AI Phone Assistant Without Hallucinations: 2026 Guide
An AI phone assistant without hallucinations answers factual questions only with content that is documented in your knowledge base – your website, your documents or your answer catalog. If a piece of information is missing, it says so honestly and offers a callback instead of guessing. After every call, a second model checks whether each answer was backed by a source and shows the result as a traffic light. In this guide you'll learn why voice AI makes things up on the phone, what that can mean for your business and how to spot a provider that actually solves the problem.
Why does an AI phone assistant make up answers at all?
An AI phone assistant makes up answers because language models are trained to produce plausible sentences – not to know when they don't know something. Without a firm knowledge boundary, the model fills gaps with guesses that sound just as confident as real facts.
In a research post, OpenAI describes hallucinations as "plausible but false statements" and names a central cause: standard training and evaluation procedures reward guessing over acknowledging uncertainty (OpenAI: Why language models hallucinate, 2025). A model that says "I don't know" scores zero on many benchmarks – one that guesses is sometimes right. According to the post, rare, arbitrary facts that can't be inferred from patterns are especially prone to errors.
Those are exactly the facts callers ask about all the time:
- Are you open on Saturday?
- How much is an initial consultation?
- Do you offer home visits?
- How long does a repair take?
On the phone, two amplifiers come on top. First, the answer has to arrive within a fraction of a second for the conversation to feel natural – there's no time to check it afterwards. Second, there's no screen: the caller sees neither a source nor a note like "AI can make mistakes". They only hear a friendly voice stating something.
What happens to your business when the AI invents answers?
Invented answers on the phone lead to wrong prices, wrong opening hours and promises you can't keep – and in the worst case to damage claims. To the caller, the assistant speaks on behalf of your company, not on behalf of a software vendor.
Typical damage in everyday business:
- Wrong prices: The assistant quotes a fixed price that doesn't exist. The customer refers to it later.
- Wrong opening hours: Someone drives to you for nothing because the AI invented Saturday hours.
- Wrong promises: "Your insurance covers that" or "That's free of charge" – sentences like these can get expensive.
- Loss of trust: One obviously wrong statement is enough for callers to distrust your entire hotline.
The Air Canada case: the company is liable for its bot
A case from Canada shows how real the risk is. Air Canada's chatbot told a customer he could apply for a bereavement discount retroactively – a rule that didn't exist in that form. In February 2024, the Civil Resolution Tribunal of British Columbia ordered the airline to pay the difference and rejected the argument that the chatbot was responsible for its own statements (Moffatt v. Air Canada, 2024 BCCRT 149; German summary at beck-aktuell, 2024). The tribunal made it clear: what the bot says is like information on the company's own website.
What about German law?
No German court has ruled on this question yet. An analysis by the law firm HLC considers liability for negligent misinformation under pre-contractual obligations (culpa in contrahendo, Sections 311 and 241 of the German Civil Code, BGB) possible in principle; it is still open whether the known risk of hallucination already amounts to a breach of the duty of care (HLC, 2024). For you, this means safeguards like a knowledge boundary and a documented review aren't just a quality topic – they're evidence of acting with due care. This article is not legal advice; for your specific case, talk to a lawyer.
On top of that, the transparency obligation in Article 50 of the EU AI Act (Regulation (EU) 2024/1689) has applied since 2 August 2026: callers must be told they're talking to an AI. Read more in our post on the AI Act for voice agents.
How does a knowledge boundary work in an AI phone assistant?
A knowledge boundary means the AI phone assistant answers factual questions only from approved sources: your website, your documents and your answer catalog. If it finds nothing suitable there, it admits it openly and offers a callback or a transfer to a human.
Technically, the principle is often called retrieval-augmented generation (RAG): before every answer, the system searches the knowledge base for matching passages and phrases the answer only on that basis. The decisive difference to a simple RAG chatbot is the rule for when nothing is found. A free model answers anyway from its general training knowledge. An assistant with a hard knowledge boundary says something like this instead:
"I'm afraid I don't have reliable information on that. May I note a callback request? Then someone from the team will get back to you."
How to build the knowledge base:
- Import your website: services, prices, opening hours, directions, FAQ – ideally several pages, not just the homepage.
- Add documents: price lists, service descriptions, excerpts from your terms as PDF or DOCX.
- Maintain an answer catalog: short, verified answers to the most common questions – phrased the way they should sound on the phone.
- Set limits: What must the assistant never promise (e.g. binding fixed prices, medical assessments, legal advice)?
Speech recognition matters too. If the assistant hears "Friday" as "holiday", even the best knowledge boundary will correctly answer the wrong question. So look for robust recognition and make sure the assistant asks back when something is unclear.
Reviewing every call: the traffic light
A traffic-light review means that after every call a second model compares each of the assistant's answers with the sources and marks the result by color. That way you see at a glance which calls were clean and where you need to step in.
| Light | Meaning | What you do |
|---|---|---|
| Green | All answers were backed by sources | Nothing – spot checks are enough |
| Yellow | Questions were left open | Check the knowledge gap and add an answer |
| Red | The assistant claimed something unsupported | Read the call, call the person back if needed, fix the source or setting |
The advantage over classic quality assurance: you no longer listen to random calls – you get the critical cases pre-sorted. How to set up metrics and review routines beyond that is covered in our post on voice agent quality management.
Approving knowledge gaps: how the assistant learns under control
Every question the assistant couldn't answer shows up as a knowledge gap in the dashboard. Identical questions are grouped and counted – so you immediately see which information is really missing often. A human answers the gap or dismisses it. Only after this approval does the answer go into the answer catalog and get used as a source from the next call on.
That's the core of a controlled learning loop: the assistant gets better every week, but never by making something up on its own. Every new answer has a human sender.
Free AI vs. AI with a knowledge boundary
The table below shows the key differences between a freely answering language model and an AI phone assistant with a knowledge boundary.
| Criterion | Free AI on the phone | AI phone assistant with knowledge boundary |
|---|---|---|
| Knowledge source | General training knowledge plus prompt | Only website, documents and answer catalog |
| When information is missing | Guesses a plausible answer | Says "I don't know" and offers a callback |
| Prices and opening hours | May be invented or outdated | Only as stated in your sources |
| Review after the call | Spot checks, manual listening | Automatic traffic-light review of every call |
| Handling gaps | Stay invisible | Collected as knowledge gaps and approved |
| Liability risk | Hard to estimate | Significantly reduced and documented |
| Small talk | Unlimited | Allowed or declined depending on the level |
A knowledge boundary doesn't make the assistant "dumber". It makes it predictable – and on the phone, that matters more than eloquence.
Checklist: how do you recognize an AI phone assistant without hallucinations?
You recognize an AI phone assistant without hallucinations by the fact that it ties answers to sources, admits gaps openly and makes every call reviewable. Ask providers these questions before you buy:
- Is there a hard knowledge boundary? Can I set it so factual questions are answered only from my sources?
- What happens when information is missing? Does the assistant honestly say "I don't know" and offer a callback or transfer?
- Which sources can I import? Multi-page website crawl, PDF, DOCX, my own answer catalog?
- Is every call reviewed? Is there an automatic review with a clear rating – not just a transcript?
- How do I handle knowledge gaps? Are open questions collected, counted and used only after my approval?
- Can calls be transferred live? Is there a handoff to a human and a callback note when nobody is available?
- Does the assistant identify itself as an AI? Article 50 of the AI Act requires it.
- Where is data processed? Is there a data processing agreement and a transparent list of providers? Details in our post on GDPR-compliant AI telephony.
- Can I test it first? A free test with your own website shows faster than any sales deck whether the assistant makes things up.
A good real-world test: in the test call, deliberately ask questions whose answers aren't on your website. A trustworthy assistant shouldn't improvise.
How anicall.io does it
The anicall.io AI phone assistant is built around exactly this principle: it only says what's on your website and in your documents, and books appointments straight into your calendar. Specifically:
- Import knowledge: website (multi-page crawl), PDF, DOCX and text files, YouTube transcripts, RSS feeds and the answer catalog form the knowledge base.
- Answer only what is grounded: before every answer, anicall searches the knowledge base for matching sources. If there are none, the agent says so honestly and offers a callback.
- Traffic-light review: after every call, a second model checks each answer against the sources – green (all grounded), yellow (questions left open), red (unsupported claim).
- Approve knowledge gaps: unanswered questions appear in the dashboard with a count. You answer or dismiss them; approved answers apply from the next call on.
- Transfer and callback: live transfer to a configured number; if nobody is available, the agent records a callback request.
- AI disclosure: the agents identify themselves as a digital assistant in the greeting and confirm it when asked.
You set the knowledge boundary per agent in four levels:
| Level | Behavior |
|---|---|
| Grounded only | Answers factual questions exclusively from the knowledge base |
| Strict | Sticks to approved tasks, declines other topics |
| Balanced | Brief small talk, then back to the core tasks |
| Open | Also answers questions outside the core tasks |
Billing works with prepaid tokens – no base fee and no subscription; every feature, including the knowledge boundary and the review, is available to every customer. An example: 300 inbound minutes per month plus one phone number come to 4,080 tokens, i.e. €40.80 net. See pricing for details and security & privacy for information on the database and providers.
Frequently asked questions
Can an AI phone assistant really never hallucinate?
No language model is error-free, so don't believe any provider who guarantees it. A knowledge boundary lowers the risk significantly because factual answers are tied to sources, and the traffic-light review makes remaining errors visible so you can react.
What's the difference between RAG and a knowledge boundary?
RAG describes the technique of searching for matching sources before answering. The knowledge boundary is the rule for what happens when no source is found: instead of improvising from training knowledge, the assistant admits the gap and offers a callback.
Do I need to rebuild my website for the assistant to give good answers?
Not necessarily. Clearly written pages on services, prices, opening hours and common questions help. Whatever is missing on the website you add via documents or the answer catalog – and after the first calls, the knowledge gaps show you what's still missing.
Who is liable if the AI assistant gives wrong information?
To the caller, the assistant speaks for your company. In the Canadian case Moffatt v. Air Canada (2024), the airline had to pay for its chatbot's wrong information; in Germany, courts haven't ruled on it yet, but liability is considered possible. When in doubt, get legal advice.
How do I test whether an assistant makes up answers?
Import your website, run a test call and deliberately ask questions that aren't answered there. Afterwards, the traffic light shows whether every answer was grounded – at anicall.io this agent test in the browser is free.
Try it yourself: in the live demo on anicall.io you can call the AI reception right in your browser or request a demo call to your phone (the demo agents speak German). Then start for free on anicall.io, import your own website and test how your assistant handles questions it isn't allowed to answer.