AI uses for records
Voice agent promise compliance: catching promises policy forbids
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Test voice agent promise compliance by tagging every commitment the agent speaks, such as a refund, waiver, callback or deadline, at its place in the audio, then checking it against the policy clause in force and the agent's tool calls. A promise passes only if policy allowed it for this caller, the facts were confirmed first and the system of record shows it was delivered.
Agent run
Voice agent promise
Key takeaways
- End-state checks miss spoken promises: an agent can commit to something out loud and change nothing in the database.
- Treat each promise as its own record, anchored to an audio span, a speaker turn and a versioned policy clause.
- Verify promise spans against the audio, because speech recognition can insert phrases that were never said.
- Order matters: many policies allow an offer only after the facts are confirmed or the caller asks.
- A promise is not compliant until the system of record shows it was kept, which needs an outcome window that has closed.
- Score scenarios across repeated runs and noisy audio, not one clean pass.
How do you test whether a voice agent makes promises outside policy?
You test it by extracting every commitment the agent speaks, pinning each one to the audio and the policy clause that governs it, and then checking two things separately: was the agent allowed to make it at that moment, and did the business actually deliver it. Task success and caller satisfaction scores do not answer either question.
Common agent benchmarks show why a separate test is needed. The tau2-bench framework gives each domain a written policy, a set of tools and a set of tasks, and it now runs voice conversations in full duplex. Its documentation says the default score for its airline, retail and telecom domains combines a database end-state check with a check that required strings appear in the agent's messages, while free-text assertions judged by a model are still marked experimental.
That design is sound for task completion, but it leaves a gap for promises. If an agent tells a caller it will waive a charge and never calls a tool, the database is unchanged; when the expected end state is also unchanged, the end-state check passes. The out-of-policy promise lives only in the audio, so the evaluation set has to capture it there.
What counts as a promise in a voice call?
A promise is any statement in which the agent commits the business to a future action, a remedy or an exception for this caller. Information about general policy is not a promise, and neither is a hedge that clearly leaves the decision to someone else, though both still need checking for accuracy.
Label the type first, because each type has a different policy test and a different place where delivery shows up.
| Promise type | Typical utterance | Policy question | Where delivery shows up |
|---|---|---|---|
| Remedy | I'll put a credit on your account for that. | Is the caller eligible, and was the remedy offered only when allowed? | Refund, credit or certificate record in billing |
| Waiver | I've taken the late charge off. | Does the agent have authority to waive, and for which reasons? | Adjustment line on the invoice or ledger |
| Deadline | You'll have the replacement by Friday. | Can the business commit to that date for this service level? | Shipment or work order dates |
| Callback or follow-up | Someone will call you back tomorrow morning. | Is a callback a supported action, and who owns it? | Scheduled task or outbound call log |
| Escalation | A supervisor will approve this for you. | May the agent predict another person's decision? | Transfer record and supervisor disposition |
| Eligibility statement | You qualify for the upgrade. | Was eligibility checked against the record before it was stated? | Account tier, entitlement or order record |
| Exception | I'll make an exception this once. | Is any exception path open to the agent at all? | Exception approval and resulting transaction |
Fields that pin a promise to the recording
The first half of the evaluation set records what was said, by whom and when, precisely enough that a reviewer can replay the moment in seconds. Each row is one promise, not one call; a single call can hold several promises and several corrections.
| Field | What it holds | Why it matters |
|---|---|---|
| call_id and promise_id | Stable identifiers; one call can yield many promises | Lets a promise be scored, corrected and audited on its own |
| audio_ref | File, channel and start and end offsets in milliseconds | A reviewer can replay the exact span instead of trusting a transcript |
| speaker_turn | Turn index and role (agent, caller, transferred human) | Separates the agent's promise from a human's after transfer |
| transcript_span | Verbatim words, recognizer confidence and a human-verified flag | Recognition errors can create or erase a promise |
| promise_type and object | Type from the taxonomy, plus what was promised as a count or duration | Makes eligibility and delivery checkable |
| stated_conditions | Any if-then the agent attached, verbatim | A conditional offer is judged on its condition |
| repair_turn | Later turn where the agent corrected or withdrew the promise, if any | A promptly withdrawn promise is a different case from one left standing |
| overlap_flag | Whether caller and agent spoke over each other during the span | Barge-in can cut a hedge or a condition out of what the caller heard |
Fields that tie the promise to policy and outcome
The second half connects each promise to the rule that governed it and to the record that shows whether it happened. Keep this part separate from the audio fields so that a policy update or a late outcome can be added without touching the transcript work.
| Field | What it holds | Why it matters |
|---|---|---|
| policy_version and clause_id | The policy text in force on the call date and the specific clause | A promise can be compliant under one version and not the next |
| facts_required | The account facts the clause depends on, such as tier, coverage or order status | Eligibility cannot be judged without them |
| facts_checked_before | Tool calls or lookups made before the promise, with timestamps | Shows whether the agent confirmed facts or guessed |
| authorization | Allowed, allowed only on request, needs approval, or never allowed | Turns the clause into a testable rule |
| transfer_context | Whether and when the call moved to a person, and that person's decision | Stops a human's decision being scored as the agent's |
| outcome_record | Identifier and status of the refund, order, task or ticket that should reflect the promise | The only evidence that the promise was kept |
| outcome_window | Days the business normally needs to deliver, and whether that time has passed | An immature outcome is unknown, not broken |
| label and rationale | One label from the scoring scheme, the reviewer and a one-line reason | Makes disagreements reviewable |
Which policy clause was in force when the agent spoke?
Judge every promise against the clause that applied on the call date, quoted from a versioned policy file, and record which facts that clause depends on. Policies that agents follow are usually rules about order and conditions as much as about content.
The tau2-bench airline policy is a clear public example. It tells the agent not to offer compensation proactively unless the user explicitly asks, to confirm the facts before offering it, and to list action details and get an explicit yes before any change to a booking. Each of those clauses produces a different test: one about who raised the topic, one about the order of lookups and offers, and one about confirmation before action.
Write each clause as an authorization value and a list of required facts. If the policy says an agent may offer a goodwill credit only to members of certain tiers after confirming a cancellation, the test for a promised credit needs the caller's tier, the cancellation record and the timestamps of both lookups. Without those, a reviewer can only say the promise sounded plausible.
When is a spoken promise confirmed as both made and kept?
A promise counts as made only once a person has checked its span against the audio, and as kept only once the system of record shows delivery. Neither the transcript nor the end of the call is enough evidence on its own.
Transcripts can invent the very sentence you are scoring. Koenecke and colleagues found that about 1 percent of transcriptions from a widely used speech recognizer contained whole phrases or sentences that were not in the audio. For a promise test that is not a rounding error: an invented commitment creates a false violation, and a dropped condition turns a compliant conditional offer into an apparent breach. Mark each span human-verified before it is labeled.
Then check what happened after the call. A refund promise is kept when the billing system shows the refund, a callback when the outbound call log shows it, and a deadline when the shipment or work order closes on time. Use the record the operations team trusts, such as a Zendesk ticket with its linked order, a Salesforce case with its approval history or the payment processor's refund entry, and note when that record was read. If the normal delivery window has not passed, record the outcome as pending rather than guessing from the caller's tone at the end of the call.
Scoring each promise and each scenario
Give every promise one label, and require all of its evidence fields before the label is final. Then score scenarios across repeated runs, because a single clean pass says little about how an agent behaves on the next thousand calls.
- Scenario rule: a scenario passes only if no out-of-policy promise appears in any of k runs. This mirrors the pass^k idea introduced with tau-bench, which counts a task as reliable only when all k independent trials succeed.
- Audio conditions: run each scenario under clean and realistic audio. tau2-bench's voice mode, for example, has a control preset with no audio effects and a regular preset with background noise, accents and interruptions, and it can save the conversation audio for review.
- Report promise rate and violation rate per hundred calls by promise type, so a rare but costly type such as exceptions is not hidden inside an average.
| Label | Definition | Evidence required | What it tells the team |
|---|---|---|---|
| Compliant and kept | Allowed by the clause, facts confirmed first, delivered within the window | Clause, prior lookups, closed outcome record | Behavior to preserve |
| Compliant, not kept | Allowed and properly made, but the business did not deliver | Clause, prior lookups, outcome record showing failure | An operations or integration fault, not an agent fault |
| Out of policy | The clause did not allow this promise for this caller or at this point | Clause, facts showing ineligibility or wrong order | The core failure this set exists to find |
| Unauthorized but honored | Out of policy, yet the business delivered it anyway | Clause plus an outcome record showing delivery | Hidden cost; policy and practice have drifted apart |
| Withdrawn | Made, then clearly corrected by the agent before the call ended | Repair turn verified in audio | Reportable, scored separately from a standing breach |
Worked example: a delayed-flight caller and a premature offer (illustrative)
This example is illustrative and uses a policy written in the style of the public tau2-bench airline domain: compensation only when the caller asks, only after the facts are confirmed, and only for certain membership tiers or fare types.
Turn 4, caller: my flight was delayed and I want to change it. Turn 5, agent: I'm sorry about that, I'll add a travel certificate to your booking for the trouble. The agent has made a remedy promise before the caller asked for compensation and before any lookup of the reservation or the membership tier. Label: out of policy, on two clauses at once.
Turn 9, after looking up the reservation, the agent finds the caller is in an eligible tier and issues the certificate. The outcome record shows it. The final label is unauthorized but honored, not compliant: the eventual eligibility does not cure the order violation, and the team learns that a generous outcome masked a process failure.
Now two exceptions. If the caller had said in turn 4 that they wanted compensation, the do-not-offer-proactively clause no longer applies, and the same promise after the lookups would be compliant. If the agent had transferred the call and a supervisor promised the certificate, the promise belongs to the human turn and is outside this agent's score. Across a hypothetical 40 runs of this scenario, one premature offer in any run fails the scenario under the all-runs rule.
Where the promise test needs a compliance or legal reviewer
Bring in a compliance or legal reviewer when a promise touches consumer protection rules, regulated disclosures or contract terms, or when the clause leaves the agent discretion that a rubric cannot settle. The evaluation set can find the statement; it cannot decide what the statement means legally.
There is a reason to treat promises seriously. In a CanLII case comment on Moffatt v Air Canada, 2024 BCCRT 149, the tribunal is reported to have found the airline liable in negligent misrepresentation for a statement its website chatbot made about the bereavement fare procedure. One tribunal decision is not a general rule, but it shows why a business cannot assume a model's words carry no weight.
The recordings themselves also need a lawful basis. U.S. federal law does not make interception unlawful for a party to the call or where one party has consented, unless the purpose is criminal or tortious, but many states require consent from every party, and using recordings to build an evaluation set raises its own consent and privacy questions. This is general information, not legal advice; counsel reviews the specific calls, states and notices.
Narrow the scope instead of stretching the rubric when policies are written as open discretion, when calls switch languages, or when outcome records cannot be linked reliably. A smaller set with verified outcomes is more useful than a large one full of pending labels.
Sourcing real calls that carry their outcomes
Simulated callers are good at exercising policy clauses, but only real operations show what callers actually ask for and whether promises were kept. The useful unit is a recording linked to its ticket or case, the policy version in force and the billing or order records that followed.
Teams that need such call sets can ask SourceX, which keeps no stock of recordings and instead looks for U.S. businesses that hold matching ones each time a request comes in. The request sets out scope, volume, language, format, timing and licensing needs, and the supplying company approves the scope and signs its agreement before any records are delivered.
Frequently asked questions
Why doesn't a standard task-success score catch out-of-policy promises?
Task-success checks usually compare the final database state with an expected state and look for required phrases. A promise the agent speaks but never acts on leaves the database unchanged, so it can pass. You need a separate label on the spoken promise itself.
Can I label promises from transcripts alone?
Use transcripts to find candidate promises, then verify each span against the audio. Speech recognition can add phrases that were never said or drop a condition the agent attached. A promise label should rest on a human-verified span.
Is a promise compliant if the customer ended up eligible anyway?
Not if the policy required the facts to be confirmed first or the customer to ask first. Label it unauthorized but honored, so the team sees that the outcome was right by luck and the process was wrong.
How long should I wait before scoring whether a promise was kept?
Wait for the normal delivery window for that promise type, such as the refund posting time or the promised callback date. Until then, mark the outcome pending rather than kept or broken.
Should promises made after a transfer to a human count against the agent?
No. Record the transfer turn and score the human's promise separately. The agent can still be scored on whether it predicted the human's decision before transferring.
Sources
- tau2-bench gives each domain a policy the agent must follow, tools and tasks, and supports voice full-duplex evaluation through realtime audio APIs. Source
- tau2-bench's default reward for its airline, retail and telecom domains combines a database end-state check and a substring check of required statements; natural-language assertions are marked experimental. Source
- The tau2-bench airline policy tells the agent not to offer compensation proactively, to confirm facts before offering it, and to get explicit confirmation before booking changes. Source
- tau2-bench voice mode has a clean control preset and a regular preset with noise, accents and interruptions, and can save conversation audio. Source
- The tau-bench paper evaluates agents that follow domain policies with tools and introduces pass^k to measure reliability across repeated trials. Source
- Koenecke et al. found that about 1 percent of transcriptions from a widely used speech recognizer contained whole phrases or sentences absent from the audio. Source
- A CanLII case comment reports that in Moffatt v Air Canada the tribunal found the airline liable in negligent misrepresentation for a chatbot statement about its bereavement fare procedure. Source
- Under 18 U.S.C. 2511 it is not unlawful for a party to a communication, or a person with one party's prior consent, to intercept it, unless for a criminal or tortious purpose. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.