Why your AI agent gives wrong answers, and how to fix it

The first week with an AI agent follows a reliable pattern. It handles ninety per cent of conversations better than expected, and then it says something wrong about your refund policy and the whole thing feels unsafe.

That reaction is correct, and the fix is usually not the one people reach for. Nearly every wrong answer we have investigated falls into three buckets, and only one of them is about the AI.

Bucket one: the document was missing

The agent answers from a knowledge base you upload — FAQs, price lists, policies, product sheets. Ask it something no document covers, and a general model will fill the gap with what is generally true of your industry, which is a plausible sentence about somebody else's business.

This is the most common cause by a distance, and the tell is that the answer is reasonable rather than mad. Nobody's refund window is thirty days. It just usually is.

The fix is a document, not a prompt. Write the answer down in the same place your staff would look for it and upload it.

Bucket two: the document was stale

The second most common. The price list is from March, the shipping partner changed in July, and the PDF nobody updated is now confidently misinforming customers at scale.

This is why it matters that the agent shows the source it used. An answer with its source attached lets you distinguish "the AI is wrong" from "our own document is wrong" in the time it takes to click. Those need entirely different responses — one is a configuration problem, the other is a business one that was quietly costing you before any AI existed.

Capacity scales with the plan on Vibot: one document on Lite, five on Basic, twenty-five on Pro, a hundred on Growth. Which is generally enough, because the businesses that struggle are not short of documents — they have eleven overlapping versions of the same one.

Bucket three: it should have escalated

The third bucket is not a knowledge problem at all. An angry customer, a disputed charge, a negotiation, an edge case with money attached: the correct answer is a person, and the failure is that the agent tried.

Every agent needs a clear boundary and a working handover. When the query is outside the knowledge base, or the customer is plainly unhappy, the conversation should escalate to a human with the full history attached, so nobody has to re-explain anything. Supervisors can also watch live and step in mid-conversation.

An agent that hands over well is trusted with more, which is the opposite of what people assume when they try to make it answer everything.

How to find these before your customers do

Test against your own hard questions. Not "what are your hours". Write down the twenty questions your team dreads and run those. On voice, ring the agent yourself and interrupt it — the Vio agent handles barge-in, and how it recovers from being talked over is more informative than any scripted demo.

Point it at the overflow first. Give the agent the after-hours line or the second queue for a fortnight before it goes near your main number. Real traffic finds gaps that no test list will.

Read the transcripts weekly, and sort by sentiment. Every Vio call comes back with a sentiment score and a goal-met rating, and every Vibot conversation is a thread you can read. Ten minutes on the worst-scored dozen tells you more than a month of dashboard averages.

Feed what you find back into the documents. The loop is the product: a question the agent got wrong becomes a paragraph in the knowledge base, and that question is then answered correctly forever. Most businesses reach a steady state in about three weeks.

The honest limit

None of this makes an AI agent right one hundred per cent of the time, and anyone selling you that number is describing a demo. What it does is make the errors legible — sourced, logged, searchable and therefore fixable — which is considerably more than can be said for the same mistake made by a tired human at 7pm with no record of the call.

If you want to see how it handles your actual FAQs before you commit to anything, send them to us and we will point an agent at them.