Grounding and Verification: Reducing Hallucinations in Voice Agents

Large language models are fluent by design. They produce the most plausible continuation of a conversation, and plausible is not the same as true. When a model generates something that sounds right but is not supported by any real source, we call it a hallucination. In a chat interface, a user might notice and ask again. On a phone call, the answer is spoken once, in a calm and confident voice, and the caller acts on it.

For businesses deploying voice agents, reducing hallucinations is not a model quality curiosity. It is the difference between an agent that customers trust and one that creates complaints, refunds and reputational damage. This post sets out the principles that matter most: grounding, verification, constraint and monitoring.

Why Hallucinations Cost More in Voice

Several properties of spoken conversation raise the stakes compared with text.

• No visual re-check. A chat user can scroll up, copy an order number, or compare what the bot wrote with a web page. A caller has only memory, and often a noisy environment.
• Spoken authority. A natural, fluent voice carries social weight. People tend to extend to a confident speaker the trust they would extend to a human representative.
• Real actions. Voice agents frequently book appointments, place orders, register complaints or collect payment details. A hallucinated confirmation ("your refund has been processed") can lead a customer to stop following up on a problem that was never resolved.
• Fewer chances to correct. Phone conversations move forward quickly. If an error is not caught in the moment, it is rarely caught at all.

There is also legal exposure. In the widely reported 2024 case Moffatt v. Air Canada, a Canadian tribunal held the airline responsible for incorrect information its website chatbot gave about bereavement fares. Whatever the jurisdiction, businesses should assume that what their automated agents say may be treated as what the business said.

A Taxonomy of Hallucinations

Not all hallucinations are the same, and different types call for different defences.

Intrinsic and Extrinsic

Research on natural language generation commonly distinguishes two categories. An intrinsic hallucination contradicts the source material the model was given: the price list says ₹499 and the agent says ₹399. An extrinsic hallucination adds information that cannot be verified from the source at all: the agent claims a store opens at 9 am when no opening hours were provided. Extrinsic errors are often harder to spot because nothing in the source directly contradicts them.

Factual Hallucinations

These are invented or distorted facts about the business or the world: product specifications, branch locations, stock availability, eligibility criteria.

Policy Hallucinations

The agent invents or bends a rule: "you can return it within 30 days" when the policy says 7, or "there is no cancellation charge" when there is. Policy hallucinations are especially damaging because customers reasonably rely on them.

Tool-Result Misreporting

When an agent calls a system, such as an order lookup or a booking service, it must accurately relay what came back. A model may say "your order is out for delivery" when the tool returned an error, or summarise a partial result as complete. This category is under-discussed and, in action-taking agents, among the most consequential.

Conversational Hallucinations

The agent misremembers what the caller said earlier: the wrong date, a different quantity, a name from a previous turn attributed to the wrong person.

Grounding in Business Data

Grounding means anchoring the agent's statements to authoritative sources rather than to the model's general training knowledge. A general-purpose model knows a great deal about the world and nothing reliable about your particular business: your prices, your service areas, your holiday schedule.

Principles of good grounding include:

• Treat the business's own data as the source of truth. Catalogues, policies, FAQs and live system data should be what the agent draws on for business-specific claims.
• Keep sources current. A grounded agent with stale data is still wrong. Ownership of each knowledge source, and a clear process for updating it, matters as much as the technology.
• Prefer answering from the source over paraphrasing loosely. The further a response drifts from the wording of a policy, the more room there is for meaning to shift.
• Make absence explicit. If the source does not contain an answer, the correct behaviour is to say so, not to fill the gap with something plausible.

The Value of "I Am Not Sure"

A well-designed voice agent needs a clearly defined, low-friction path for uncertainty. Phrases such as "I do not have that information right now, let me connect you to someone who can help" or "Main confirm karke aapko callback arrange karwa deta hoon" are not failures. They are the system working correctly.

Models tend towards answering because their training rewards helpful, complete responses. That tendency has to be counterbalanced by making the honest alternative both permitted and practical: a human handoff, a callback request, a message taken for follow-up. If the only way to end a turn gracefully is to produce an answer, the model will produce one.

Businesses should also measure this path. An agent that never says "I am not sure" is not necessarily accurate; it may simply be overconfident.

Read-Back Confirmation

Some of the most useful ideas for voice agents come from fields that learned the cost of spoken errors long ago.

In aviation, pilots read back air traffic control clearances, such as altitudes, headings and runway assignments, so that the controller can catch any misunderstanding before it matters. In healthcare, many hospitals require staff receiving verbal or telephone orders, particularly for medications, to read the order back to the prescriber for confirmation. In both cases, the principle is the same: when a spoken detail is critical and errors are costly, repeat it and get explicit confirmation.

For voice agents, read-back is valuable for:

• Numbers: phone numbers, pin codes, account or policy numbers, amounts. "Aapka number nau aath saat shunya..." is easy to mishear, for both the agent and the caller.
• Addresses and names: especially locality names, landmarks and spellings.
• Orders and bookings: items, quantities, dates, time slots.
• Commitments: anything the caller will act on, such as a callback time.

Read-back serves two purposes at once. It catches speech recognition errors on the caller's side, and it catches reasoning or transcription errors on the agent's side before any action is taken. The cost is a few seconds of conversation; the benefit is avoiding a wrong delivery or a missed appointment. Good read-back is selective, reserved for details that truly matter, so that the conversation does not become tediously repetitive.

Structured Tool Outputs Over Free Text

When an agent interacts with business systems, how information flows back into the conversation strongly affects reliability.

• Structured results reduce ambiguity. A result with explicit fields for status, amount and date is harder to misread than a paragraph of prose.
• Errors must be unmistakable. A failed lookup or a timeout should be represented as a clear failure, never as an empty result that a model might interpret as "nothing found" or, worse, as success.
• Actions should be confirmed by the system, not asserted by the model. An agent should only tell a caller that a booking is confirmed if the booking system actually returned a confirmation.
• Identifiers should come from systems. Order numbers, ticket IDs and reference codes should be taken from real tool results, never generated in conversation.

The general principle is to let deterministic systems own facts and state, and to let the language model own the conversation around them.

Constraining What Can Be Promised

Certain statements carry much higher risk than others. A voice agent should be deliberately limited in what it can commit to on the business's behalf.

• Prices and discounts should come only from authoritative sources, and the agent should not negotiate or improvise offers unless explicitly authorised.
• Refunds, cancellations and compensation should be described according to policy and executed only through proper systems, never promised informally.
• Delivery dates and timelines should be stated only when a system provides them, and phrased with appropriate uncertainty when they are estimates.
• Eligibility decisions, for loans, insurance or offers, are often regulated or sensitive and generally should not be implied by an automated agent without clear grounding.

A useful exercise is to list every category of statement that would cause harm if wrong, then decide for each whether the agent may state it freely, state it only from a verified source, or must hand it to a human.

Monitoring and Continuous Review

No combination of techniques eliminates hallucinations, so production monitoring is essential.

• Review real conversations. Sample calls regularly, with extra attention to calls involving commitments, complaints or escalations.
• Check claims against sources. For conversations that referenced policies or data, verify whether what was said matched what was available.
• Track the uncertainty path. Watch how often agents hand off or decline, and whether those decisions were appropriate.
• Close the loop. Every confirmed hallucination is a signal: a missing knowledge source, an unclear policy, an ambiguous tool result or an instruction that needs refinement.
• Listen to customers. Complaints that begin "but the agent told me..." are among the most valuable data a business can collect.

Building Agents People Can Trust

Hallucinations are a property of how language models work, not a bug that will simply disappear with the next model release. Better models reduce the rate, but the responsibility for what an agent says to a customer remains with the business deploying it.

The organisations that succeed with voice agents treat reliability as a design discipline: ground the agent in real business data, make "I am not sure" a first-class answer, read back what matters, let systems of record own facts, limit what can be promised, and keep reviewing real calls. At Cirio, we consider these principles foundational, because a voice agent is only useful if callers can believe what it tells them.