and one that asks about nothing. Neither works. The first wastes people’s time, and the second makes them nervous enough to check everything anyway.

We think about every action the agent can take along two lines: how reversible it is, and how costly a mistake would be.

Three questions before the agent acts

1. Can it be undone? Filing a ticket, tagging an owner or posting an update can all be undone in seconds. Sending money, granting access or emailing a customer cannot.

2. Is there a rule that covers it? If a written policy already says yes, the agent follows the policy. If there isn’t one, the agent asks.

3. Would a teammate check first? If a thoughtful new hire would double-check before doing it, so does the agent.

An action that is reversible, covered by a policy and routine gets done straight away. Anything else becomes a clear question for the right person, with the context already attached.

What asking looks like

Asking well is most of the work. A good approval request from the agent reads like a note from a sharp colleague: what was requested, by whom, what the policy says, what is unusual about this one, and a single button to approve or decline. The approver should never have to scroll up a thread to understand what they are saying yes to.

We measure this closely. If approvers regularly open the original thread before deciding, the agent’s summary isn’t good enough yet, and we treat that as a bug.

The agent never approves its own exceptions

One line we will not cross: when a request falls outside a policy, the agent cannot decide that this case is close enough. It can recommend, and it can explain why, but a person makes the call. Over time, the decisions people make on exceptions become the raw material for better policies.

Trust compounds

Teams usually start with the agent asking about almost everything. As they watch it summarize requests accurately and follow policies to the letter, they write more rules and the agent asks less. On mature teams, most routine approvals now go through on policy alone, and people spend their attention on the few requests that genuinely need it.

That trust only exists because people could see every decision along the way. Every action the agent takes is logged with the reason behind it, and anyone on the team can read why something happened.

The rule of thumb

Automate the doing, not the deciding. When in doubt, the agent asks, and it asks well.