Before an AI support agent discusses or changes an account, it must confirm who it is talking to. An identity verification script sets exactly how: what it says, which factors it asks for at each access level, how many attempts are allowed, what it says on failure, and what it records.
What is the script for, and when is it used?
Support agents handle requests that range from "what are your opening hours?" to "change my phone number". Each needs a different level of certainty about who is asking. The script ties access levels to verification factors, so the agent never discusses more than the check supports. Whether an AI agent can verify a caller's identity covers the principle; this template is the working document.
Use it at the start of any conversation that will touch account information, and again before any change that raises the access level. A customer verified to discuss a balance is not yet verified to change their address.
How do you fill in the script?
Set three or four access levels, then for each level write the steps.
- Access level. What the customer can do once verified at this level.
- Factors required. Something they know, something they have (such as a one-time code to a registered device), or a check through a verification service. Higher levels need more, or stronger, factors. Connecting to such a service is covered in how to connect an AI agent to your identity verification service.
- Exact wording. What the agent says at each step, so wording cannot drift into hints.
- Attempts. How many failures before the agent stops.
- On failure. What the agent says, without revealing which factor failed.
- Record. What is filed: factors used, result, time.
Three rules belong in every script. The agent asks; it never confirms details the caller reads out ("is my address 12 High Street?"). It never says which answer was wrong. And it does not accept information offered unprompted in place of a factor it asked for.
The identity verification script template
Part 1: access levels
| Level | What the customer can do | Factors required |
|---|---|---|
| 0 | General information only | None |
| 1 | ||
| 2 | ||
| 3 |
Part 2: script per level
| Step | Agent says | Agent checks | On pass | On fail |
|---|---|---|---|---|
| 1 | ||||
| 2 | ||||
| 3 |
Part 3: rules. Attempts allowed; lockout and route to a person; signals of social engineering that end the check; records filed.
A filled sample script: retail bank
Access levels
| Level | Can do | Factors |
|---|---|---|
| 0 | Product information, branch hours | None |
| 1 | Balance, recent transactions, card status | Two knowledge factors from the account record |
| 2 | Block card, order replacement to registered address | Level 1 plus one-time code to registered phone |
| 3 | Change phone, email or address | Not available to the agent: handed to a person |
Level 2 script
| Step | Agent says | Checks | On pass | On fail |
|---|---|---|---|---|
| 1 | "Before I can help with your account, I need to check it's you. Can you tell me your date of birth?" | Matches record | Step 2 | Count failure; step 2 |
| 2 | "Thanks. And the first line of your address?" | Matches record | Step 3 | Count failure |
| 3 | "I've sent a code to the phone number we hold for you. Please read it to me." | Code matches | Level 2 granted | "I can't verify you this way just now" |
Rules. After two failures at any step, the agent stops and offers a route to a person or the app. It ends the check if the caller says they are calling for someone else, asks the agent to send the code to a different number, or presses for urgency to skip a step. Each check is filed with factors used, result and time.
Notice level 3. Changing contact details is the classic step in account takeover, because it redirects future codes. The bank decided no agent handles it at any depth, so the script simply hands over.
How should the agent refuse?
Refusals should be calm and give no detail that helps an impostor. "I'm not able to verify you on this call, so I can't discuss the account. You can verify in the app, or I can arrange a call back to the number we hold." A library of approved refusals keeps wording consistent; see the support agent refusal reasons library template. The wider process is in how to verify a customer's identity before discussing an account with an AI agent, and the stance behind it in the AI support agent that knows when to say no.
Questions people ask
Should the agent tell the customer which answer failed?
No. Saying "your date of birth was right but the address wasn't" helps an impostor narrow the answer. Say only that verification did not succeed.
Can verification carry over between conversations?
Only within a short, defined session. A new conversation starts at level 0. Record the session length in the rules section.
How do we check the agent follows the script?
Score a sample of verifications every week on a QA scorecard; the support QA scorecard template includes a verification section. Any step skipped is a failure, even if the customer was genuine.