AI is moving quickly from assisting people to leading work. Anthropic internal reports state that work in which AI assists peaked in November 2025, work in which it collaborates under close human direction peaked in May 2026, and work it leads, completing most of a task end to end while a person supervises, has risen from under 1% to 26% since February 2026.
Previously, models summarised documents, drafted emails and suggested next steps, and a person decided what happened next. However, AI agents change that relationship, with the ability to update records, issue refunds, send communications, modify configurations and trigger workflows across connected systems, often completing multi-step tasks with little or no human involvement.
This shift turns a familiar governance question into a design decision. When an agent can act, someone has to determine what it is allowed to do, within what limits and with whose approval. Organisations already answer these questions every day for their people, through delegated authority, approval limits and segregation of duties. Agents need the same treatment, and many organisations are discovering this only after something has gone wrong. In May 2026, Gartner predicted that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps identified only after production incidents.
This article sets out how technology leaders can approach agent authority deliberately, why capability is a poor guide, which dimensions of authority can be set, how consequence and reversibility should shape those choices, and how oversight and accountability can keep pace as agents take on more of the work.
Capability is a poor guide to authority
In many agent deployments, authority follows capability. Agents built for exploratory proofs of concept are often given loose permissions, such as a service account with broad access, a full set of tools and few limits on what they can change. At that stage the risk is low, volumes are small and someone is usually watching. Problems tend to surface later, when the same agent moves into production with the same permissions and meets an unusual input, a manipulated instruction or an edge case its designers did not anticipate. Its authority has been inherited from the pilot and from what it is technically able to do, without ever being deliberately decided.
Researchers at the University of Washington argued in 2025 that this gets the relationship the wrong way round. In their framework for levels of autonomy for AI agents, they treat autonomy as a deliberate design decision, separate from what an agent is technically able to do. A highly capable agent can still be deployed to operate with limited independence, and the choice should reflect the context and risk of the task.
The security community reaches a similar conclusion from a different direction. OWASP’s guidance on LLM applications identifies excessive agency as a significant risk, tracing it to agents being given more functionality, permissions or autonomy than their task requires. Gartner’s May 2026 analysis found that failures are most likely when organisations do not distinguish between an agent’s ability to act and the scope of access it is granted, and its 2027 technology trends name overprivileged agents among the operational risks of agentic AI.
The common thread is that authority deserves a decision of its own, taken with the business consequences of an error in view.
The five dimensions of authority
Authority is easier to govern when it is broken into parts. The following five dimensions cover most of what an organisation needs to decide for any agent:
1. Access:
The data and systems the agent can see. For example, an agent summarising customer correspondence needs read access to that correspondence and very little else.
2. Actions:
The tools the agent can use and the changes it can make. Reading a record, updating it and deleting it are different powers, and each should be granted separately.
3. Limits
The boundaries on scale and value. For example, a maximum refund, a cap on the number of records changed in a single run, or a restriction to particular customer segments or regions.
4. Approvals
The actions that require a human decision before they take effect, and who is authorised to give that decision.
5. Escalation and stopping
The conditions under which the agent hands a task back to a person, and the means by which it can be paused or halted quickly when something goes wrong. This includes limits on attempts, time and token spend, so an agent unable to complete a task does not loop indefinitely at growing cost.
None of these concepts are new to large organisations. Finance teams operate delegated authority matrices, IT teams apply least-privilege access, and regulated firms maintain segregation of duties. Technology leaders can extend those existing frameworks to agents, so that an agent’s authority is documented, approved and reviewed in the same way as a new member of staff. Authority also works alongside context – an agent needs both accurate knowledge of the business and clear limits on what it may do with that knowledge.
A spectrum from observing to acting alone
Governance also needs a way to describe how independently an agent operates. Gartner’s May 2026 research proposes four autonomy levels, each representing a different trust boundary with its own governance requirements.
Level 1: Observe:
They have read-only access to defined data sources, and their outputs are visible only to the person who asked.
Level 2: Advise:
Agents generating recommendations, drafts or proposed actions that people review and carry out themselves.
Level 3: Act with approval
Agents executing actions such as writing data or sending communications only after explicit human sign-off for each one.
Level 4: Act autonomously
Agents act autonomously within defined guardrails, with people reviewing exceptions, audit logs and aggregated outcomes.
The research warns in particular against uniformity. Enterprises that treat agent governance as binary, with agents either locked down or fully trusted, tend to encounter two failure modes:
- Over-restriction of simple agents, which slows delivery and encourages shadow development;
- Under-restriction of more autonomous agents, which increases operational, security and compliance risk.
A proportional approach matches controls to the level at which each agent operates.
To further refine, a single agent rarely needs one level for everything it does. For example, a customer service agent might act autonomously when updating a delivery address, act with approval when issuing a goodwill credit, and only advise when a complaint involves a vulnerable customer. Setting the level for each action allows organisations to automate routine work fully while keeping people close to the decisions that matter most.
Consequence and reversibility
Deciding where each action should sit is a risk decision. The impact of an agent getting an action wrong, and the likelihood that it will, determine how much independence that action can safely be given. How easily an error can be put right forms part of its impact, since a mistake corrected in seconds carries far less risk than one that cannot be recalled.
Reversibility should be considered because it determines how much an error costs once it has happened. For example, a misrouted ticket can be corrected in seconds, whereas a payment to the wrong account, an email sent to thousands of customers or a submission to a regulator may be impossible to recall, and an agent operating at machine speed can repeat such an error many times before anyone notices. All of these hold different risk profiles.
It is also important to consider volumes, as an action that is trivial and reversible on its own can become significant when an agent performs it thousands of times in an hour, which is why limits on scale belong alongside limits on value.
Consequence also has a regulatory dimension. Under UK GDPR, significant decisions about individuals made solely by automated means carry additional obligations, including safeguards that allow people to contest a decision and obtain human intervention. These rules were revised by the Data (Use and Access) Act 2025, and the ICO has been updating its guidance in response. For agents involved in decisions about credit, employment, benefits or access to services, the authority question is also a compliance question.
Keeping human oversight meaningful
Approval is among the most widely used controls on agents, and also one of the most fragile. Gartner cautions that approval protects an organisation only while reviewers exercise real judgement, and that time pressure and approval fatigue can wear it down into a false sense of safety. A reviewer facing hundreds of requests a day, each presented with little context, will approve most of them by habit.
The Gartner research states that people can also fall to automation bias, where they tend to over-trust outputs from automated systems, observing that even advisory agents can anchor judgement, creating risk when inaccurate outputs are accepted without scrutiny. The ICO’s Tech Futures report on agentic AI makes a related point about scale, noting that as oversight is reduced, whether intentionally or gradually, unintended consequences may persist for longer before they are noticed and corrected.
Meaningful oversight has to be designed, with approval points working best when they are reserved for a small number of decisions that warrant them, when reviewers see enough context to judge each request, and when approval rates and response times are monitored for signs of rubber-stamping.
For agents operating autonomously, oversight shifts to ongoing, proactive monitoring. Live evaluations assess the quality of an agent’s outputs and decisions as it works, and alerts flag any drift in accuracy, behaviour or cost, giving teams early visibility of emerging issues. Exception queues, audit logs and outcome monitoring add to that picture, and thresholds can halt the agent automatically when its behaviour moves outside agreed bounds.
Authority that is earned over time
Each expansion should also be justified by need, with the value of the extra autonomy weighed against the additional risk it brings. A new agent can begin at an advisory level, with its recommendations compared against the decisions humans make. Running an agent in this kind of shadow mode, where it proposes actions that are logged and assessed, without being executed, produces evidence about its reliability before it is trusted with anything that matters.
As accuracy, error rates and incident history build a record, specific actions can move to approval, and later to autonomous operation within limits. When conditions change, through a new model version, a change to the underlying process or a rise in errors, authority can be reduced again. This mirrors how organisations extend responsibility to people, and it gives risk functions and boards a clear basis for each expansion.
The ICO’s Tech Futures report states that giving AI systems agency leaves responsibility for data processing with people and the organisation. Every agent needs a named business owner accountable for its outcomes, and every change to its authority should be recorded and approved in the same way as any other change to delegated authority.
Management system standards such as ISO/IEC 42001 provide a structure for documenting these responsibilities, alongside the governance foundations covered in our earlier articles, Why AI Governance is Key to Scaling AI and The Building Blocks of AI Governance. The ICO’s call for evidence on agentic AI, opened in October 2026, signals that regulators will take a growing interest in how organisations control the agents they deploy.
Delegation by design
Organisations have long experience of deciding how much authority to give the people who act on their behalf. They set limits, require approval for significant decisions, separate duties and widen responsibility as trust is earned. AI agents call for the same discipline, applied deliberately and at a specific detail, because they act faster, at greater volume and without a person’s instinct for when something looks wrong.
Technology leaders who treat agent authority as a design decision will find it easier to scale agents with confidence, to answer the questions boards and regulators are beginning to ask, and to avoid the costly rollbacks that follow when authority is granted by default.
Key takeaways
- Set an agent’s authority as a deliberate decision, separate from what the agent is capable of doing.
- Break authority into access, actions, limits, approvals, and escalation and stopping, and document each one.
- Match the level of autonomy to each action, guided by its consequence and how easily it can be reversed.
- Design approval points so they remain meaningful at volume, and monitor them for signs of fatigue.
- Widen authority on evidence, reduce it when conditions change, and keep a named owner accountable throughout.




