Trust and control
Adopt AI without losing the audit trail.
For anyone who has to answer for what an agent did: every action it takes here is limited, traceable, and signed off by a person you can name. An agent can't step outside what you allowed - the action isn't blocked after the fact, it was never available to it. You decide what needs sign-off, and every decision, human or AI, lands on the same record.
The problem
The next AI pilot has to be defensible before it ships, not after.
Your board asks who is accountable for a decision an agent made months ago. An auditor asks you to show the agent could not have gone beyond what it was allowed, not that it happened not to. Most AI deployments can't answer either question, because the agent's boundaries live in a prompt and a policy document, not in the system it acts through.
How it works
Every agent action plans01 from what the operation actually looks like, decides02 inside limits a person already set, and acts03 onto one record: human and AI, same trail.
01
The objective isn't guessed
What an agent may propose comes from where it sits in the work, not from a prompt someone wrote once and forgot.
02
Limits that actually hold
An out-of-bounds action isn't rejected after the fact. It was never something the agent could pick.
03
One record, always
Every AI and human decision, including the ones you turned down, lands on the same record.
Accountability
You draw the line. We hold it.
You own
- What an agent is allowed to propose at all.
- Which of those actions need a human sign-off.
- The sign-off itself, recorded against your name.
The AI owns
- Staying inside the limits you set, every time.
- Putting every step on the same record as human decisions.
- Showing the evidence behind every proposal.
Single sign-on, on-premise deployment, data residency and full audit-trail export are part of the Enterprise tier. What that involves, and what we do not claim yet .
What we can show today
Enforced by architecture, not by a policy someone can bypass.
Because the limits sit in the platform, an agent's bounds are the same whether or not anyone is watching, and you can show an auditor what an agent could not have done, not only what it did not do. An operator running fluvial logistics on the Seine works this way today: every dispatch confirmed by a named owner, dismissals recorded alongside approvals, one trail for both. What we cannot show you yet is an external audit of that trail. The seven constraints behind it are set out in the manifesto.
What this is not
- Isn't this just SSO and an export button?
- We have both, but that's not the point. The point is architectural: an out-of-bounds action can't be expressed by the agent in the first place, not caught after the fact.
- How is this different from logging AI decisions after they happen?
- A log tells you what already went wrong. The constraints make the wrong action impossible to propose in the first place.
- What happens when a human overrides the AI?
- It's recorded in the same audit trail as every other decision, with the reason attached.
See where the line sits.
Put one workflow through it. Set what an agent may propose and what needs your sign-off, then read the record it leaves behind - approvals, dismissals and the evidence attached to each.
Metronome is in beta. You create a workspace and configure it against your own workflow - nothing to install, no data migration to begin.