Article 14 of the EU AI Act expects the people overseeing a high-risk system to stay aware of automation bias, the habit of trusting its output too readily. For agent drafts, that means designing approval so the reviewer must judge rather than click, and measuring whether they catch errors, so you can show a supervisor the review is real.
What does Article 14 say about automation bias?
Article 14 requires high-risk AI systems to be designed so that natural persons can oversee them effectively while they are in use. Among the abilities the oversight measures must support, the Article names one directly: the people assigned to oversight should remain aware of the possible tendency to automatically rely, or over-rely, on the system's output. It singles out systems that provide information or recommendations for decisions taken by people.
The same Article expects overseers to understand the system's capacities and limits, interpret its output correctly, and be able to disregard, override or reverse it. The companion duty to halt a system is covered in Article 14 and the stop button. For the concept itself, see what automation bias is.
Who carries this duty?
The provider must build the system so that oversight is possible and bias can be resisted, for example by presenting output in a way that invites scrutiny. The deployer, under Article 26, must assign oversight to people with the competence, training and authority to do it, and must use the system as instructed.
In practice the deployer carries most of the weight for agents, because it decides who approves, how many drafts they see and how their work is checked. A lender whose underwriters approve agent-drafted decisions, or an insurer whose handlers approve settlement letters, owns the approval design.
Why are agent drafts especially prone to rubber-stamping?
An agent at depth 2 drafts and a person sends. The draft arrives complete, well written and confident. It is usually right. Each of those features pushes the reviewer towards approval.
Volume makes it worse. A reviewer who approves a few hundred correct drafts learns that approval is the right answer, and stops reading closely. The rare wrong draft then goes out with a human signature on it, which is worse than no review at all because it carries false assurance.
The design of the screen matters too. If the draft fills the view and the sources sit behind a link, the reviewer checks the prose rather than the facts. If an approve button is the default action, the path of least effort is to press it.
How do you make the review real?
Build friction where judgement is needed, and measure whether judgement happens.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Overseers understand limits | Approvers trained on the agent's known failure modes for this capability | Training record per approver and capability |
| Output interpreted correctly | Draft shown beside its sources, each marked with provenance; external facts flagged | Screen design and approval records showing sources viewed |
| Awareness of over-reliance | Approver must confirm the decisive facts, not just the draft | Field-level confirmations in the approval record |
| Ability to disregard or override | Edit, reject and return options equal in weight to approve | Override and edit rates per approver |
| Bias detected in practice | Known-wrong test drafts inserted into queues, never sent to customers | Detection rate per approver over time |
| Workload supports care | Limits on drafts per approver per hour; time-to-approve monitored | Queue and timing reports |
| Second line checks the first | Sample of approved drafts re-reviewed independently | Sampling results and follow-up actions |
The seeded test is the most persuasive evidence. If an approver catches deliberately flawed drafts, the review is working. If they approve them, you know before a customer does, and you can retrain, reduce volume or pause that approver.
What will a supervisor ask?
- How often do approvers change or reject the agent's drafts? Has that rate fallen over time?
- How long does an approval take, on average and at the fastest?
- How do you know an approver read the sources and not just the draft?
- What happens when an approver approves a flawed draft?
- Who decides that an approver is competent to oversee this agent?
- Can a customer who complains find out whether a person reviewed their letter?
An edit rate close to zero is a warning, not a success. It means either the agent is perfect or nobody is looking. The second is more likely. For the customer-facing side of explanation, see the ICO and explaining AI agent decisions to people.
What does Article 14 not cover?
It does not set a minimum review time, a maximum queue or a required edit rate; those are your design choices. It does not require a person to review every output of every system; it sets oversight measures proportionate to the risk. It does not make human approval an exemption from other duties: a reviewed decision can still be a solely automated one under GDPR Article 22 if the review is token. And it applies as a binding duty only to high-risk systems, although the same controls make sense wherever agents draft for people. For the full control map, see AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
In Pryme, a held action names the person it is waiting for and shows the facts behind it with their provenance, so external claims stand out. Every approval, edit and rejection is filed with the original draft, which lets you measure edit rates and detection on seeded drafts without building a separate system.
Questions people ask
Does showing the agent's confidence score help?
Often it hurts. A high score encourages approval without reading. Showing which facts are unverified, and where they came from, gives the approver something specific to check.
Should approvers rotate between agents?
Rotation can help keep attention fresh, but only if approvers are trained on each agent's failure modes. An untrained rotating approver is more likely to defer to the draft.
Is a token human review worse than none?
It can be. It adds a signature that suggests scrutiny without providing it, which can mislead customers and supervisors and may not satisfy rules on automated decisions.