Colour theme
Book a consultation

Preparing for AI agents

Before an AI agent changes business records, define its permissions, approval steps, test cases and stop procedure. Start with one limited workflow.

Blue letters AI surrounded by looping wires in a computer-generated scene.

An AI assistant that drafts a reply leaves someone to review and send it. An agent connected to your business software may also update a record, schedule an appointment or send a message. Those permissions change what a mistake can affect.

Before connecting an agent, choose a specific task and write down the actions it needs. For example, preparing a proposed appointment from an enquiry needs different permissions from changing a customer’s account. Check whether a form or a rules-based automation can already do the job.

Define the work and the decisions

Describe the inputs, expected result and exceptions. Who checks an unclear request? What happens if a customer appears twice, a required field is missing or a connected service is unavailable?

Separate reading information, drafting a change and committing it. For an initial appointment pilot, the agent could propose a time and a reply while a staff member confirms both. Keep payment approvals, changes to bank details and deletion of business records outside that pilot.

Appoint a business owner who can decide whether the result is useful. Include the people who handle exceptions today; their examples will help you build realistic tests.

Limit access in the connected systems

The Canadian Cyber Centre’s agentic-AI guidance recommends limited privileges, human oversight and controls around tool use. An instruction telling an agent to be careful doesn’t restrict an API key. Configure the connected systems to allow only the required actions and data.

Use a dedicated identity where supported, keep credentials out of prompts, and establish a way to revoke access. Test that the agent is refused when it tries an action you haven’t allowed. Require approval through a control the agent can’t bypass or approve for itself.

Review the provider’s data handling, retention and connector terms before sending customer or employee information. The existing shadow-AI review helps identify tools and inputs already in use.

Test misleading inputs and failed actions

Agents can encounter prompt injection: instructions embedded in content such as an email or document that try to redirect their behaviour. Include a test message asking the agent to ignore its task, send information to an outside address or use a tool it shouldn’t have.

Also test ordinary failures. A timeout after a record was created can lead to a duplicate if the agent retries. Ask the implementer how it detects completed actions and handles repeated requests. Check the result in the business application, including any unwanted changes.

Use synthetic examples in a test environment first. Keep a list of expected outcomes so a model, connector or configuration update can be checked against the same cases.

Measure the pilot and practise stopping it

Record completed tasks, corrections, missed exceptions, costs and staff review time. Compare them with the existing process. A fast draft that requires extensive checking may offer little improvement.

Keep action records that let the person checking the agent trace what changed, without collecting unnecessary personal information. Test the stop procedure: revoke the connection, cancel queued work where supported and confirm what still needs a person to finish or reverse.

Expand the pilot only when the owner accepts the results and the controls work. Talk to us if you need help choosing a workflow and planning its integrations and approval steps.


Featured Image Credit