The most common complaint about a voice agent is: "it gave wrong information". On investigation the finding is usually surprising — the agent did not give wrong information; it gave old information. The document was updated, but the update never reached the agent.
Knowledge reaches an agent through three paths
Knowing the distinction makes diagnosis easy:
- Information inside the prompt. Things written directly into the agent's instructions: opening hours, forms of address, campaign name. Effective the moment you save.
- Documents in the knowledge base. Uploaded files, pages, price lists. These are processed to become searchable; until they are, the agent sees the old version.
- Live queried data. Stock, availability, order status. Fetched through a tool at request time, so freshness is never an issue.
Freshness problems live almost exclusively in the second group.
Why a processing step exists
A document uploaded to a knowledge base is not read end to end by the agent. It is split into parts, each stored in searchable form; when a question arrives the agent retrieves the relevant parts.
This makes hundreds of pages usable quickly and cheaply. The cost is this: changing the document is not enough, it has to be reprocessed. Without that, the agent keeps finding the old parts.
Symptom: the agent insists on the old answer
If you see this picture, the cause is almost certainly the processing step:
- The agent gives the same old answer every time.
- The answer is consistently wrong — not random.
- When you open the document, the correct information is there.
Answers that vary randomly point to a different problem, usually contradictory rules in the prompt.
Contradiction is more dangerous than absence
If the knowledge base holds two documents on the same topic — the old price list was never deleted and a new one was added — the agent finds both and has no rule for choosing. The result is answers that change from call to call.
This is why deletion matters as much as addition in knowledge base maintenance. Expired campaign documents, old price lists and pages for discontinued products should be removed.
Verifying an update
Testing an agent after an update should be more than checking whether it "works". Three steps that help:
- Ask something specific to the new information. Not "how much is it", but a detail that exists only in the new version. An agent parroting the old answer can pass a generic question.
- Start a new conversation. An ongoing conversation can carry the old answer in its context.
- Ask the same question three times. If the answer is stable, the knowledge has settled; if it varies, there is a contradiction in the knowledge base.
Name a content owner
Even in technically sound setups, freshness problems recur because ownership of content is unclear. It should be written down who updates the knowledge base when a price changes. Otherwise everyone assumes someone else did.
Tomorrow: what happens when a call ends — turning conversations into work automatically.