Why your AI agent gave a wrong answer, and how to find the cause

An AI agent gives a wrong answer for one of a handful of reasons, and most of them are not the model. Before rewriting the prompt or switching platforms, find out what the agent was given on the turn it went wrong. That one check tells you whether the fix belongs in the knowledge base, the prompt or the platform's settings.

We learned this on live calls: a voice agent we built quoted one product's price for another, twice, and each time the model had done exactly what it was told with what it was given. This guide is the checklist we now use after every bad call. How to write the knowledge base itself is in our chatbot knowledge base guide.

First, look at what the agent was given

Most platforms show, for each turn of a call, the text they retrieved from the knowledge base and passed to the model. Retell AI shows it under Knowledge Base Retrieval in the call history; Synthflow shows every lookup in the call's Actions tab; Bland AI has a Source testing tab. Find the turn where the agent went wrong and read exactly what it had. Then match what you see to one of the seven causes below.

What you find on that turn	Likely cause	Where the fix is
A price or time with no product name next to it	1. A number without its owner	Knowledge base
Nothing relevant retrieved, and the knowledge base never covers it	2. The knowledge base is silent	Knowledge base
Nothing relevant retrieved, but the knowledge base does cover it	3. The search missed	Knowledge base wording, then retrieval settings
No search happened at all	4. The agent never searched	Prompt
Two different answers to the same question	5. Two answers in the source	Knowledge base
The right fact, but the agent said something else or added to it	6. The agent went beyond its source	Prompt
A fact that was true once	7. The fact is out of date	Knowledge base, and a refresh routine

1. A number without its owner

What happened to us: a caller asked the price of a day package, and the agent answered with 3,000 rupees, a villa's price. The retrieved text gave the amount without naming the villa, so the agent attached it to the package the caller had asked about. On another call, a weekend cottage price was quoted for a day visit the same way.

Platforms cut documents into chunks and hand the agent a few of them. A chunk that holds a number without the name of what it belongs to is an invitation to attach it to whatever the caller just mentioned. Retell AI's own documentation warns about this: avoid words like "it" or "this", "because prior chunks may not be present".

Fix: rewrite the source so that every sentence with a price, time, size or capacity names its product. "A first consultation with Dr Mehta costs 800 rupees" survives being cut out; "Consultation: 800" doesn't. The same goes for a school's fee per grade, a builder's price per unit type or a salon's price per service. Tables are the worst offenders: a row loses its column headings when chunked.

2. The knowledge base is silent

What happened to us: a caller asked what a package included, and the agent said "everything is included". The website listed meals and activities, and never said that some activities cost extra. The agent filled the silence with the friendliest answer.

When the knowledge base doesn't answer a question, the model doesn't say nothing. It says something plausible. For a clinic that might be "yes, medicines are included"; for a school, "yes, transport is available".

Fix: write a sentence for every question callers are likely to ask, including the ones the website doesn't answer: "Whether medicines are included in the IVF price has not been confirmed; please call the clinic." Then the agent has something true to say. Knowing which questions to cover is the hard part: it depends on the industry, which is why we work from a list of the questions callers of each kind of business ask.

3. The search missed

The fact is in the knowledge base, but it wasn't retrieved on that turn. Common reasons:

• The caller didn't repeat the product's name. "And how much is that one?" searches for "how much is that one". Bolna, for example, searches with the caller's latest message; Retell condenses the recent conversation into a query, which helps.
• The caller's words aren't the website's words. "Fees" versus "tuition", "possession" versus "handover", "consultation" versus "appointment".
• Too few chunks. Retell and Voiceflow retrieve 3 by default. For a business with ten similar products, the right one may be the fourth.

Fix: first the wording: make each section name its product and use the words callers use. Then the settings: raise the number of chunks retrieved a little, or use the platform's query instruction (Retell has one) to carry the product's name into the search.

4. The agent never searched

On some platforms, searching the knowledge base is a tool the agent decides whether to use. Vapi's documentation is explicit: "You must explicitly instruct your assistant in its system prompt about when to use the query tool." Bland AI's agent decides from a source's name and description; Synthflow searches when a condition you write applies. If the agent answered a price question without a search, it answered from general knowledge.

Fix: a prompt line naming the tool and when to use it, for example "Always search the knowledge base before answering questions about prices, timings or policies." On Bland and Synthflow, rewrite the source description or search condition in the words callers use.

5. Two answers in the source

The website gives the clinic's Saturday hours as 10 to 2 on one page and 10 to 4 on another, or last year's fee sheet is still uploaded next to this year's. The search retrieves whichever scores higher that turn, so the agent is right on one call and wrong on the next. Synthflow's docs put it plainly: with conflicting documents, the answer "may be inconsistent and depend on which document it selects".

Fix: settle every contradiction before the knowledge base is loaded, with the business if needed, and keep exactly one version of the file on the platform. Delete the old one when you upload the new one.

6. The agent went beyond its source

What happened to us: the agent promised a caller that details would follow on WhatsApp. The business had no such process, and nothing in the knowledge base said it did. The model was being helpful.

This one is the prompt's job, not the knowledge base's. The facts were fine; the agent added an action.

Fix: prompt lines that fence the agent in:

Answer questions about the business only from the knowledge base. If it doesn't have the answer, say you will check, and give the business's phone number.
Whenever you give a price, say the name of the product or service with it.
If the caller hasn't said which product or service they mean, ask before answering.
Don't promise messages, callbacks or bookings unless the knowledge base says the business offers them.

7. The fact is out of date

An offer that ended last month, last season's prices, a doctor who has moved. The knowledge base was right when it was written. Websites change without anyone telling the person who built the agent.

Fix: mark anything dated in the knowledge base as something to confirm (an offer with no end date should say so), rebuild the knowledge base whenever the business changes its website, and before busy seasons.

Test for all seven before a client does

Before an agent goes live, make ten calls that aim at these causes:

1. Ask the price of every product by name, one after another.
2. Ask a follow-up price question without repeating the name.
3. Ask about two products with similar names.
4. Ask something the website never says. The right answer is "I'll check", not a guess.
5. Ask what's included, and whether anything costs extra.
6. Ask for something the business doesn't do, like a callback or a WhatsApp message.
7. Ask about an offer or a seasonal price.

Then read what was retrieved on every turn, not just what the agent said. A right answer from the wrong chunk is luck.

Where KB Builder fits

Causes 1, 2, 5 and 7 are knowledge base problems, and they are the ones KB Builder is built to prevent. It reads a business's website, writes every fact with the name of what it belongs to, writes a sentence for each question callers of that kind of business ask (saying plainly when the website doesn't answer it), shows you contradictions to settle, and flags dated facts to confirm. Causes 3, 4 and 6 are in your hands on the platform; our platform setup guides cover each one.