The named human who owns the outcome. In MING Labs’ operating model the assignment is made before the agent starts working, not reconstructed after something goes wrong: an agent can be Responsible for a work unit, but it is never Accountable. Accountability stays with the human whose name is on the firm.[S1]
01The rule, on a real workflow
Abstract principles about AI accountability are cheap, so here is the RACI of an actual workflow in MING Labs’ own fleet: proposal generation. The agent drafts the proposal from the meeting transcript, as Responsible. A senior partner is Accountable and owns the outcome; they win the deal or lose it. A co-founder is Consulted on positioning when it matters. The operator is Informed through the fleet logs.[S1]
Nothing in that assignment is novel, which is the point. It is the same RACI the firm would write for a junior consultant drafting the same proposal. Hiring an agent into the role changed who does the drafting. It changed nothing about who answers for the result.
02Three mistake cases, three written answers
The reason the question “who is accountable” stays boring at MING Labs is that each mistake case already has a written answer when it happens.
| Mistake case | Who answers for it | Where that is written down |
|---|---|---|
| A bad draft reaches review | The Accountable owner, whose review exists to catch it | The workflow RACI in the agent’s job description |
| A mistake made while acting autonomously | The named human who granted that autonomy level | The authority log: every grant carries a timestamp, a reason, and an approver |
| Repeated failure in the role | The owner, who decides demotion or retirement | The incident record; two incidents in a quarter revert the level |
The second row does the heavy lifting. Autonomy is where accountability usually goes to die in agent programs, because autonomy is treated as a configuration value that nobody remembers setting. In MING Labs’ fleet every change to an agent’s authority, promotion, demotion, scope expansion, or scope reduction, is a logged event with a named approver, the same discipline as a human role change.[S1] When something goes wrong at level three, the interesting question is not philosophical. It is a lookup.
03What regulation expects
The EU AI Act draws the same line for high-risk systems. Article 26 requires deployers to assign human oversight to natural persons who have the necessary competence, training and authority, and to keep the system’s automatically generated logs for at least six months.[S2] Named humans with real authority, plus a record that shows what happened: that is the regulatory shape of accountability, and it is the shape an agent fleet has to produce on demand, whatever the jurisdiction. An operating model where accountability is assigned per workflow and autonomy is a logged grant produces it as a side effect of normal operation, not as a compliance project.
04The claim, scoped
Scope
- The rule: an agent can be Responsible for a work unit and own it operationally. Organisational accountability remains with a named human.
- Where it runs: MING Labs’ own production fleet, every agent under a named human owner, since January 2026.
- Enforcement artefacts: per-workflow RACI in each agent’s job description; an authority log for every autonomy change; incident-triggered reversion (two per quarter drops the level).
- Not claimed: legal advice, or a statement about how liability is allocated in any specific jurisdiction. Legal liability regimes differ; this page describes the operating design that keeps a named human answerable.
Accountability is only real while the owner could still do the work. An owner who cannot reproduce what their agent produces is accountable on paper and helpless in practice, which is why every named owner at MING Labs rehearses exactly that, quarterly, under the Comprehension Obligation . The same ownership discipline governs the other end of the agent’s tenure: retirement is the owner’s decision too, on a published criterion, as described in when to shut down an AI agent . Both are what it looks like to run agents as colleagues inside a hybrid organisation : the org chart stays honest, and someone’s name is always on the outcome.
