Every voice agent must answer one question before it can help: is this the right person? Ask too little and you risk giving one customer's details to someone else. Ask too much and the caller feels interrogated, gives up or hangs up. The answer is to verify in proportion to the risk, using the lightest check that protects the request in front of you.
Caller identity verification with AI works by matching the check to the risk of the request. Simple information needs little or no check, a change to an account needs a one-time code sent to the contact details on file, and sensitive actions need a stronger step. Agents should never ask for passwords, PINs or full card numbers on a call.
Why is verification hard on a phone call?
A phone call gives you very little to go on. Caller ID is not proof of identity, because numbers can be faked or shared, and a name and a date of birth can often be found by anyone who looks. Each piece of information on its own is weak, so the strength comes from combining checks in a sensible way.
The verification use case describes the calls where this matters most, and the principle is the same across sectors: a check should cost the honest caller very little, and make life hard for the dishonest one.
How much verification does each request need?
Most calls can be sorted into three tiers. The tier decides the check.
- Tier 1, general information. Opening hours, a product question or a published status. No account detail is shared, so no check is needed. If the reply starts to mention anything about the customer, move up a tier.
- Tier 2, account service. Checking an appointment, an order or a booking. Use a light check: the caller's number matches the account, and they confirm one detail on the record, such as the date of their last booking. Treat knowledge questions as a weak signal on their own.
- Tier 3, sensitive changes. Changing contact details, cancelling a service, moving money or unlocking access. Use a one-time code sent to the contact details already on file, and add a further step if your policy requires one.
Some sectors set stronger rules for certain services. Banks and lenders are the clearest example, as covered in AI voice agents for banking. Those rules apply on top of this approach, not instead of it.
Which caller identity verification methods work best?
No single method is strong enough for everything. The table compares the usual options.
| Method | Strength | Good for | Weakness |
|---|---|---|---|
| Caller ID match | Weak alone | A first filter for tiers 1 and 2 | Numbers can be spoofed or shared |
| Knowledge questions | Weak alone | Light checks alongside another method | Details leak, and callers may guess |
| One-time code by SMS or app | Strong | Tier 2 and tier 3 changes | A shared phone can pass the code on |
| Callback to the number on file | Strong | Changes that can wait a few minutes | Slower, and it needs a number on record |
| Signed-in channel | Strong | Sensitive requests from a logged-in customer | The customer has to be signed in |
| Human check | Flexible | Anything the rules leave to judgement | Costs staff time |
How does the one-time code flow work?
The one-time code is the workhorse for sensitive changes. It works like this:
- The agent says it will send a code to the number or email already on the account, and confirms which one, without reading the whole address out.
- The code is sent by message, and the caller reads it back to the agent.
- The agent checks the code against the record. It accepts only the current code, and only for a short time after it was sent.
- If the code is right, the agent carries out the change or moves the request to the next step.
- After a set number of failed attempts, the agent stops, locks further attempts for a period and offers a secure route or a person.
The agent must never speak the code itself. A code that is read out on a call can be heard by anyone nearby, and it defeats the point of sending it.
How do you handle a failed verification?
A failed check should be handled calmly and without giving anything away:
- Do not say which part was wrong. Say that the details could not be confirmed.
- Do not read out any account detail to be helpful, even to a caller who sounds genuine.
- Offer a secure route, such as signing in through the app or website, or a callback to the number on file.
- Log the attempt and the outcome, but not the answer the caller gave.
- Pass the call to a person if the caller seems distressed or says they need help.
What should you never ask on a call?
Some information should never be spoken on a voice call, whatever the risk:
- passwords, PINs or any code that secures the account
- full card numbers or card security codes
- full national identity or passport numbers, unless a specific secure process needs them
- anything the caller would have to read out that is better handled in a secure channel
If the agent needs one of these, it should ask the caller to use a secure channel and explain where to go.
How do privacy and the audit trail fit in?
Record the outcome of a check, such as passed, failed or locked, rather than the secret the caller gave. Redact any sensitive value that appears in a transcript, and keep an audit trail that shows who was verified, when, and what the agent was allowed to change. Keep recordings and transcripts only as long as your policy needs. The AI voice agent security checklist lists the questions to ask a vendor on these points.
How VoiFlow handles this
VoiFlow enforces permissions and limits outside the model, so a rule such as a cap on failed attempts holds however the conversation goes. When a check fails and a person is needed, staff receive the full summary. Redaction and the audit trail are part of the platform, and every call is recorded and scored, so verification problems show up in review. The trust centre explains how these controls are documented.
Frequently asked questions
Is caller ID enough to verify someone?
No. Caller ID can be faked or shared, so use it only as a first filter. For anything that changes an account, add a one-time code or a callback to the number on file.
What if the caller does not have their phone?
Offer a secure route that does not need that phone, such as signing in on the website or app, or a callback to a contact detail already on the account. Do not lower the bar just to finish the call.
Should an AI agent ever read out a one-time code?
No. The caller reads the code back to the agent, and the agent never speaks it aloud. If the code has to be shared another way, the process should be redesigned rather than the rule relaxed.
How do you test verification before going live?
Test each method with colleagues, including failed attempts, lockouts and a shared phone, before the line opens. You can hear how the checks sound on a live demo agent.
