Platform
Make responsibility explicit, even when AI does the work
Responsible, accountable, consulted, informed. AI can be responsible for a task without becoming accountable for the outcome — accountability stays human.
Accountability, counted
Systems
11
Accountable people
6
Every system has exactly one accountable person. Nothing here is answered for by a team, a rota or an alias.
The distinction that matters
An agent can be responsible. It cannot be accountable.
R is about doing the work, and work is delegable — including to software. A is about answering for the outcome, and that has never been delegable to anything without something at stake. Most AI governance confusion is these two letters being treated as one.
- R — Responsible
- Does the work. This is the letter an agent can hold, and increasingly does. It is about execution, and execution is delegable.
- A — Accountable
- Answers for the outcome. Always a named person, never a team and never software. Exactly one per system, because two accountable people is none.
- C — Consulted
- Must be asked before it proceeds. This is the letter that becomes a co-signer when an action crosses a bound.
- I — Informed
- Told afterwards, from the record rather than by someone remembering to send an email. The evidence ledger is what makes this letter automatic.
Automating the work does not automate answering for it.
This is why OpsAI attaches a named person to every AI system before it attaches a policy. A rule with no owner is a preference, and an incident with no accountable party becomes an organizational problem rather than an operational one.
Why A stays human.
Three separate reasons, and each is sufficient on its own. They are worth separating because teams usually concede the third and then design as though the first two were negotiable.
- It has nothing at stake
- Accountability works because consequences land somewhere. Software has no licence to lose, no performance review and no reputation. Assigning it A does not distribute risk, it deletes the record of who holds it.
- It cannot exercise judgement about its own scope
- An accountable party has to be able to say "this was within what I agreed to, and that was not". An agent evaluates its bound; it cannot hold an opinion about whether the bound was the right one.
- Regulators and courts do not accept it
- Every regime that matters here attaches obligations to a person or a legal entity. An estate that has assigned A to a system has not transferred the obligation, only obscured where it sits.
Every system, assigned
One accountable person per row, and no blanks.
The two columns to read together are Accountable and Runs for. When they disagree, the row will pass an audit and fail an incident — which is precisely the kind of thing a register exists to surface rather than smooth over.
RACI assignment
11 systems · 1 owned across a team boundary| AI system | R · Responsible | A · Accountable | Runs for | Chain depth |
|---|---|---|---|---|
| refund-resolver | refund-resolverL3 Act with approval | Priya NairSupport Ops | Support Ops | 1 of 3 |
| order-lookup | order-lookupL1 Observe | Priya NairSupport Ops | Support Ops | 2 of 3 |
| ap-invoice-agent | ap-invoice-agentL2 Advise | Rahul MenonFinance | Finance | 1 of 3 |
| payout-runner | payout-runnerL4 Act autonomously | Rahul MenonFinance | Finance | 2 of 3 |
| vendor-onboard | vendor-onboardL3 Act with approval | Anita RaoProcurement | Procurement | 1 of 3 |
| inventory-sync | inventory-syncL2 Advise | Anita RaoProcurement | Procurement | 1 of 3 |
| tier1-support | tier1-supportL2 Advise | Priya NairSupport Ops | Support Ops | 1 of 3 |
| crm-hygiene | crm-hygieneL1 Observe | Sneha IyerRevOps | RevOps | 1 of 3 |
| claims-triage | claims-triageL2 Advise | Devang ShahClaims | Claims | 1 of 3 |
| contract-reader | contract-readerL1 Observe | Kabir SenLegal | Legal | 1 of 3 |
| dunning-agent | dunning-agentL2 Advise | Sneha IyerRevOps · cross-team | Finance | 1 of 3 |
The other direction
Sorted by person, it is a load report.
A register sorted by system tells you nothing is unowned. Sorted by person it tells you something different and more useful: whether accountability is distributed or has quietly concentrated on whoever was willing to sign.
- Priya NairSupport Ops
- refund-resolverL3
- order-lookupL1
- tier1-supportL2
3 systems · highest L3 Act with approval
- Rahul MenonFinance
- ap-invoice-agentL2
- payout-runnerL4
2 systems · highest L4 Act autonomously
- Anita RaoProcurement
- vendor-onboardL3
- inventory-syncL2
2 systems · highest L3 Act with approval
- Sneha IyerRevOps
- crm-hygieneL1
- dunning-agentL2
2 systems · highest L2 Advise
- Devang ShahClaims
- claims-triageL2
1 system · highest L2 Advise
- Kabir SenLegal
- contract-readerL1
1 system · highest L1 Observe
Accountable people
6
named, not aliased
Most systems held
3
by one person
Carrying an L4
1
unattended autonomy
Unassigned systems
0
a blank is a finding
IllustrativeThe OpsAI sample estate. The same rows, counted differently.
How far R can travel
Responsibility delegates. Accountability stays where it started.
An agent handing work to another agent is R moving down a chain. A does not move with it — it remains attached to the person at the root, which is what stops a four-hop chain from ending in nobody being answerable.
One chain, in full
depth 1 of 3- Priya NairSupport Ops
- refund-resolversubset of parent · L3
- depth 1 of 3
- L1ObserveReads. Never writes.Light, targeted checks3
- L2AdviseProposes. A person commits.Output quality checks5
- L3Act with approvalActs inside a bound. Above the bound, a person co-signs.Meaningful review · audit trail2
- L4Act autonomouslyActs inside a bound, watched continuously, circuit breaker armed.Circuit breaker armed1
Autonomy is how much R an agent may hold without a person in the loop. It is a separate axis from risk: a level-1 system can still request a critical action, it simply cannot commit one. Raising a level is an accountable person's decision, and dropping one is automatic when a bound is broken.
payout-runner is the only system here at L4. It runs unattended, so it is watched continuously and its circuit breaker is armed — and it still answers to Rahul Menon. What a breaker trip looks like.
C, when it matters
Approvals
The consulted party becomes a named co-signer the moment an action crosses its bound.
What A agreed to
Policies
A rule is owned by the team carrying the risk, which is what makes accountability concrete rather than nominal.
How I gets informed
Audit
From the record, automatically, rather than by someone remembering to circulate it.
Where to start
Name one person per system, then argue about the rest.
Assigning A is the cheapest governance work available and the hardest to schedule, because it requires someone to accept something. It is also the step every other control depends on.