Resources, Governance
Human oversight for AI agents: a practical design guide
Human oversight means an employee can understand, check and stop what an AI agent does. This guide shows how to design it so it works in daily use, not only on paper.
Updated 20 November 2026. 9 minute read.
In short
- Article 14 of the EU AI Act requires high-risk AI systems to be built so that the employees overseeing them can understand the output, avoid over-reliance, override decisions and stop the system safely.
- Article 26 requires deployers to give oversight to employees with the competence, training and authority to act. These duties apply to Annex III high-risk systems from 2 December 2027.
- Decide which agent actions need approval by asking four questions: can it be undone, does it spend money, does it reach customers, and does it touch personal data?
- Every agent needs a stop control, a named owner and a log of what it did. Review the log on a fixed schedule.
- Too many approval requests lead to rubber-stamping. Approve by risk, batch low-risk work and track how often approvers change the agent's proposal.
What Articles 14 and 26 expect
The EU AI Act splits oversight between the provider that builds a high-risk AI system and the deployer that uses it. Both duties apply to Annex III use cases, such as recruitment and performance management, from 2 December 2027, after the Digital Omnibus on AI moved the date.
Article 14: built for oversight
The provider must design the system so that it can be effectively overseen while in use. The aim is to prevent or minimise risks to health, safety and fundamental rights. Article 14(4) says the employees assigned to oversight must be able to:
- understand the system's capabilities and limits and monitor its operation, including spotting anomalies and unexpected behaviour;
- remain aware of automation bias, the tendency to rely too much on the output;
- correctly interpret the output;
- decide not to use the system, or disregard, override or reverse its output;
- intervene or interrupt the system through a stop button or a similar procedure that brings it to a halt in a safe state.
Article 26: oversight in use
The deployer must use the system in line with the provider's instructions and assign human oversight to employees who have the necessary competence, training and authority, with the support they need. The deployer must also monitor the system, keep its logs for at least six months and suspend use if it presents a risk.
GDPR Article 22
Where an agent makes decisions about individuals that have legal or similarly significant effects, GDPR Article 22 also applies, whatever the AI Act risk class. The guidelines on automated decision-making, endorsed by the European Data Protection Board, say human involvement must be meaningful, not a token gesture. An employee who signs off every output without looking does not count.
Even when an agent is not high-risk, the same design works as good practice. It is also the easiest way to show a customer or auditor that your agents are under control.
Levels of autonomy
The AI Act does not define levels of autonomy. A simple scale helps you decide, for each agent and each action, how much it may do alone. The four levels below are a working model, not a legal category.
- Suggest: the agent proposes, and an employee does the work. Example: an agent drafts a reply that an employee edits and sends.
- Act with approval: the agent prepares the action and waits for an employee to approve it. Example: an agent prepares a refund and an employee approves it.
- Act and report: the agent acts within set limits and reports afterwards. Example: an agent tags and routes incoming tickets.
- Act alone: the agent acts within its role and is reviewed periodically. Example: an agent updates internal meeting notes in the company memory.
One agent can work at different levels for different actions. A support agent may route tickets alone, draft replies for approval and never issue refunds above a set amount. Set the level per action, not per agent.
Deciding which actions need approval
Ask four questions about each action an agent can take. A yes to any of them is a reason to move the action up to approval, or to put a limit on it.
- Reversibility: can the action be undone fully and cheaply? Sending an email, deleting data or signing a contract cannot. Saving a draft can.
- Money: does the action commit or move money, or run up significant AI usage? Payments, purchases, refunds and price changes belong here.
- Customers: does the action reach a customer, a candidate or the public? Anything sent outside the company carries your name.
- Personal data: does the action read, change or share personal data, especially special categories such as health data? GDPR principles of purpose limitation and data minimisation apply.
Two further checks apply. First, if the action contributes to a decision about an individual in an Annex III area, such as hiring or promotion, the AI Act high-risk rules may apply. Second, if the decision is based solely on automated processing and has a significant effect, GDPR Article 22 applies. In both cases, an employee must take the decision, with the agent in a supporting role.
A practical default: internal and reversible actions can run alone, external or irreversible actions need approval, and actions on money or personal data get a limit as well as an approval rule.
Who approves, and what they need
Article 26(2) asks for three things in the employee who oversees an AI system: competence, training and authority. Authority is often missed. An approver who cannot reject an agent's proposal, or who fears doing so, is not providing oversight.
- Competence: the approver knows the business task well enough to judge the result. A refund is approved by the support lead, not by IT.
- Training: the approver knows what the agent does, where it tends to fail and how automation bias works. This also meets the AI literacy duty in Article 4.
- Authority: the approver can reject, change or stop the work without further sign-off, and is backed by management when they do.
- Capacity: the approver has time in the working day to review properly.
Name a primary approver and a deputy for each approval rule, so work does not stall during holidays. Record the names. For high-risk use, the deployer must be able to show who was assigned to oversight.
Stop controls, logging and review
Stop controls
Article 14(4)(e) names a stop button or a similar procedure that halts the system in a safe state. For an agent, that means more than closing a browser tab. Plan three levels of stop:
- pause one task, so the agent stops a single piece of work;
- pause one agent, so it takes no new actions until an employee restarts it;
- revoke access, so the agent loses its connection to an app or tool at once.
A spending limit that pauses the agent automatically is also a stop control. Test every stop control before go-live and after each change.
Logging
Log what the agent was asked to do, what it did, which tools it used, who approved each step and what it cost. For high-risk systems, Article 26(6) requires deployers to keep the logs under their control for at least six months, unless other law, in particular data protection law, sets a different period. Logs that contain personal data are themselves subject to GDPR, so set a retention period and limit access.
Review
Set a review cycle per agent. Monthly works for most. Look at a sample of actions, the approvals that were rejected or changed, incidents and costs. Review again whenever the model, the instructions or the tools change. Under Article 26(5), a deployer that sees a risk must inform the provider and suspend use.
Avoiding approval fatigue
If every action needs approval, approvers stop reading. This is the automation bias Article 14 warns about, and it turns oversight into a formality. Design against it from the start.
- Approve by risk, using the four questions above. Low-risk, reversible work should not wait in a queue.
- Batch similar low-risk items into one review, such as a daily list of routed tickets.
- Show the approver what matters: the proposed action, the reason, the data used and the cost. Do not ask them to read a full transcript.
- Use limits instead of approvals where you can. A refund limit of a fixed amount is easier to manage than approving every refund.
- Track the rate at which approvers change or reject proposals. A rate near zero over many weeks can mean the rule is too strict or the approver is no longer checking.
- Rotate approvers on high-volume rules, and give them time for the work.
Relax a rule only on evidence. When an agent has handled an action well for a set period, move it down one level and keep reviewing the log.
A worked example: a support agent
A company uses an agent to handle customer support emails. It can read the shared inbox, look up orders, draft replies and issue refunds. This use case is not in Annex III, so it is not high-risk, but the company applies the same oversight design.
- Tagging and routing incoming emails is internal and reversible. Level: act and report.
- Replying to customers reaches customers and cannot be unsent. Level: act with approval. The support lead approves replies in a queue.
- Refunds move money and cannot be undone easily. Refunds up to a fixed amount need approval from the support lead. Anything above that needs the finance manager.
- Order lookups touch personal data. The agent can read only the order system, not the full customer database, and only for the ticket it is handling.
- A monthly AI usage budget sends a warning at 80% and pauses the agent at the limit.
- The support lead can pause the agent at any time. IT can revoke its access to the inbox.
- Every action and approval is logged. Each month, the support lead reviews 20 sampled replies and every refund, and checks how often replies were edited before approval.
After three months with few edits, the company lets the agent send replies alone for order status questions only, and keeps approval for complaints and refunds. The customer is told in each reply that it was drafted with AI assistance.
How Blentera helps
Blentera is a platform where companies run AI agents with employees in charge. Each agent has a role in the org chart, and approval rules can be set per agent, project, app or tool. Budgets warn at 80% and pause the agent at the limit. Agents share one company memory, and an activity log records their actions and can be exported. An EU AI Act use-case register checks requirements such as human oversight and review dates before an agent goes live. It is hosted in the EU. The oversight design itself, and meeting the law, remain your company's responsibility.
Questions
- The specific duties in Articles 14 and 26 apply to high-risk AI systems, which for Annex III use cases means from 2 December 2027. For other agents, oversight is not mandated in the same form, but the AI literacy duty in Article 4 and GDPR Article 22 can still apply.
- Article 14(4)(e) requires that the employees overseeing a high-risk system can interrupt it through a stop button or a similar procedure that brings it to a halt in a safe state. For an agent, this usually means pausing tasks, pausing the agent and revoking its access to tools.
- No. Approving everything leads to approval fatigue and automation bias, which Article 14 asks you to guard against. Approve by risk, use limits and batch low-risk work so that approvers can look closely at what matters.
- An employee with the competence to judge the task, training on how the agent works, and the authority to reject or stop it, as Article 26(2) requires for high-risk systems. Usually this is the business owner of the task, with a named deputy.
- For high-risk systems, deployers must keep logs under their control for at least six months, unless other law provides otherwise. Set the retention period with your data protection officer, because logs often contain personal data.
Sources
- Regulation (EU) 2024/1689 (AI Act), consolidated text of 27 July 2026, EUR-Lex
- Article 14: Human oversight, AI Act Service Desk
- Article 26: Obligations of deployers of high-risk AI systems, AI Act Service Desk
- AI Act: regulatory framework and timeline, European Commission
- Regulation (EU) 2016/679 (GDPR), EUR-Lex
- Guidelines on automated individual decision-making and profiling (WP251rev.01), endorsed by the EDPB
- Endorsed WP29 guidelines, European Data Protection Board
Related
This guide is general information, not legal advice. Check how the rules apply to your company with your own advisers.
Get started
Put your first agent to work in minutes.
Three steps, and your first agent is working for your company.
Step 1
Create an account
Sign up with your work email and invite your team. No sales call needed.
Step 2
Add your first agent
Create an agent, give it a role, and connect your own AI key and the apps it may use.
Step 3
Give it a task
Ask in Blentera or your chat app, and follow the work until it is done.