Part 1, Understanding AI agents
Part 3, Introducing an agent
Principle 3
Authority
Scope and limitations of this guide
This guide provides general information on organising work with AI agents, not advice tailored to a particular organisation, system or decision. It is not legal or regulatory advice.
Its examples, checklists and templates do not establish that a proposed use is safe, lawful or suitable. You need to assess and test the proposed use in your own circumstances. Following the guide does not by itself meet the requirements of any law or standard.
The guide is provided in good faith, using information available at the time of publication or revision. AI systems, provider terms and legal requirements can change. Before acting, check the current requirements that apply to your organisation and country, confirm what the system can do and test the proposed use. You are responsible for how you apply the guide and should obtain professional advice where appropriate.
Reading, downloading or using the guide does not by itself create an adviser–client relationship with Tom Clowry or Considered Ideas Ltd.
To the fullest extent permitted by law, Tom Clowry and Considered Ideas Ltd accept no liability for loss or damage arising from use of, or reliance on, this guide. Nothing in this notice excludes or limits liability for death or personal injury caused by negligence, fraud or fraudulent misrepresentation, or any other liability that cannot lawfully be excluded or limited.
What can the agent do without further approval, and what needs a person’s decision?
Agree what the agent can do without further approval. State when it needs a person’s decision and what it should do while it waits.
- What to decide
- What the agent can do, what needs approval and how the work can be stopped.
- Key point
- Specify which actions the agent can take without further approval and which need approval first.
Why it matters#
Proposing an action and implementing it have different consequences#
A proposal can still be revised. An action that has been implemented can commit the organisation, affect other people and be difficult to reverse. For example, a reply can be revised until it is sent.
Background#
From economics
The principal–agent problem#
Economists use the term ‘agent’ for anyone who acts on behalf of another, known as the principal. The principal–agent problem arises because the agent’s interests can differ from the principal’s, and the principal cannot fully monitor its work.
Organisations have long managed this with clear instructions, limited authority, records and checks. An AI agent does not have interests as an employee does, but you still cannot see all of its work, and its objective can differ from yours if its instructions are incomplete. Principle 3 applies the same remedies.
Source Jensen, M. C. and Meckling, W. H. (1976), Theory of the firm: managerial behavior, agency costs and ownership structure.
Actions and decisions#
Reading, changing and acting#
The risk depends on the information the agent can access and the actions it is allowed to take. Reading can expose sensitive information to the AI service. Changing records can introduce errors. Sending, publishing or paying can commit you to something that is difficult to reverse.
Give the agent the lowest level that the task needs. Permission to prepare work does not include permission to act on it.
Agents that buy and pay#
Shared standards are also emerging that allow agents to find products, place orders and make payments. Some support purchases authorised in advance, in which a record states what the person has authorised and the limits within which the agent can act.
Agents that make payments are at an early stage. The standards, payment systems and regulation that support them are still being developed and tested, and there is little practical experience to draw on. An agent that can make payments can also make costly mistakes, such as buying the wrong item, paying the wrong supplier or acting on instructions hidden in a web page. At this stage, an agent is better used to prepare purchases for a person to approve than to make them itself. Any use of an agent to buy or pay on behalf of an organisation needs to be approved through its usual process for spending. The people responsible for finance and for the systems involved need to agree to it. The approval should set out what the agent can buy and the most it can spend. Payment controls should enforce those limits, and any purchase outside them needs further approval.
Decisions about people#
This section is not legal advice. Its examples are drawn from UK law at the time of writing. The law can change, and the rules may differ in the EU and elsewhere. Ask the person who advises your organisation which rules apply to you.
Some decisions have legal or similarly significant effects on a person, such as refusing credit, rejecting a job application or deciding whether someone receives a benefit. Do not let an agent make such a decision without meaningful human involvement.
The law restricts these decisions. In the UK, data protection law requires safeguards where they are made without meaningful human involvement, and the rules are stricter where the decision uses special category information, such as information about a person’s health. In the EU, such decisions are allowed only in limited circumstances. Only the person responsible for data protection can confirm that an exception is lawful and that the safeguards are in place.
Draft guidance published by the ICO says that human intervention cannot be tokenistic. The person reviewing the decision should have the discretion and authority to change it.
Source ICO, Automated decision-making, including profiling, draft guidance, 31 March 2026.
In practice#
5 steps
01Examine the actions the task involvesPermission to prepare a recommendation need not include permission to act on it.
Before giving an agent a task, examine the actions it would involve. Permission to prepare a recommendation need not include permission to act on it. Some actions require further approval because of their consequences for other people.
Before granting permission:
Who could be affected by a mistake?
Include people who may be overlooked by the available records.
How serious could its effects be?
Could they be put right?
Set the controls
The actions that need prior approval, and those that require closer checks, should be agreed by those responsible for the work and for the agent.
02Set out the permissions in the briefTell the agent the result required, what it can use and what it can change.
The Brief The instructions given to the agent, including the purpose, permitted information and actions, checks and arrangements for seeking advice. should tell the agent:
The result required
Include an approved example of the required standard where you can.
Which information and tools it can use
What it can change
When to ask for help
What records and evidence to provide
03Match the agent’s access to the workBeing able to use a tool does not mean having permission to use it for every purpose.
Being able to use a tool does not mean having permission to use it for every purpose. Give the agent access that matches the work that has been authorised.
InstructionsExplain the expected behaviour.
An agent does not always follow an instruction as intended.
Access controlsLimit the actions available to the agent.
Where an error could have serious consequences or be difficult to reverse, do not rely on instructions alone. Enforce limits through access controls or approval requirements.
Before use, test that approval requirements, access restrictions and stopping arrangements work as intended.
04Permission can be given in advanceA person need not approve every small step.
A person need not approve every small step. Where testing shows that the agent meets the agreed requirements, it can be given permission to repeat the task under stated conditions. Agree in advance:
The conditions
Be specific about when the agent can act.
How many actions or attempts it can make
A mistake can be repeated many times before anyone notices.
How the result will be checked
Who can stop the work
Which actions can be stopped or cancelled
How completed actions will be checked and errors corrected
Review permissions in light of experience#
Experience can justify extending, retaining or reducing permission.
05Decide whether one agent can pass work to anotherState in the brief whether this is allowed.
An agent can be configured to pass part of a task to another agent. State in the brief whether this is allowed.
The record should show which agent carried out each action.
See also · 1.2 Tools and actions · 1.3 Permissions and controls · 1.2 How connections are made
Examples#
The opening hours example
What the agent can and cannot do#
A local service is open from 9 am to 5 pm on weekdays. Some service users have asked for evening appointments, and the team must decide whether longer hours would improve access. An agent examines twelve months of request records and compares the costs of the options. See chapter 1.5.
The brief given to the agent at the start sets these limits.
Returns the comparison, its sources, the assumptions that matter and any unanswered questions to the service lead.
Decides after the team has discussed the comparison.
Illustrative example
Approval to draft and approval to send#
An organisation allows an agent to prepare communications but reserves publication for approval by the person responsible. Its rules distinguish permission to prepare a draft, approval to send it once and permission to repeat the action.
A request for a draft permits the agent to prepare it, but not to send or publish it.
A one-off approval does not create standing permission.
Nothing is sent, published, uploaded or scheduled without the explicit approval of the person responsible.
In other work
Risks#
Material that the agent reads can contain instructions#
An agent can encounter instructions embedded in emails, documents or web pages that conflict with the instructions it should follow.
Attempts to redirect an AI system in this way are a form of Prompt injection An attempt to redirect an AI system through instructions that override or conflict with those it should follow, including instructions embedded in material it is meant to read or process.. The material can come from inside or outside the organisation. The UK National Cyber Security Centre warns that prompt injection may never be fully mitigated. Limit what the agent can access and do if it follows such instructions.
Sources National Cyber Security Centre, Prompt injection is not SQL injection (it may be worse), 8 December 2025. OWASP, GenAI LLM Top 10 2026, LLM01: Prompt Injection.
Notes an agent keeps between tasks
An agent can keep notes, lessons or preferences from earlier work and use them in later tasks. It can draw on these as well as its instructions.
Exercise#
Check when the agent needs approval#
Choose an action the agent is allowed to take and suppose that a condition for taking it has not been met. State what work can continue, what must stop and who can decide how to proceed.
Check the instructions against a case involving incomplete information or conflicting requests. An approaching deadline should not give the agent permission to exceed its authority.
Describe what the agent can do without asking, what needs approval and what it must do when a condition is not met. Explain who could be affected by an error and whether its effects could be put right. Include how someone can stop the work and check what has already happened.
Further reading · Autonomy expands accountability · clowryai.com/thinking
© 2026 Considered Ideas Ltd. Licensed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0). This guide provides general information and is not legal or regulatory advice. The guide’s home page gives the suggested credit.
Scope and limitations of this guide
This guide provides general information on organising work with AI agents, not advice tailored to a particular organisation, system or decision. It is not legal or regulatory advice.
Its examples, checklists and templates do not establish that a proposed use is safe, lawful or suitable. You need to assess and test the proposed use in your own circumstances. Following the guide does not by itself meet the requirements of any law or standard.
The guide is provided in good faith, using information available at the time of publication or revision. AI systems, provider terms and legal requirements can change. Before acting, check the current requirements that apply to your organisation and country, confirm what the system can do and test the proposed use. You are responsible for how you apply the guide and should obtain professional advice where appropriate.
Reading, downloading or using the guide does not by itself create an adviser–client relationship with Tom Clowry or Considered Ideas Ltd.
To the fullest extent permitted by law, Tom Clowry and Considered Ideas Ltd accept no liability for loss or damage arising from use of, or reliance on, this guide. Nothing in this notice excludes or limits liability for death or personal injury caused by negligence, fraud or fraudulent misrepresentation, or any other liability that cannot lawfully be excluded or limited.

