Who Is Accountable When AI Prepares the Decision and a Person Signs It?

In 96 percent of AI use cases the AI prepares and a person decides. That makes the human in the loop a role that HR has to design, staff and support.

MetroMinds · 18 October 2026 · 7 min read

A grid connection analyst at a power utility approves an AI-prepared capacity exception on a busy Thursday afternoon. The agent had gathered the load study, checked the capacity register and staged the approval request. Three weeks later it turns out the load study was outdated, and a connection date has to be withdrawn. In the review meeting the question comes up quickly: when the approves what an agent prepared, who is accountable for the outcome? The analyst who clicked approve, the manager who set the thresholds and the team that configured the agent all point in different directions.

Questions like this are becoming routine as AI moves from answering questions to preparing decisions. Most organisations say they keep a "human in the loop", but far fewer have worked out what that human is responsible for, what authority they have and what support they need. That is work for HR as much as for IT or risk.

AI prepares, people decide

Every use case in the MetroMinds Library is scored on : how much room the use case gives the AI to act without a person. In 4,800 of the 4,980 use cases, or 96 percent, the score is 1 or 2 on a five-point scale. In practice, the AI gathers, checks, ranks and drafts, and a person makes the decision.

Those decisions are often heavy. Almost half of the use cases (2,288) score 4 or 5 for the impact of errors, and 2,144 combine high error impact with low action latitude. That combination describes a very specific job: a person who decides on consequential matters, using material that an AI has prepared.

Bubble chart of 4,920 MetroMinds Library use cases by action latitude and impact of errors. All use cases score 1 to 3 on action latitude; the area for scores 4 and 5 is empty. The largest group, 998 use cases, combines latitude 2 with the highest error i — Source: MetroMinds Library, 4,980 use cases (August 2026). 4,920 use cases with both scores; bubble size shows the number of use cases.

The largest single group in the chart, 998 use cases, combines an action latitude of 2 with the highest possible error impact. That is where the human in the loop carries the most weight, and also where the risk of a quick, uncritical approval is greatest.

From the MetroMinds Library

Should a feeder capacity exception move without a human signature?

Industry Energy and Utilities, Power Utilities
AI domain Agentic AI, human-governed autonomy
Role Grid connection analyst, with the network engineer above a threshold
Capability Operations and Delivery > Service Fulfillment > Order and Request Fulfillment

The agent reads the connection case, the capacity register and the connection policy, prepares the studies it may run alone and turns anything that changes a committed capacity position into an approval request. The analyst accepts, corrects, rejects or overrides every staged step, and the exception stays open until the decision and its reason are recorded.

15 to 35% handling time reduction
10 to 25% review time reduction
8 to 20% rework reduction

Library note: accountability stays with the human approver because the AI prepares action rather than owning authority. Workforce impact: less status chasing, more formal supervision of exceptions and override logs.

Book a demo

Accountability needs a name, authority and support

The Library descriptions are explicit about who carries responsibility. In 3,960 use cases, 80 percent of the total, the workforce text states that accountability remains with a named human role, such as the business owner who approves the action. In 4,712 use cases, 95 percent, a second role is defined that joins only when a threshold is crossed, for example a cost limit, a safety concern or a legal question.

Regulation points in the same direction. The EU AI Act requires of high-risk AI systems to "assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support" [1]. The OECD AI Principles state that organisations and individuals deploying or operating AI systems "should be held accountable for their proper functioning" [2]. Both texts treat oversight as a role that has to be designed, staffed and supported, which is classic HR territory.

For a CHRO, that turns an abstract principle into concrete design work. The authority of the approving role needs to be written down, including where it ends. The role description should mention the review of AI-prepared work. People need training to recognise when the AI is likely to be wrong, and the performance system should reward careful decisions rather than speed alone.

The quiet risk: approving without really looking

A human in the loop only adds value if that human actually judges the output. The EU AI Act names the risk directly: people assigned to oversight must be able "to remain aware of the possible tendency of automatically relying or over-relying on the output" of an AI system, which the Act calls automation bias [1]. They must also be able to decide "not to use the high-risk AI system or to otherwise disregard, override or reverse the output" [1].

Workload is one of the strongest drivers of automation bias. When the volume of AI-prepared decisions grows faster than the time people have to review them, approval turns into a routine click. The example below shows how quickly that can happen.

Scenario

A sign-off chain for supplier payment holds

A finance team introduces an agent that reviews supplier payments and prepares a decision for each one it puts on hold: release, keep on hold or escalate. The numbers in this example are illustrative. An accounts payable specialist may release holds up to 10,000 euros. Above that, the controller decides, and above 250,000 euros the CFO signs. The agent never releases a payment itself.

In the first month the agent prepares about 40 holds a day per specialist, and each one gets real attention. After the rollout to all suppliers, that number rises to 120. Over an eight-hour day, a specialist then has an average of four minutes per decision, including reading the evidence the agent attached. The approval rate climbs, override reasons become shorter and nobody notices until an audit sample finds a duplicate payment that was released without a check.

120 holdsPrepared per specialist per day after the full rollout
4 minutesAverage time left per decision in an eight-hour day
3 levelsOf authority, each needing its own limits, training and override rules

The fix sits largely with HR and line management: realistic volumes per role, clear authority limits in role descriptions, training on typical AI errors and performance goals that value good overrides.

What changes in the role itself

When AI prepares decisions, the content of the approving role shifts. The comparison below shows the difference for a typical operational role, based on the patterns in the Library descriptions.

That second version is a different job. It asks for other skills, which we cover in our article on the skills people need when AI takes over routine work. It also deserves a different role description, other performance measures and a realistic workload.

Questions to settle before go-live

Which role approves each type of AI-prepared decision, and is that authority written down with clear limits?
Which second role takes over above the threshold, and does that person have time for it?
How many AI-prepared decisions can one person review well in a day, and who monitors that volume?
Do people know the typical errors of the AI they work with, and are they trained to override it?
Does the performance system reward careful review and well-reasoned overrides, or mainly speed?

The answers belong in the design of the use case, not in an incident review afterwards. For a broader view of how AI changes jobs outside HR's line of sight, read what AI will change in your workforce, and why HR hears it last.

How MetroMinds makes accountability visible

MetroMinds shows the workforce impact of every use case, including which role keeps decision authority, which second role joins above a threshold and how the work of each role changes. The profile includes the scores for action latitude and impact of errors, so you can see early which use cases put the heaviest decisions on people.

In the portfolio, decisions about use cases follow the same principle. A use case can only be rejected or put on hold with a written reason, and roles in the platform are fixed: contributors submit, portfolio managers decide and customer admins manage access. That gives HR, risk and the business one place to agree who is accountable before a use case moves to .

Want to see how MetroMinds maps roles and accountability across your AI portfolio? Book a demo

Sources

European Union, Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14 "Human oversight" and Article 26 "Obligations of deployers of high-risk AI systems", Official Journal, 12 July 2024.
OECD, Recommendation of the Council on Artificial Intelligence (OECD AI Principles), Principle 1.5 "Accountability", adopted 2019, revised 2024.
MetroMinds Library, 4,980 AI use cases with benchmark ranges based on real-world implementations, August 2026.