As AI agents take on more responsibility, a critical question emerges: How do we ensure they’re acting with the right authority? The answer starts with establishing clear boundaries around what an agent can do on someone’s behalf. Here, I’ll cover how identity verification, human approval, and signed consent can help protect people and their information.
If I ask an AI agent to book a trip within a budget, I expect it to automatically handle the details. However, if it needs to spend more or use a different account, I expect it to ask me first.
That leaves the business receiving the booking request with a specific problem. It must check what I authorized and decide whether this request is covered. Simply having a valid AI agent credential doesn’t answer that question on its own.
This challenge becomes particularly important in agentic commerce, where AI agents can manage the buying journey, from finding a product or service to placing an order and arranging payment.
In financial services, AI agents might also help people review accounts, initiate transfers, or dispute charges. A bank may receive requests from a customer’s agent or use its own agents to carry out tasks.
The two can meet in the same transaction. The agent booking my trip may also ask my bank to approve a payment. Permission to buy the ticket should not give it general access to my accounts. The merchant and the bank each need to check the authority that applies to their part. The same principle applies to insurance changes and access to company information.
Set clear limits on AI agent permissions
Telling an AI agent to pay this month’s bills shouldn’t give it permission to add new recipients or access any account it chooses. That’s where clear AI agent authorization becomes essential. Permissions should define which accounts the agent can access, how much it can spend, and when its authority expires. The person or business granting authority also needs a way to withdraw it.
Routine activity within those limits can follow the organization’s normal policies. A new supplier, an unusually large purchase or a change in account access may require someone to approve it. Device or behavioral signals might also raise concerns, even when the amount is small.
If every step comes back for approval, the agent isn’t saving much work. I’d focus on identifying the points where someone needs to make a fresh decision, while keeping the ordinary checks in place.
Verify who approved an AI agent’s action
That approval needs to be linked to the right person. Liveness can help assess whether a real person is present. Biometric matching can help establish whether that person matches an enrolled identity. However, neither tells us that the person approved a particular purchase or account change.
At Mitek, we help establish identity during onboarding and authenticate people when they return. Our MiPass offering supports biometric matching and liveness checks. Digital Fraud Defender adds protection against deepfake, injection, and document template attacks. These capabilities provide identity evidence for the organization’s decision.
The organization still must show the person the action being proposed and record their response. If an employee is approving on behalf of a company, it also needs to check that employee’s authority.
Protect personal data in agentic workflows
I’m equally concerned about the information we hand to the agent. It shouldn’t need to carry someone’s selfie, biometric template, or reusable authentication secret from service to service to prove an action was approved.
Privacy-preserving verification should let us confirm the person without handing their biometric data to the agent. The agent should receive only the authorization evidence needed for the task. We also need clear rules on who can access the underlying verification data and how long it is kept.
How signed consent could work for AI agents
Take a business that lets an agent order supplies from approved vendors within a budget. An illustrative integration could work like this:
- Agent initiates. The agent proposes a larger order from a new supplier, outside its existing authority.
- Approval is requested. The business holds the order and asks an authorized person to review the supplier, items, and cost through an authenticated channel.
- Human approves. The person authenticates using Mitek capabilities and explicitly approves or rejects the order.
- Signed consent. The business or its authorization provider creates a digitally signed record linking the approval to the specific agent, order, intended service, and expiry.
- The request is checked. Before ordering, the business validates the consent, authenticates the agent, and checks current permissions and risk. It holds or rejects changed requests and authorizations that are expired, revoked, or already used.
The business keeps an audit trail of the approval and execution decision. It must also enforce revocation so the agent can’t keep using a withdrawn permission. Revocation won’t undo an action that has already been completed.
A bank could apply the same approach to a transfer above an agreed limit or a change in account access. The approval would need to cover that exact action and come from someone authorized to make it.
Before expanding an agent’s access, I’d want confidence that the organization can trace every action back to the permission granted and stop requests that fall outside those boundaries. That’s the practical test I would apply to human trust. The more autonomy we give AI agents, the more important it becomes to establish exactly where that authority begins and ends.