Voice agents without reliable answers: missing or conflicting sources

Four problems, four decisions

Missing information differs from a tool failure. A policy absent from the knowledge base needs a source. A failed query leaves the result unavailable for that attempt. Conflicting sources require the team to establish the current information.

Fix the source before compensating in the prompt

Check document processing, attachment to the agent and content. Remove or replace outdated versions as part of curation. Writing ‘always answer correctly’ does not fix contradictory policies.

Use focused documents and test a question answered by each selected source. Assign an owner and record policy changes; do not assume the agent automatically identifies the authoritative version.

An incomplete answer can still help

Fictional example: ‘Do you open at 8 or 9?’ The sources disagree. Expected response: ‘I could not confirm the hours reliably. I can point you to the team's channel’. That channel must exist in the approved information.

Do not fill the gap with a probable time. Ask a clarifying question only if it can resolve the uncertainty, such as identifying the branch. Do not request personal data when an outdated source is the actual problem.

# Insufficient information
Use selected sources and tool results to answer.
If information is missing or conflicting, do not guess.
Briefly explain what could not be confirmed.
Ask for clarification only when it can resolve the question.
Provide only an approved continuation channel.

A referral is not a completed action

Explaining where to find help differs from transferring a call or creating a request. Configure and test the necessary tool when the task requires either action. Confirm record creation only after the destination returns success.

Without an integration, provide the available channel without claiming the team received the question. Test transfer destinations and failure behavior before a pilot.

Test absence and recovery

Prepare four questions: one answered by the source, one missing, one conflicting and one outside scope. Add a tool failure and a valid question after an inconclusive answer. Record expected and observed behavior.

After correcting the source, repeat the original question and a control question. Check accuracy, absence of invented facts and usefulness of the next step. This example is not a completed authenticated test.

Which source should prevail when documents disagree?

Accountable staff should define authority, scope and validity. A recent date alone does not establish that one document replaces another: it may concern a different branch or product. Deselect conflicting sources or clarify their conditions before retesting.

At a fictional shop, two documents list opening at 8am and 9am for the same branch. Preserve uncertainty and provide the approved channel until curation resolves the source. “Always answer correctly” does not establish valid hours. After correction, test the original question and another rule in the same document.

Identify the kind of uncertainty before answering

Not finding an answer and finding two incompatible answers are different problems. In the first, required evidence is missing. In the second, material exists but its authority or applicability is unclear. A third situation occurs when the answer exists but depends on a condition the caller has not supplied. Treating all three as “I do not know” wastes information and can lead to unnecessary questions or premature conclusions.

Consider a fictional school. One document says enrollment ends on Friday; another says it remains open until Monday. Before selecting a date, check the class, location, term, and document versions. Different courses may have different rules without an actual contradiction. If both texts refer to the same enrollment and no approved rule establishes precedence, the agent should not choose the more convenient date. It should explain that it could not confirm the deadline for that enrollment and offer the defined next step.

Now consider “Can I take part in the activity?” The policy may require a particular age range. Here the answer depends on a missing fact, and a specific question may resolve the uncertainty: “How old is the participant?” Immediate routing is unnecessary, as is collecting a full name and identity document merely to explain a general rule. The requested information should correspond to the condition that actually changes the answer.

Describe these distinctions in the instructions with examples from the service being offered. Classification should lead to a concrete action: ask for the missing condition, consult the appropriate source, or state that confirmation is unavailable. This makes responses more useful and prevents a confident manner from taking the place of evidence. It also gives reviewers something more precise to evaluate than whether the agent sounded hesitant or decisive during the call.

Define authority and validity outside the conversation

The agent should not invent a document hierarchy during a call. The company must decide beforehand which material represents the current rule and which systems represent individual state. A cancellation policy may explain general conditions, while the order API indicates whether a particular purchase has already shipped. Those sources answer different questions. A file's modification date alone does not establish that it supersedes an approved policy.

Prepare a simple source inventory recording owner, subject, audience, validity period, and agents using the material. An internal training file may contain fictional examples that should never be presented as actual sales terms. A branch manual may be correct for one location and incorrect for another. Source selection should reflect that scope. Adding many files to “provide more knowledge” can reduce clarity when versions and exceptions are mixed without identification.

In Tigy, processing and attachment are separate checks. A document must finish processing and be selected among the agent's sources. When updating a policy, verify that the new file is available and the old version has been removed from the relevant selection. Do not assume uploading an updated document automatically changes every agent that used its predecessor. Test a question affected by the change and another whose answer should remain unchanged.

When a rule changes on a future date, state its applicability period in the approved material. A caller may ask about an earlier purchase, and applying a new condition to that older case would be incorrect. If interpreting validity requires analysis the agent cannot perform with available sources, define a routing option. Documentary maintenance resolves some uncertainty before it ever reaches the conversation, and provides a clear owner for corrections when a contradiction is discovered.

Express the limit without abandoning the caller

A useful uncertainty response explains what could not be confirmed and what remains possible. “I do not know” may be truthful, but rarely tells the caller how to proceed. “I found different information about this class's deadline and cannot confirm which applies” defines the uncertainty. Then provide the real channel or procedure for obtaining confirmation. Do not invent a response deadline simply to end the conversation on a more pleasant note.

Separate uncertain content from confirmed content. If documents agree about an activity's location but disagree about its opening time, the agent can provide the location and explain the uncertainty about the time. The entire answer need not become useless. Equally, it should not present two times as though both were valid options. A caller might choose one and later discover that attendance was impossible during that period.

Avoid presenting a hypothesis as a recommendation. “The latest date probably applies” hands the customer an inference the operation has not approved. When a hypothesis helps staff investigate, it may be recorded explicitly as a hypothesis, separate from commitments communicated to the caller. The spoken answer should contain what the source supports, the missing information, and a next step that actually exists.

During longer calls, check whether the caller understood the situation. Someone may interpret “I will check” as an already completed lookup or “I will route this” as immediate human service. Use verbs corresponding to actual state: “I could not verify,” “I recorded the request,” or “I can provide the responsible team's contact channel.” If no additional lookup is configured, do not describe a search the agent cannot perform. Transparency prevents both false promises and frustrating endings while preserving the portions of the request that can still be resolved.

Turn recurring uncertainty into source corrections

Track unanswered questions as maintenance signals rather than treating them only as agent failures. If many customers ask whether a fee is included, the pricing policy may state a price without explaining its components. Repeatedly rewriting the prompt does not fix missing commercial information. The team responsible for pricing should approve the answer, which can then be added to the appropriate document. The agent receives evidence instead of an instruction to sound more conclusive.

Record the question, the reason for uncertainty, and the source involved. Useful categories include missing information, an uncollected condition, incompatible versions, unattached sources, and unavailable lookups. These categories point to different owners. An attachment issue can be corrected in configuration; a policy conflict requires a content owner's decision; an unavailable API requires integration investigation. Combining everything into one rate makes it harder to prioritize meaningful corrections.

After each change, repeat the failed question, a different way of asking it and a similar case that still has no answer. For example, updating the main location’s opening hours does not authorize the agent to give a branch the same hours. These tests check both the correct answer and the boundaries of the new information.

Measure reductions in avoidable uncertainty without classifying every stated limit as a defect. A well-founded decision not to invent information protects callers from misleading guidance. The aim is to close preventable gaps and identify genuine gaps more quickly while preserving the ability to say that something could not be confirmed. Well-maintained sources, integrations with clear states, and instructions distinguishing absence from conflict offer more value than demanding absolute confidence in every response. Reviewers should reward accurate limits as well as successful answers.