Give agents access. Keep your business in control.
AI generated editorial imagery. Illustrative, not a customer testimonial or product screenshot.
An agent becomes useful when it can do more than explain what should happen. It may need to read a customer record, prepare a document or update a system. Giving it that access should be a deliberate business decision. The objective is to let routine work move without constant supervision, while keeping a clear boundary around spending, commitments and sensitive changes. Here is a practical way to design that boundary.
In this article
What to take away
- Give each role the access needed for its job and distinguish reading from changing information.
- Make approval requests specific enough for a person to understand the action and its effect.
- Keep evidence of completed work and a clear route for exceptions, changes and stopping the workflow.
Start with a role and access map
List the systems each agent needs and the reason for each connection. A reporting agent might read campaign results and CRM stages without changing either. An onboarding agent might update a defined set of customer fields and create delivery tasks. Those are different permission profiles, even if both agents use the same CRM.
Ask the implementation team to show how those boundaries are enforced. Where the connected platform supports separate accounts, roles or scopes, use the appropriate controls. An instruction saying not to do something is useful context, but it is not a substitute for restricting the underlying access where that restriction is available.
| Access level | What it permits | Illustrative business use |
|---|---|---|
| Read | Retrieve approved information | Prepare a report from campaign results |
| Prepare | Create a draft for review | Assemble an onboarding pack or proposed reply |
| Act within rules | Complete an authorised, bounded change | Update an approved status field |
| Request approval | Pause before a consequential action | Ask before publishing a new claim or changing spend |
Define approvals around consequences
A useful approval policy reflects the effect of the action. Consider who is affected, whether the action can be reversed and what it might cost. OpenAI's practical guide recommends human oversight for sensitive or consequential actions and intervention when an agent cannot complete a task reliably. Your implementation should translate that principle into decisions people recognise.
For example, a content agent might organise a draft calendar autonomously but request approval before publishing a new claim about your service. A finance assistant might assemble records but leave payment authorisation with the relevant person. The policy should name the approver and explain what happens if nobody responds.
Three clear operating boundaries
- Work independently
Routine actions that fit the agreed role, tools and limits.
- Ask for a decision
A defined exception or consequential change with context attached.
- Stop and escalate
Missing access, conflicting facts or an action outside the approved scope.
Make approval requests easy to judge
An approval message should explain the proposed action, the relevant evidence and what will change if you accept it. It should identify the customer, campaign or record involved and show any cost or commitment. A vague request to continue forces the approver to investigate the whole task again.
Use a clear expiry or recheck rule where facts can change. A proposed appointment time may no longer be free. A draft prepared from an older customer record may need updating. The implementation should check the relevant conditions before carrying out the approved action, rather than treating yesterday's approval as permission for a different situation.
Keep business knowledge separate from incoming requests
Customers, documents and websites can supply useful information, but they can also contain mistakes or instructions that conflict with your rules. An incoming message should not be allowed to redefine what the agent may do. Ask how the system distinguishes business instructions from material it is processing.
Maintain a small set of authoritative sources for important facts such as services, prices, policies and approved claims. Name the person responsible for updating them. If two sources disagree, define how the agent should resolve the conflict or ask for help. This makes the system easier to maintain as the business changes.
Ask for an activity history you can use
The history should let a reviewer connect a request to the action and its result. Useful records include the relevant business reference, the action attempted, the response from the connected tool and any approval. An agent saying a task is complete is not the same as the destination system confirming the change.
Anthropic's guidance on evaluating agents highlights the difference between an agent's account of its work and the actual outcome in the environment. For business review, inspect the saved record or confirmed result. Sample ordinary completed work as well as exceptions so unnoticed mistakes do not become the accepted routine.
Give exceptions an owner and a return route
An exception queue should explain what blocked the task, what has already happened and what a person needs to decide. Avoid sending every uncertainty to the founder. Route support cases, campaign questions and finance issues to the people who own those responsibilities.
After the person resolves the issue, decide whether the agent can continue safely or whether the task should remain with the team. Repeated exceptions are also useful information. They may reveal an unclear instruction, missing business data or a connection that needs improvement. Reviewing those patterns can reduce the supervision burden over time.
Test control as carefully as completion
Before launch, test the stop conditions as well as the successful path. Try an out of scope request, an expired connection and conflicting records. Check that the agent pauses when required, preserves the work already done and alerts the correct owner. Practise disabling access or pausing the workflow so the team knows what to do during an incident.
At handover, document the permissions, approval policy, knowledge owners and support route. Review the setup when a role, tool or business process changes. Good control makes autonomy more useful: the team can trust routine work to move because the boundaries and exception process are understandable.
Common questions
Does every action need my approval?
No. The aim is to agree which routine actions can run independently and which need a decision. Requiring approval for everything can leave the business with nearly as much work as before.
Can one person supervise several agents?
Possibly, depending on the volume, complexity and exception rate. Measure the actual supervision required. Do not assume that one person can manage an unlimited amount of automated work simply because the agents operate continuously.
Sources and further reading
Primary sources support the technical distinctions. Business examples, suggested workflows and buying questions are our editorial guidance.
What would this change in your business?
If you want help with the implementation, we can scope the work, build the connections and train your people at handover.
Check free audit eligibilityLearn with us
Understand AI agents.
Put your ideas into practice.
Join our AI agents and automation community for practical resources, shared learning and a place to explore what is possible for your business.