Before You Start: Are You Using an Agent or Building One?
You can use an AI agent through an existing app, customise one using a no-code platform, or build one using code. The amount of control you have over its setup will depend on the service.
If you’re using an existing agent, you’ll usually work with the permissions and settings the provider offers. If you’re building one, you may be responsible for choosing the tools it can use, connecting accounts and deciding when it needs approval.
The goal is the same in both cases: give the agent the access it needs for its job, understand what it can do with that access and keep a way to stop it.
1. Be Clear About the Job You Want It to Do
Before connecting an account or adding a tool, write down the task you want the agent to complete. Include what it should produce and which decisions you want to make yourself.
For example, if you want help with travel planning, you could ask it to compare flights and create a shortlist. You can then decide whether to book. If you want help with email, it could prepare drafts for you to review rather than send replies immediately.
A clear instruction might look like this:
‘Find three hotels in York for 12 to 14 June, under £150 per night. Show me the total price and cancellation terms. Don’t make a booking.’
Why it matters: A broad request such as ‘organise my trip’ leaves the agent to interpret how far it should go. Setting out the task and its limits helps you decide what access is needed and gives you something to check its work against.
If you’re building an agent, use those limits when choosing its tools and permissions. An instruction alone doesn’t prevent an agent from using a tool it has been given.
2. Choose Your Platform and Connections Carefully
Check who provides the agent, the platform you’re building it on and any extensions or integrations you plan to connect. Look for clear information about the company, its security and privacy practices, and how to get help if something goes wrong.
If you’re building an agent, you may also come across third-party connectors that let it interact with other services. These can be useful, but each connection adds another service or component to consider. Check who maintains it, what it can access and whether you need it for the task.
Download apps and extensions from the provider’s official site or a trusted marketplace, and check the developer’s name. If an invitation arrives unexpectedly, open the provider’s website yourself instead of using the link in the message. Our guide explains how to check if an email is fake or a scam.
Why it matters: The agent is only one part of the setup. A connected app or fraudulent extension could also have access to the information you share or the tasks you allow it to perform.
3. Check What Each Permission Allows
When an agent asks to connect an account, read what it can do after you select ‘Allow’. A connection can cover more than simply viewing information.
For example:
- An email connection might allow the agent to read, draft, send or delete messages.
- A calendar connection might allow it to see, create, change or cancel events.
- A file connection might allow it to read or edit documents.
- A shopping connection might give it access to purchase history or saved payment methods.
- A financial connection might allow it to view account information or initiate transactions.
Choose the narrowest access available. If you want an agent to prepare an email for review, check whether it can work with draft access rather than permission to send.
If you’re building an agent, check the permissions of each tool and connection separately. A read-only research tool and a tool that sends emails perform different jobs and should have access that reflects that.
Why it matters: Permissions set the boundaries of what an agent can do if it makes a mistake or follows a malicious instruction. The fewer unnecessary actions available, the less it can affect.
4. Limit the Information It Can Reach
Think about the information behind each connection, not just the account name. An inbox can contain bills, password-reset messages and private conversations. A cloud folder might hold identity documents or files shared by other people.
Look for ways to share only what the task requires. Depending on the service, you might be able to select one folder, upload a single document or use a separate calendar instead of connecting an entire account.
If you’re building an agent, keep sensitive files and accounts outside its reach unless they’re necessary. A tool that can search every folder on your computer has a much wider view than one restricted to a folder you’ve chosen for the task.
Why it matters: Information the agent can access may be processed, stored or shared as part of its work. Limiting access reduces the amount of personal information involved if something goes wrong.
Before sharing files or messages, check the provider’s privacy settings and deletion options. Disconnecting an account later does not necessarily delete information the service has already collected.
5. Protect Passwords, API Keys and Other Credentials
Agents and the tools connected to them need a way to access services. Depending on the setup, that might involve an account connection, an API key or another credential.
If you’re using a ready-made agent, check where the connection takes you and what it requests. Be cautious if a tool asks you to enter a password or security code directly when you expected to connect through the account provider.
If you’re building an agent:
- Use the platform’s approved method for storing credentials, such as a secrets manager.
- Don’t paste passwords or API keys into instructions, public documents or code that others can access.
- Give each credential only the access required for its task, where the service allows it.
- Remove or revoke credentials you no longer need.
An API key is a credential that lets software connect to a service. Treat it like a password: someone who obtains it may be able to use the access attached to it.
Why it matters: Credentials can give an agent access beyond the words you type into a prompt. If a key or password is exposed, an attacker may be able to use that access too. The UK’s National Cyber Security Centre recommends considering the reach of the credentials available to an agent.
6. Require Approval Before Important Actions
Check when the agent asks for confirmation and what it can do without asking. Settings might appear under names such as ‘Require approval’, ‘Automatic actions’ or ‘Confirm before sending’.
Keep an approval step for actions that affect your money, accounts or other people, such as:
- Sending an email or message.
- Making a purchase, booking or payment.
- Deleting files or changing account settings.
- Sharing personal or confidential information.
- Publishing content.
If you’re building an agent, make sure the approval happens before the connected tool carries out the action. Asking the agent in its instructions to ‘check with me first’ is useful, but the platform or workflow should enforce the approval step where possible.
Why it matters: A confirmation screen gives you an opportunity to check the recipient, amount and other details before the action happens. Review those details each time; simply approving a request without reading it won’t catch a mistake.
Security organisation OWASP recommends human approval for consequential agent actions, alongside controls that check what the agent is allowed to do.
7. Set Spending Limits and Check Payment Access
If an agent can shop or book for you, find out whether it has access to a saved payment method. Check whether the service offers spending limits and approval for each purchase.
Be specific about your budget. A price shown during a search may not include delivery, taxes or other charges. Ask to see the full cost and seller before approving payment.
For example:
‘Find a pair of walking shoes for under £100, including delivery. Show me the retailer and returns policy before buying anything.’
If you’re building a shopping agent, set spending limits in the payment service or workflow where available. A limit written in the prompt is an instruction to the agent; a limit enforced by the payment service is a separate control.
Why it matters: A spending limit can restrict the financial impact of an unwanted purchase, while an approval step lets you check whether the product, price and seller are right.
8. Test the Agent Before Using It for Important Tasks
Start with a task where you can easily check the result and correct a mistake. Ask the agent to compare public information, prepare a draft or organise sample data.
If you’re building your own agent, test it before connecting sensitive accounts or letting it act on real information. Try situations where your instructions are incomplete, the information it reads is misleading, or a requested action should require approval.
For example, if an agent is supposed to draft emails without sending them, check that it stops at the draft stage. Then check that it still stops when a message it reads contains instructions telling it to send a reply.
Why it matters: Testing helps you find problems in the setup before they affect your real accounts, files or payments. It also shows whether approval controls work in practice.
A successful test is a useful start, but continue checking the agent’s work as you add new tools or permissions.
9. Check How You Can Monitor and Stop It
Find out whether the service keeps an activity history and what that history shows. You should be able to review the agent’s completed actions and, where available, the tools it used.
Also look for a way to pause or stop it. If you’re using a ready-made service, this might be a stop button, an automation setting or an option to disconnect an account.
If you’re building an agent, plan how you’ll disable the workflow and revoke its connections. Keep records that help you understand what it did, but avoid storing passwords, API keys or unnecessary personal information in those records.
After the agent completes a task, check the result against the original source. Read the retailer’s order confirmation, check the calendar event or open the document it changed.
Why it matters: An activity history helps you spot unexpected actions and understand what happened. Knowing how to stop the agent gives you a way to limit further activity while you investigate.
10. Review Connections and Remove Access You No Longer Need
An agent you used for one task may still have access after that task is finished. Review its connected accounts and remove any permissions you no longer need.
In the connected account, look under Security, Connected apps, Third-party access or a similar section. Select the service and choose the option to remove or revoke its access.
If you built the agent, also check for unused tools, credentials and automations. Remove them from the setup and revoke access through the connected service where necessary.
Why it matters: Every unused connection is another route to your information or account. Removing access reduces the number of services that can reach it.
If you shared a password directly with a service and no longer want it to have access, change that password and review active sessions. Check the provider’s deletion process separately if you want it to remove information it has already collected.
What Should You Do if an AI Agent Takes an Action You Didn’t Approve?
Stop the agent or pause the workflow, then remove its access to the affected account. Check the activity history, account notifications and confirmation messages to establish what happened.
Contact the relevant company as soon as possible. A retailer may be able to cancel an unwanted order; a travel provider can explain your options for a booking. If you see an unfamiliar payment, contact your bank using its official app or phone number.
If you believe a password or API key has been exposed, change or revoke it and review the access it provided. Keep relevant confirmations and messages in case you need them when reporting the issue.
How Often Should You Check an AI Agent’s Permissions?
Review permissions whenever you connect a new service, add a tool or change the tasks the agent performs. Check existing connections periodically and remove those you no longer use.
A change that seems small can give an agent a new level of access. For example, an email assistant that moves from preparing drafts to sending replies needs a fresh decision about when it should act and when it should ask.
Should You Let an AI Agent Work Without Approval?
That depends on the task and what the agent can access. You may be comfortable letting it compare public information or organise a set of test files independently.
For actions involving money, private information or other people, keep a point where you can review what it plans to do. If you’re building the agent, put that approval into the workflow. If you’re using an existing service and it doesn’t offer the control you need, complete that step yourself.

