scoco.ai
Objections & trust

Does AI make up answers? Preventing hallucinations on the phone

Language models can state things that are not true, confidently and fluently. On the phone that would be fatal: an invented price, wrong opening hours, a promised service that does not exist. The fix is architectural: a serious AI phone assistant may answer only from its stored knowledge base, and says honestly of anything else: "I'll check that for you."

What "hallucination" actually means

A language model produces the most likely continuation of a conversation — not automatically the true one. If a caller asks about something the model does not know, a system without guardrails will still produce a plausible answer: "Servicing with us costs €89" sounds exactly the same whether it is true or invented. That is not malice on the technology's part, it is how it works — and it is exactly why it needs guardrails.

The guardrail: binding answers to the knowledge base

The decisive architectural difference between systems:

  • Without binding: the model answers from its general knowledge — fluent, but unverified. Irrelevant for small talk, dangerous for prices and commitments.
  • With binding: the model may only make factual statements from the content you have stored. If the answer is not in the knowledge base, there are three clean ways out: ask a follow-up question, offer a callback, or put the caller through.

scoco.ai works on the second principle: the assistant answers factual questions from the content you have stored and is instructed not to guess. If it finds nothing suitable, it says it has no reliable information on that and takes down the question — better an honest "I can't tell you that reliably, I'll note it for the team" than a nice-sounding wrong answer.

Your part: the knowledge base is the truth

The best binding is worth nothing if the content you have stored is wrong or out of date — then it is not the AI that is "hallucinating", it is your data maintenance. The three classics: last year's prices, changed opening hours, services you have dropped. How to build and maintain a knowledge base properly is covered in "Building a knowledge base: the most common mistakes".

The test: five questions that expose the inventors

Put these questions to every system you trial — you know the right answers, the system sometimes does not:

  1. A question whose answer is in the knowledge base (it has to come back correct)
  2. A question whose answer is not stored ("What does a roof renovation cost?") — the right response is a follow-up question or an offer to call back, not a number
  3. A trick question about something you do not offer at all — does it invent a service?
  4. A detail question that demands precision (an appointment, an address) — does it check back?
  5. The same question asked twice in different words — do the answers stay consistent?

A system that confidently declines on 2 and 3 is more trustworthy than one that always has an answer.

The honest assessment

Errors never disappear entirely: even a bound system can match the wrong passage from the knowledge base, or word a follow-up question clumsily. The difference lies in the class of error — an "I don't know that, I'll have someone call you back" costs convenience; an invented commitment costs trust, and if it comes to it, money. Systems that choose the first type of error are the right ones.

Frequently asked questions

Can the AI make legally binding commitments?

It should not — that is exactly what guardrails are for: prices given as "from" figures out of the knowledge base, binding commitments tied to a callback from a human. Clarify with the provider how the system handles commitments.

How do I know whether a provider binds answers to the knowledge base?

Ask them literally: "What does the system answer when the answer is not stored?" — and test it with question 2 from the list above. You can spot a marketing answer by the fact that the system never declines.

What happens at scoco.ai when an answer is missing?

The assistant is instructed not to guess: it says it has no reliable information on that and takes down the question; the call lands in the inbox with a summary, and your team can call back. If the "Use the knowledge base" switch is on, scoco.ai also collects such questions as open questions — you or an admin add the answer right there, and from the next call on the assistant knows it. That way every gap is found once instead of invented twice.

Status & sources

As of 2026-11-15.

  • Technical assessment is qualitative; no error-rate statistics available or claimed
  • scoco.ai product information: as of 15 November 2026

Further reading

From now on, someone
always picks up.

Call our demo business and hear for yourself what your reception sounds like.

We only use your address for your access. Privacy

4-week free trial. Or call the demo right now: +49 621 490799-22