Threll.ai
Knowledge sources

The agent answered. Where did it get that from?

An answer from a voice agent sounds equally certain whether it comes from an approved price list or from a note nobody has read since last year. Seven questions that decide which one the customer gets.

4 min read
A woman in her thirties in a dark teal jumper leans against the counter in an office staff kitchen at blue hour, a mobile phone to her left ear and her gaze lifted to the right, her right hand resting on a blank sheet of paper on the worktop, a colleague out of focus and turned away in the foreground

An agent that answers wrongly sounds exactly like an agent that answers correctly. That is the whole problem.

On a web page the customer can see where the text sits. She can see whether it is a pricing page or a blog post from last year, and often when it was last changed. In a phone call none of that exists. One voice arrives, in one tone, and everything it says carries equal weight – the price you approved last week and the sentence someone wrote in an internal note last autumn.

So the question is not whether your agent answers well. It is where the answer came from, and who owns that place.

The five places an answer can come from

Source Good for How it fails
The system instruction Rules, tone, what the agent must never say Grows long and contradicts itself without anyone reading it end to end
Uploaded documents Price lists, terms, procedures Nobody has any relationship with them after week three, and the version is rarely dated
A lookup in a back-office system Order status, free appointments, balance – whatever changes hourly Requires integration, and fails loudly when the system is down
Your website What the customer could have read anyway May have been ingested once, and does not refresh itself
The model itself Ordinary language and coherence It fills the gaps. It does not know that it is doing so

The first four are sources someone chose. The fifth was chosen by nobody. It steps in the moment the others do not cover the question, it sounds exactly as certain as the rest, and in the log it looks like any other answer. The day someone asks what the agent actually said, that is the row that is hardest to account for.

Row three deserves a note. It is the most expensive one to put in place, and far fewer calls need it than people assume.

Want to talk it through with someone who has done this before?

Fifteen minutes on a call. We will tell you honestly whether a voice agent fits your setup – and what has to be settled before it goes live.

Book Meeting (opens in a new window)+47 24 12 66 30

15 minutes · no obligation · pick your own time

Seven questions to settle

1. Which sources are actually switched on? Not the ones you planned. The ones marked active in the configuration today. A knowledge-base search that was turned on during testing and never removed is still answering customers.

2. Who owns each source? A name, not a department. A source with no person behind it does not get updated. It gets discussed.

3. What is the oldest date in the material? Find the oldest document the agent can answer from, and read it. Was it right when it was written, and is it right now? Opening hours are the classic case: they stop being true on their own, without anyone touching a setting.

4. What happens when two sources disagree? The price list says one thing, the terms another, and a procedure note from last year a third. If nobody has set the order of precedence, the model sets it – slightly differently from call to call. Write down which source wins, and what the agent should do when it meets a contradiction.

5. What does the agent say when the answer is nowhere? This is the most important of the seven, and the one most often left open. “I don’t have a reliable answer to that, but I’ll put you through to someone who does” is not a defeat. It is the sentence that separates a knowledge base with edges from one without.

6. Which questions has the agent been asked that no source covers? They are already sitting in your calls. Take one week, sort the questions, and find the ones the agent answered with nothing to answer from. That list tells you what is missing – a workshop does not.

7. Who gets told when something changes? When you change the free-shipping threshold, someone has to change the text the agent answers from. If that is nobody, it takes weeks. And what the agent says in the meantime can bind you.

It does not throw an error

Five of the seven questions can be answered in a morning, and none of them needs a new platform. What makes them hard is that none of them announces itself.

A poor knowledge base does not produce errors. It produces answers – in the same calm tone as all the correct ones.

Threll.ai builds voice agents in Norwegian, Swedish and Danish. Try question five today: call your own number and ask about something you know is written down nowhere. What the agent tells you is what it tells everyone else.

Related Articles

Continue reading more articles

See all articles