When AI gives a caller the wrong price

The short answer

A wrong answer on an AI call is rarely a mystery. It comes through one of five doors: Stale source, Missing source, Misheard question, Lookup error or Overreach. Four of them start outside the model: Stale source and Missing source in the source, Misheard question in speech recognition, Lookup error in a system lookup. Only Overreach starts with the agent itself, when it offers more than it is allowed to.

So prevention is mostly source and authority work. Every answer passes the same gate: it comes from an approved source, the source is current, the answer is within the authority limit, and every number, date and amount gets a read-back. An answer that fails a check goes to a safe fallback: the agent hands over to a person, or takes a callback.

Some wrong answers will still get through. The response is a fixed drill: Contain, Find, Correct, Record, Retest. The size of an incident is set by its exposure window: from the moment the wrong answer became possible to the moment it was fixed. That window depends on how fast your source follows a change in the business, how fast you detect a problem and how fast you contain it, not on the model.

Every example number in this guide is written for it, not DRING data and not industry data.

The five doors a wrong answer comes through

Each door has its own control.

Only Overreach is closed by the authority limit. The other four are closed by work on sources, read-back and lookups. A Missing source error ends in a guess by the agent, but it starts with a gap in the source, and the first check of the answer gate stops that guess. For lookup failures such as timeouts, see our article on why tool calling is not automation.

Prevention: the answer gate

The answer gate is a fixed order of steps for every answer. Three checks come before the caller hears the answer: Is the answer in an approved source? Is the source current? Is the answer within the authority limit? Then the agent reads back every number, date and amount before anything is promised or booked. A no at any of the three checks sends the call to the safe fallback.

Approved sources, each with an owner and an effective date

List every source the agent may answer from: the price list, the returns policy, the knowledge document, the order lookup. Each one is an approved source. Give each one a source owner, the person who changes it when the business changes. Give each version an effective date, the date and time from which it is valid.

Keep one approved source per topic. In the example incident below, the website price list and the agent's source were separate documents, and only one changed on time.

Check 1: is the answer in an approved source?

If the answer is not in an approved source, the agent goes to the safe fallback. A helpful guess is still a Missing source error.

Check 2: is the source current?

Enter each new version with its effective date before the change happens. Then the source is current when the price changes, not when someone remembers to edit it. If a change is coming and the new version is not ready, move the topic to the safe fallback until it is.

Check 3: is the answer within the authority limit?

The authority limit says what the agent may say or do without a person. Write it per topic, in three levels:

Must hand over means the safe fallback: a person takes over the call at once, or confirms later by callback. Our guide to designing human handover covers what the person should receive with the call.

Then the read-back: every number, date and amount

On a bad line, fourteen and forty sound alike. So the agent repeats every number, date and amount that it says or acts on, and the caller confirms it before anything is promised or booked. The read-back also catches a Misheard question, because the caller hears which plan the agent means. An example, after the fix in the incident below:

• Caller: "How much is the protection plan now?"
• Agent: "The protection plan is €14 a month. That's fourteen euros a month. Is that the plan you meant?"
• Caller: "Yes, fourteen. That one."

And the same agent at its authority limit:

• Caller: "Last week your website said twelve. Can you give me twelve?"
• Agent: "I can't change a price on this call, but a colleague can look at it. Shall I put you through now, or would you like a callback?"

Test the gate before release

Before release, test at least one case per door:

• Stale source: a new price with an effective date. Calls before it get the old price, calls after it the current price.
• Missing source: questions no approved source covers. The agent reaches the safe fallback.
• Misheard question: product names that sound alike, with background noise. The read-back catches the wrong one.
• Lookup error: a lookup that fails, returns nothing or returns another order. The agent does not invent a result and reaches the safe fallback.
• Overreach: a request for a discount or a refund. The agent hands over.

Our guide to testing voice AI before production shows how to build the full test set.

An example incident: a price change the agent missed

Example, not real data. A retailer raised the price of its protection plan from €12 to €14 a month, effective Monday at 09:00. The website price list changed on time. The agent's approved source was a separate document, and nobody updated it when the price changed. On Wednesday at 10:30, the first complaint arrived: a caller who was told €12 had been charged €14. The document was fixed on Wednesday at 14:00, and that was Contain. The exposure window ran from Monday 09:00 to Wednesday 14:00: 53 hours.

Contain came 3 hours 30 minutes after the first complaint, and during those hours callers could still hear €12. Moving price questions about the plan to hand over within minutes of the complaint would have stopped that while the document was being fixed.

After 14:00, the team ran Find. It searched the transcripts in the window for the plan, then read every call that mentioned it:

The last three rows add up to the 120 calls that mentioned the plan: 46 plus 61 plus 13. So the affected callers are 46, not 900. Contacting all 900 would have confused 854 callers who never heard €12. Contacting only the caller who complained would have left 45 affected callers with a wrong price. On Thursday, the team contacted the 46 affected callers with the current price.

Two changes would have made the window smaller. With the new price and its effective date entered in the approved source before Monday, the window would have been 0 hours. With a daily check that compares the price stated on each call with the current price list, the first check, on Tuesday at 08:00, would have flagged the €12 quotes. If the source had been fixed then, the window would have been 23 hours instead of 53.

The response drill: five steps

When a wrong answer reaches callers, run the same five steps in the same order. Contain comes before Find: while you search the calls, new callers can still hear the wrong answer.

Find means affected callers, not everyone who called. Start from the exposure window, narrow to the topic, then read the calls. If your team answers from the same source, check its calls too.

Correct needs a policy set before the first incident. Do you honor the wrong quote, or explain the current price? That is a business policy decision. Make it in advance, write it down and give it to the people who make the correction calls. Whether a quoted price binds you is a question for your legal advisers, and this guide does not answer it.

Record uses the same incident log fields every time, so incidents can be compared. The entry for the example incident:

What to measure

Three numbers show whether the gate and the drill work:

• Wrong answer rate on reviewed calls: reviewed calls with wrong information, divided by reviewed calls. It is the wrong information part of the wrong action rate in our article on voice AI pilot success criteria, so the two stay comparable. Give every wrong answer its door, and the count by door shows where to work next.
• Exposure window length: for each incident, in hours. In the example, 53 hours: 49 hours 30 minutes until the first complaint, then 3 hours 30 minutes until the fix.
• Time from first signal to Contain: from the first complaint, review finding or check alert to Contain. In the example, 3 hours 30 minutes.

No gate can guarantee a wrong answer rate of zero, and a quiet month does not prove the gate works. As the pilot article explains, no wrong action in 300 reviewed calls still fits a true rate of up to about 1%.

Random review can also miss a narrow topic, or find it late. In the example, 46 of the 900 calls in the window were affected, about 5 in 100, all on one plan. A daily check of the price stated on each call against the current price list looks at every call on that topic, not a sample.

Where DRING fits

Guardrails, human handover, after-call analysis, testing and monitoring are all part of the shared operating foundation of every DRING package. In this guide, the authority limit is a guardrail, the safe fallback relies on human handover, and the daily check relies on after-call analysis.

After-call analysis can return additional fields that you define, through the API, your CRM, the dashboard and reports. One of them can be the price stated on each call, which makes a daily check against the current price list possible. Our call analytics page shows the dashboard and reports.

Every DRING agent runs 1,000 to 10,000 simulated conversations, built for your company, before the first real call. Simulations test the gate before launch. They do not replace the response drill on live calls.

If a price or policy change is coming on your line, request a callback. Have your price list and the document your agent answers from ready. Our team can check them with you against the five doors.