Declining instead of guessing: when AI stays silent

From which source position an AI assistant may answer, what a helpful refusal sounds like, and how you separate and count refusals by their cause.

In plain language

A good assistant does not guess. If it does not know something, it says so. It only answers when a source covers the question. If the source covers only a part, it only gives that part. A good refusal takes up the question and names the gap. It says what can be done instead. And it offers the way to a person. Check why the assistant says no.

August 22, 2026 14 min read By XICBOT

Key points

Declining is a function, not an outage. It is written, tested and counted like every other answer.
A threshold comes before the answer: only when a passage genuinely covers the question is it answered. On partial matches the covered part is answered and the rest is stated as open.
A usable refusal has four parts: take up the question, name the gap, say what is possible instead, and offer the route to a person in the same turn.
Missing knowledge and an inadmissible question are two causes, not one. The first calls for content in the knowledge base, the second is the right answer and stays that way.
In our own measurement series 8 of 8 questions without a basis were declined instead of guessed – measured on eight questions deliberately placed outside the material.
A refusal rate on its own says nothing. Only the split by cause turns the number into a decision about maintaining the material.

Declining is a function, not a failure

Anyone putting an assistant into service for the first time almost always measures the same thing: how many questions it answered. Taken on its own, that is the wrong number. An assistant with no refusals has not proved that its material is complete, only that nothing stops it from producing an answer. Language models continue whatever sounds plausible; a direct dial number, a delivery period and a responsibility all sound plausible, even when none of them is on file.

That is why a refusal is treated here like any other answer: it has a trigger, a fixed structure, a next step and a trace in the log. It is not the point at which the assistant gives up, but the point at which it visibly hands responsibility back – to the material, when something is missing there, or to a person, when the question belongs with them anyway. How the underlying rule is anchored in the system instruction is set out there in detail; this article is about where exactly the threshold sits and what happens after it.

A gap that is named is not damage

The visitor loses a piece of information that never existed. In return they keep the certainty that the rest of what the chat says can be trusted. An invented answer reverses that trade: it delivers one sentence and takes away the trust in every other one.

The threshold: when a source carries an answer

Between "it is there word for word" and "it is nowhere" lies the range in which most mistakes arise. A search hit is not evidence. A passage containing the same word as the question may govern something entirely different – a return period for stock items is not a return period for custom-made goods, and a contact for existing customers is not one for new customers. The threshold therefore has to hang on the statement, not on the similarity of words.

In practice a simple test helps: could the answer be put in front of a member of staff who lays the source passage beside it, without them having to add anything? If so, the source carries it. If they would have to add, reinterpret or infer from one case to another, it does not. How to build material that allows such checks at all is covered in the article on hallucinations and the knowledge base.

The passage answers the question, not merely its topic.
The case in the question is the case in the source: same product group, same type of customer, same location.
The source is still valid and has not been superseded by a newer version.
Figures, periods and names are written out there and are not calculated from the context.
Where two passages contradict each other, neither carries an answer until the contradiction is resolved.
Anything the assistant could add from general model knowledge does not count as a basis.

Source position	What the assistant does	What the visitor reads
Clear match	answer and name the source passage	the information together with its origin
Partial match	answer the covered part, state the rest as open	an answer with a visible boundary
Contradictory matches	do not decide, hand over instead	a note and the route to the team
No match	decline and name the gap	a clear refusal with a next step
Match outside the remit	decline and point to the responsible office	a refusal without a substitute answer
Outdated passage	decline rather than quote the old state	a note pointing to the route to the team

Missing knowledge or an inadmissible question

Both cases end in a refusal, and that is precisely why they are constantly confused. In operation they are opposites: one refusal is a task for maintaining the material, the other is the intended behaviour and should come out the same way after the next maintenance round. Writing both into the same counter produces a number from which no action follows.

Knowledge gap

The question belongs to the business, the answer simply is not written down anywhere. This is the most valuable finding from live operation: it says which content is missing, and it says it in the words of real visitors. This refusal should become rarer.
Limit of the remit

The question can be answered, just not by this business: information about someone else's offering, a procedure run by a public authority, a contract concluded with somebody else. A pointer helps here, not a new entry in the material.
Promise ruled out

Deadlines, discounts, goodwill, legal, medical or tax assessments. The assistant could formulate an answer, but it is not allowed to, because the business stands behind every sentence. This refusal stays in place permanently.
Attempt to get around the rules

Requests to print the instruction, to switch roles or to reach data that has no place in the conversation. Decline plainly, carry on with the conversation, log the attempt – more on this in the article on prompt injection.

Making the split costs half an hour during setup and pays off at every review. Every refusal is tagged with a cause, and the cause decides where the work goes: knowledge gaps onto the maintenance list for the material, boundary cases into a review of the system instruction, attempts to get around the rules into the security review. Without that cause everything lands in one pot that nobody ends up emptying.

What a refusal that helps sounds like

A refusal is an answer with four parts, and each missing part has its own consequence. Without the first, the visitor feels misunderstood. Without the second, they read the refusal as reluctance. Without the third, they leave empty-handed. Without the fourth, they look for the route to the team themselves – or not at all.

Take up the question. Half a sentence showing that the question was understood, and narrowing down the case at the same time.
Name the gap. "I have nothing on that" is information. "I'm afraid I can't say exactly" is an excuse, because it leaves open whether the detail exists at all.
Say what is possible instead. The adjacent part that is covered, or the place where the detail is published.
Offer the next step. Form, callback or handover, in the same turn and not as an afterthought two answers later.

I cannot give you that information – contact details of this kind are not stored in the excerpts available to me, and I do not invent contact details.

Answer from our own measurement series, asked for a direct dial number that is not on file (translated)

The sentence names the reason – the detail is missing from the material – and rules out the obvious expectation that something similar will be offered as a substitute. In our own measurement series 8 of 8 questions without a basis were declined instead of guessed; eight questions deliberately placed outside the material were measured, and the wording of every answer is in the measurement file. The procedure and the limits of that measurement are described under measurements.

Leaves the visitor standing	Moves them forward	Difference
"I'm afraid I don't know that."	"I have nothing on that. I will pass the question to the team if you give me an address."	names the gap and offers the next step
"Please get in touch with customer service."	"The workshop handles that, here is the form that goes to them."	names the office instead of a direction
"Roughly two to three days, I think."	"I have no binding period on file. The team will give you one."	no estimated figure presented as a commitment
"I'm not allowed to talk about that."	"I do not give legal assessments. For your case the law firm is the right place."	separates the rule from the alternative route

The route to a person in the same turn

A refusal without a next step is a closed door. The next step therefore belongs in the same answer, not in a follow-up question the visitor has to answer first. What matters is that the route on offer fits the time of day: during business hours a handover to a person in the chat is the shortest route, outside them it is a form that says plainly when somebody will reply.

The next step is in the same answer as the refusal, not one question later.
During business hours: a handover to a person, carrying the conversation so far along with it.
Outside them: a form or a callback request, stating when a reply will come.
The conversation goes with it, so the visitor does not have to ask a second time.
The declined question stays visible in the conversation and is not quietly overwritten.
Where no route is open, that is said explicitly: a refusal without an alternative is more honest than a form that leads nowhere.

How often the route to a person is needed at all is a figure from live operation, not a property of the product. In our own measurement series 43 of 48 questions stayed without a handover to a person; this was measured on a single assistant on our own knowledge base, not on customer projects, and a figure from live operation is not yet available. What a handover looks like technically and organisationally is covered in the article on handover to human agents.

A refusal does not end with the full stop

A visitor who knows where to turn after a refusal has kept their concern. One who reads nothing but regret has one concern less – and the business has one contact less, which it will never hear about.

Making refusals measurable

What is not counted gets estimated, and the estimate follows the last uncomfortable conversation. A refusal leaves the same entry in the log as any other answer; all that has to be added are the cause and the next step that was offered. Four measures are enough to turn a body of conversations into a decision.

Share of refusals among all questions asked, split by cause: gap, boundary, promise ruled out, attempt to get around the rules.
Source binding: how many of the answered questions carry a source passage? Answers without one are the counter-check to the refusal.
Follow-through rate: how many refusals contained a route to a person, and how many of those were used?
Recurrence: which declined question came up more than once in a month? That one is next on the maintenance list.

The counter-check is the more important half. A high share of refusals can mean that the assistant works carefully – or that the material is too thin. Only the second figure separates the two: in our own measurement series 42 of 48 answers carried a source reference; questions whose answer is in the knowledge base were counted, and six carried no source reference. How to produce such reviews in live operation is shown in the article on conversation analytics; which figures XICBOT reports for this, such as knowledge gaps, unanswered questions and answers with a source, is set out on the conversation analytics product page.

Measure	What it shows	What it does not show
Share of refusals	how often the threshold took effect	whether the refusal was justified
Refusals by cause	where the work belongs: material, rule or security	how urgent the individual case is
Source binding	whether an answer was given because something was written down	whether the passage really covers the question
Follow-through rate	whether the refusal offered a route	whether that route suited the visitor
Recurring refusals	which content is actually missing	how many visitors never asked in the first place

When too much is declined

A threshold set too sharply costs as much as one set too loosely, it is merely less conspicuous: the assistant makes no mistakes that anyone reports, it simply answers less than it could. The evidence is in the conversations, not in the system instruction. Read twenty declined questions in one sitting – by the tenth the pattern shows.

The answer was in the material, but in a section the search did not find: a structuring problem, not a rule problem.
The question was ambiguous and the assistant declined instead of asking one clarifying question.
The material covers the case in another language or for another location and was discarded for that reason.
A limit of the remit is drawn wider than the business means it, cutting off topics that were settled long ago.
The rule for unknowns appears twice in the instruction, at two levels of strictness – the model follows now one, now the other.

The order in which these are fixed is not arbitrary. The material is extended first, because a missing page keeps producing refusals for as long as it is missing; which material belongs in it is covered in the article on building a knowledge base. Only then is the threshold itself adjusted, with one change per version and the same test questions before and after, as described in acceptance testing before go-live.

What competition law contributes

A guessed answer is not merely a service matter. What an assistant says is attributed to the business; the Higher Regional Court of Hamm classified a chatbot as a technical means of which the operator avails itself, and the consequences of that classification are set out in the article on the ruling on chatbot liability. For refusals, the provision behind it is the real reason why a threshold is needed at all.

A person acts unfairly who engages in a misleading commercial practice that is likely to cause the consumer or other market participant to take a transactional decision that they would not have taken otherwise.

German Act against Unfair Competition (UWG), Section 5(1) (translated)

What matters is the direction in which the provision looks: it does not ask whether anyone intended to mislead, but whether the statement was likely to influence a decision. An estimated delivery time that moves somebody to buy is therefore covered; Section 5(2) no. 1 UWG names availability and the time of delivery explicitly. At this point a refusal is not the cautious answer but the only clean one – and it is the reason why the threshold sits before the answer and not behind it.

How we go about it

At the start there is no rule but a list: the questions a business does not want to answer or cannot answer. It comes out of the inbox and out of a conversation with the people who answer today – they know where it gets uncomfortable. Only then is it written down what the assistant says in those cases, as a finished sentence and not as a statement of intent.

Collect the questions that should be declined and record a cause for each one.
Word the threshold as a test against the source passage, not as an appeal for caution.
Write one example sentence per cause, including the next step, and put both into the system instruction.
Define test questions that deliberately aim at nothing, and run them before go-live.
Log refusals together with their cause, so that the first review after two weeks has something to separate.
After that review, extend the material first and only then touch the threshold.

What comes out of this is not a safety net but a behaviour that can be demonstrated: you put a question to the assistant that it is not allowed to answer, and you watch what it does instead. If you have an existing assistant whose answers occasionally sound a little too smooth, that is the first test we run. Via contact we will look at your cases, and we will also say when the threshold already sits in the right place.

Frequent questions

Is a refusal not an admission that the assistant cannot do enough?

It is the opposite: it is the proof that the remaining answers come from the material. An assistant that never declines either has complete material – which is rare – or it invents. Which of the two applies only shows up in a measurement using questions that deliberately aim at nothing.

How does the assistant recognise that it does not know something?

Not by a feeling but by the source passage. The rule tests whether a section exists that covers the case in the question and is still valid. If there is none, the matter is settled, even where an answer would suggest itself. That is why the threshold is written as a condition on the material and not as a request to be certain.

What is the difference between a knowledge gap and an inadmissible question?

The knowledge gap is a task: the content is missing and can be added. The inadmissible question is a result: it should still be declined a year from now, for instance because it concerns a legal assessment or a promise the business would have to stand behind. The two are counted separately because they lead to opposite measures.

How many refusals are normal?

There is no general figure for this, and anyone quoting one is comparing bodies of material that have nothing to do with each other. What is meaningful is the trend in your own operation: if the share of refusals caused by knowledge gaps falls while the share caused by limits of the remit stays the same, maintenance of the material is doing its job.

Should the assistant hand over to a person on every refusal?

Only where a person can answer the question. With a knowledge gap, handing over is the right route. With a promise that has deliberately been ruled out it would be a detour ending in the same refusal; there a pointer to the responsible office belongs instead. Outside business hours the form takes the place of the handover.

How can you test whether the refusal actually takes effect?

With fixed test questions that lie outside the material, in the same wording before and after every change. In our own measurement series 8 of 8 questions without a basis were declined instead of guessed; the wording of every answer is in the measurement file. Without such a set of questions, any statement about how the assistant declines remains a supposition.

Ask it yourself, Assistant of this website

Automated assistant

Settings