RACI for Machines - When the agent acts, who carries the consequence? VII - Internal Hesitation - Act II - When Our Agents Meet

When an agent acts, who is accountable?

An agent drafts the response, checks the policy, updates the record, triggers the workflow and sends the notification.

Something goes wrong.

The first governance question is not:

What did the AI do?

It is:

Who owned the action?

That question will become harder to answer as agentic AI moves from clever demonstration to ordinary work. A chatbot suggests. An agent acts. It may search, summarise, classify, draft, decide the next step, update a system or hand work to another agent. The more useful it becomes, the more it starts to occupy a space that used to belong to people, teams and processes.

That is where the old governance tools begin to creak.

RACI is one of those tools. It is simple, useful and often slightly unloved. Responsible. Accountable. Consulted. Informed. Four words that try to stop work becoming fog. Who does the work? Who owns the outcome? Who needs to be asked? Who needs to know?

In human organisations, that is already difficult enough.

With machines, it becomes stranger.

Can an AI agent be Responsible? In one sense, yes. It may perform the task. It may draft the email, extract the data, update the record or trigger the workflow. If we are only asking who carried out the mechanical action, the agent may be the obvious answer.

But that is not enough.

A machine can be assigned a task. It cannot be assigned accountability.

That sentence matters because accountability is not just about causation. It is not only about identifying which thing made something happen. Accountability belongs to the world of judgement, authority, explanation and consequence.

A machine can produce an output. It cannot appear before a board and explain why the action was appropriate. It cannot understand the public duty behind a decision. It cannot carry reputational responsibility. It cannot apologise in a meaningful way. It cannot be disciplined, trained through experience or asked to exercise moral courage next time.

The organisation can use the machine. The organisation cannot hide inside it.

This is why “the AI did it” must never become an acceptable governance answer.

If an agent acts, the organisation needs to know who authorised the action, who designed the process, who approved the control, who monitored the result and who is accountable if harm follows. Otherwise, agentic AI becomes a perfect fog machine. Work happens, but ownership dissolves.

Classic RACI was not designed for this.

It assumes that the work is done by people or teams. It assumes that responsibility and accountability can be attached to human roles. It assumes that if something needs to be challenged, escalated or explained, there is someone in the structure who can answer.

Agentic AI breaks that neatness. Not because machines become accountable, but because machines can now do enough work to make accountability harder to see.

So perhaps the answer is not to abandon RACI, but to extend it.

For AI-enabled processes, we may need RACI+E:

• Responsible: the human role responsible for ensuring the work is completed properly
• Accountable: the human role answerable for the outcome and consequences
• Consulted: the people or functions whose expertise, consent or challenge is needed
• Informed: the people who need to know what has happened
• Executed by: the system, agent or automation that performs the action

That extra category matters.

It allows the agent to be named without pretending it is accountable. It shows where the machine acts while preserving the human chain of responsibility around it. It stops the organisation from quietly moving the work into an automated space and then losing the line of ownership.

For example, imagine an AI agent used to draft a response to a resident.

The agent may be Executed by.
A customer services officer may be Responsible for checking and sending.
A service manager may be Accountable for the quality and appropriateness of the response.
Legal, policy or safeguarding colleagues may be Consulted for certain categories.
The resident, case owner or partner service may be Informed.

That is very different from saying:

The AI writes the response and a human reviews it.

That sentence sounds reassuring, but it hides too much.

Which human? What are they reviewing for? Accuracy? Tone? Legal compliance? Equality impact? Safeguarding risk? Financial consequence? Does the human have enough time to review properly? Can they reject the output? Are they expected to check the source material or merely approve a polished draft?

A human in the loop is not a role. It is a hiding place unless the organisation defines what that human can approve, challenge, stop and be held accountable for.

This is where RACI+E becomes useful. It forces the process to name the point of control.

If the agent drafts, who checks?
If the agent recommends, who decides?
If the agent updates a record, who owns the accuracy?
If the agent triggers a workflow, who can stop it?
If the agent hands work to another agent, who owns the chain?

These are not technical questions. They are governance questions.

In public services, this matters because many actions are not just transactions. They are exercises of authority, discretion or care. A letter, a case note, a triage decision, a translation, a benefit recommendation or a referral may look like administration, but it can still affect someone’s life.

The risk is not only that the agent gets something wrong.

The risk is that the organisation cannot clearly say who was meant to catch it.

That is why agentic AI needs more than model evaluation. It needs process ownership. It needs audit trails. It needs decision rights. It needs escalation routes. It needs a clear line between assistance, execution and judgement.

Those distinctions matter.

An agent that assists is helping a human do the work.
An agent that executes is carrying out an action within a defined boundary.
An agent that recommends is shaping judgement.
An agent that decides is crossing into authority.

The further along that line the agent moves, the stronger the governance needs to be.

Some uses of AI will be low risk. Summarising a long public report for internal briefing may be relatively safe if the source is visible and the summary can be checked. Drafting a first version of a meeting agenda may be harmless. Sorting documents into broad categories may be useful if errors are easy to reverse.

Other uses are different.

A case note that becomes part of a person’s record.
A prioritisation decision that affects service access.
A financial assessment.
A legal interpretation.
A safeguarding referral.
A response to a vulnerable resident.
A workflow that closes, escalates or delays someone’s request.

In those cases, the organisation needs to know not just that AI was involved, but how it was involved.

Was it a drafting tool?
Was it a recommender?
Was it an executor?
Was it allowed to update the system?
Was the action reversible?
Was the human review meaningful?
Was the evidence visible?
Was the decision recorded as human judgement or machine output?

This is also where the link to the hallucination ceiling becomes important.

The problem is not only that an AI can hallucinate facts. It can hallucinate completeness. It can make a half-checked process look finished. It can turn ambiguity into fluent prose. It can create a plausible chain of action that nobody fully owns.

RACI+E is one way of resisting that.

It says: before we are impressed by what the agent can do, let us name who remains responsible for the work.

There is a useful discipline in that. It prevents the demo from becoming the governance model. It stops “the agent will handle that” from replacing thought about risk, evidence and accountability. It keeps human authority visible at the points where the organisation owes people an answer.

This does not mean every AI-assisted task needs a huge governance framework. That would be absurd. The point is proportionality.

Low-risk, reversible, internal tasks need light controls. Higher-risk tasks need clearer ownership. Decisions that affect rights, access, safety, money or legal standing need explicit accountability.

The more real-world consequence the agent can create, the less acceptable it is for ownership to be vague.

That is the practical test.

Not “Can the agent do it?”
Not “Is the model accurate enough?”
Not “Would this save time?”

But:

If this goes wrong, can we explain who owned the action and why the control was good enough?

If the answer is no, the agent is not ready for that process.

This is why councils and other public bodies should be cautious about adopting agentic AI through productivity alone. Saving time is attractive, especially when capacity is thin. But the fastest process is not always the safest process. And when public authority is involved, safety includes fairness, explainability, auditability and the ability to challenge what happened.

RACI for machines is not about giving machines a place in the hierarchy.

It is about stopping them from creating a gap in it.

The machine may execute. The human organisation remains accountable.

That may sound obvious now. It will become less obvious as agents become more capable, more embedded and more ordinary. Once the workflow runs smoothly, people may stop noticing where the machine enters the chain. The output will simply arrive, formatted and ready, carrying the quiet authority of completed work.

That is when governance matters most.

Not at the edge of failure, but at the point where automation begins to feel normal.

Because the question is not whether machines can act.

They already can.

The machine may execute the work, but it must never become the place where accountability goes to disappear.