Approvals and Confirmations
An AI agent should never change your records on its own. Achilles splits every change into two steps: the agent proposes it, and a person approves it. Nothing happens until the person says yes, and the platform, not the AI, holds the approval. This page explains what customers and staff see, and how administrators decide which actions need approval.
What the person sees
Suppose a broker asks the agent to create a health insurance lead for a customer.
- The agent works out which ability to use and what details to fill in.
- Instead of acting, it shows a confirmation card listing the exact details, for example the customer name, the product and the contact number.
- The person chooses Confirm or Cancel.
- On Confirm, the platform carries out exactly what was shown and the agent reports the result. On Cancel, nothing happens.
The card appears in the website widget and in the console Playground. In voice conversations the same approval is passed through the session.
Why the approval is trustworthy
The approval is not just a "yes" in the chat. When the agent proposes a change, the platform creates a one-time approval that is tied to four things.
| Tied to | Why it matters |
|---|---|
| The exact action | An approval for one action cannot be reused for another |
| The exact details | The details cannot be quietly changed after the person has seen them |
| The person who was asked | Someone else cannot approve on their behalf |
| A short time window (15 minutes) | An old approval cannot be used later |
An approval works once. Once it is used or cancelled, it is gone, so it cannot be replayed. If the platform does not recognise an approval, it treats it as expired and refuses.
Choosing which actions need approval
You have two levels of control.
| Setting | Effect |
|---|---|
| Require confirmation for all changes | Every ability that changes data asks for approval. This is the recommended default for regulated work. |
| Require confirmation for a specific ability | That ability asks for approval even if it only reads data, for example a lookup of sensitive medical information. |
Abilities that only read data do not need approval unless you mark them. The platform prevents a change ability from opting out when the all-changes setting is on, so an agent cannot quietly weaken your policy.
How to configure it
- Open Agents in the Build section of the console and choose your agent.
- Open its settings for safety rules and find the confirmation setting.
- Turn on confirmation for all changes, and list any additional abilities that should always ask.
- Check the result on the Health page under Operate, which flags configuration problems before you deploy.
- Try it in the Playground: ask the agent to make a change and check that a card appears.
If you built the agent with the Agent Builder, confirmation for changes is normally proposed for you. You can also ask your Achilles contact or IT team to review the setting. Technical detail is in the definition reference.
The card shows the fields your ability collects. Use clear ability descriptions and field names, so staff and customers can tell at a glance what they are approving.
What is recorded
Two events are kept for every approval: when the approval was requested, and how it ended, whether confirmed, cancelled, expired, or refused because it came from the wrong person. Reviewers can find both in the conversation record. See Audit trail.
Limits to know about
- Approval is by the person in the conversation. A second person, such as a manager co-signing a broker's change, is not part of the flow today. For high-stakes cases, use Human handoff instead.
- The card shows the details of the change, not a simulation of its full effect.
- Approval covers changes the agent makes. It does not review what the agent says.
Next, see how sensitive fields are protected in Personal data.