Skip to content

AI uses for records

Business call outcome alignment: a good call, then a cancellation

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Keep the conversation and the outcome as two linked layers and never let one overwrite the other. Preserve what was said, by whom, when and with which corrections; then attach later account events, such as a cancellation with its request date, effective date and stated reason, through a predefined window. Business call outcome alignment means labeling the gap, not hiding it, and never treating the link as proof of cause.

Call to transcript

Business call outcome

This page covers business call outcome alignment: a good call, then a cancellation.Illustrative

Key takeaways

  • Keep the conversation layer and the outcome layer separate, and add a third layer that records how they link.
  • A call disposition describes the call's logistics, not whether the customer stayed.
  • Billing systems separate the cancellation request date from the end date; store both and time the link from the request.
  • Fix outcome events and the window before joining, and label open windows pending and truncated data censored.
  • Exclude or separately label involuntary cancellations such as failed payments.
  • Check transcription and diarization around every labeled signal before trusting the link.

How should a call dataset account for a cancellation after a call that sounded successful?

Record the call as it happened, record the cancellation as it happened, and join them with an explicit rule written before anyone looks at the results. The call keeps its own labels: the issue raised, whether the agent resolved it in the conversation, the disposition the agent logged. The account keeps its own events: the cancellation request, when it takes effect and the reason the customer gave. The join adds a third layer that says how the two relate.

The tempting shortcut is to relabel the call as a failure because the customer left. That throws away the most useful thing the example teaches: a conversation can be locally successful, with the stated problem fixed and a polite close, and still sit one day before a cancellation. A model trained to treat a polite close as success learns the wrong lesson, and a model trained to treat every pre-cancellation call as a failure learns a different wrong lesson. Keep both facts and the time between them.

Why the call's own label and the business outcome disagree

They measure different things at different times. The call label is written by the agent or the phone system at the end of the conversation, and it usually describes the call's logistics. In HubSpot's calls API, for example, the call outcome property records values such as connected, left voicemail, no answer or wrong number, set through internal identifiers. None of those says whether the customer stayed.

The outcome lives in another system with its own timing rules. Stripe's subscription object, as defined in its public OpenAPI specification, separates canceled_at, the date of the cancellation request when a subscription is set to cancel at the end of the period, from ended_at, the date the subscription actually ended, and carries a cancel_at_period_end flag. Its cancellation details distinguish a reason such as cancellation_requested, payment_failed or payment_disputed from customer feedback values such as switched_service, too_expensive or missing_features.

So a sentence like the customer canceled the next day can mean three different things: the request was made the next day, the subscription ended the next day, or a failed payment closed it the next day. Each changes what the call can be said to have touched.

What has to be preserved from the call before any outcome is attached?

Preserve the raw evidence of the conversation first, because every later label depends on it and it cannot be reconstructed. That means the audio and its rights basis, speaker turns with times, a transcript aligned to those times, the moments where speakers corrected themselves or each other, and any transfer to another agent or queue.

At a glance

What has to be preserved from the call before any outcome is attached

Speaker turns come from diarization. The open-source pyannote.audio toolkit, for example, returns turns as start and stop times with a speaker label, which is the right shape to store even when another tool produces it. For the transcript, a time-aligned format such as WebVTT carries cue start and end timestamps and lets each cue name its speaker with a voice span, so a reader can jump from a sentence to the matching audio. Media Fragments' temporal syntax, such as t=10,20 for the interval from second 10 to second 20, gives a stable way to cite a moment in the recording from a label.

  • Audio rights: the recording disclosure that applied, the parties' locations as recorded, and any restriction on reuse.
  • Speaker turns: start, end and speaker role (customer, agent, third party, IVR) for each turn.
  • Transcript alignment: cue-level timestamps tied to the audio file's hash and duration.
  • Repair turns: self-corrections and misheard items, with the turn that repaired them.
  • Transfer context: which queue or person the call came from or went to, and what was handed over.
  • Commitments: promises made in the call (callback, credit, escalation) with their stated deadlines.

The call-to-outcome linkage protocol, step by step

The protocol below turns a call archive and an account event log into linked, gradeable records. Each step names what to store and the mistake it prevents. Window lengths are illustrative and should be fixed per dataset before linkage.

The call-to-outcome linkage protocol, step by step
StepWhat to doFields storedMistake it prevents
1. Freeze the callHash the audio, store duration, recording disclosure and call metadata as loggedcall_id, audio_sha256, duration, disposition_as_logged, rights_basisLater edits to the call record changing the evidence
2. Segment and alignDiarize, transcribe, align cues to audio; keep tool versionsturns[start, end, role], cues[start, end, voice, text], tool_versionsQuotes that cannot be traced to a moment in the audio
3. Mark conversational signalsLabel issue, in-call resolution, commitments, repair turns, sentiment at closeissue_code, resolved_in_call, commitments[], repairs[]Treating the disposition as the conversation's result
4. Define outcomes before joiningWrite the outcome events, the window (for example 30 days) and exclusionsoutcome_spec_version, window_days, excluded_reasonsChoosing the window after seeing which one makes calls look predictive
5. Join on account and timeLink events for the same account that start after the call and inside the windowaccount_key, event_type, requested_at, effective_at, reason, feedbackLinking a sibling account's event or one that predates the call
6. Grade outcome maturityMark each link mature, pending (window still open) or censored (account data ends)maturity_status, observed_untilCounting customers as retained only because time ran out
7. Adjudicate relevanceA reviewer marks whether the event plausibly relates to the call topic, without claiming causerelevance_label, reviewer_role, notePinning involuntary or unrelated churn on the conversation
8. Document the releasePublish the spec, counts per label, tool versions and known gaps with the datasetdatasheet_version, label_counts, known_limitsBuyers mistaking correlation in the set for a causal claim

When is an outcome mature enough to attach, and which cancellations belong to the call?

An outcome is mature when its window has closed with the account still observable; until then it is pending, and it should be stored that way rather than as retained. If the account's data stops before the window ends, the link is censored and needs its own label.

Not every cancellation belongs to the call. A cancellation closed for a failed or disputed payment is a billing event, and it should usually be excluded or labeled separately even if it lands the day after a support call. A cancellation requested by someone who was not on the call, such as a finance lead acting on a budget review, may relate to the account but not to the conversation. Record the requester's role when the source system has it.

Renewals and downgrades need the same treatment as cancellations. A customer who renews on reduced seats after a good call has not simply been retained, and a customer who pauses rather than cancels may return. Decide in the outcome specification which of these events count, how partial changes are coded, and whether a later reversal, such as a cancellation withdrawn before the end date, replaces the earlier event or is kept beside it. Keeping both, with their dates, is usually safer for training and evaluation alike.

Use the request date, not only the end date, for timing. Where a subscription is set to end at period close, the request may arrive long before the end date, and the request is what sits next to the call in time.

Worked example (illustrative): the login fix that preceded a cancellation

This example is illustrative, not taken from a real company. A customer calls a scheduling software vendor about a login loop. The 11-minute call ends with the issue fixed, the customer saying the fix works, and the agent logging the call as connected and resolved. The next morning the customer requests cancellation at period end, choosing switched service as the reason in the cancellation flow.

Reviewing the aligned transcript shows two things the summary missed. At minute 8 the customer mentions they are trialing another tool for their second location. At minute 9 the agent misheard the plan name and the customer corrected it, a repair turn that delayed the fix by about two minutes.

Worked example (illustrative): the login fix that preceded a cancellation
LayerFieldValue (illustrative)Evidence pointer
Conversationresolved_in_callYes: login loop fixed and confirmed by the customerCue at t=610,640, customer voice
ConversationsignalsCompetitor trial mentioned; repair turn on plan nameCues at t=480,505 and t=540,560
Call recorddisposition_as_loggedConnectedPhone system log, unchanged
OutcomeeventCancellation requested day 1, effective at period end day 19Billing record request and end dates
Outcomereason and feedbackRequested by customer; switched serviceBilling record cancellation details
Linkmaturity and relevanceMature at day 30; reviewer: plausibly related, not caused by the login issueReviewer note with role and date

Does the recording support the transcript and the claimed outcome?

Check both before trusting either. For the transcript, score a random sample of calls against careful human reference transcripts; NIST's SCTK toolkit includes sclite, which aligns reference and hypothesis text to score speech recognition output, and its word error figures show whether names, plan terms and numbers survive transcription. Errors cluster on exactly the words that matter for outcomes, so report them for those terms separately where possible.

Diarization errors are the more dangerous ones for this task. If turns are assigned to the wrong speaker, a statement such as we are thinking of switching can be attributed to the agent instead of the customer, and the conversational signal disappears. Spot-check turns around every labeled signal against the audio, and store the reviewer's correction rather than editing the tool output in place.

Keep the scoring sample separate from the calls used to tune the transcription or diarization settings, and record which tool versions produced the released turns and cues. If the tools are upgraded later, re-run the sample and publish the new figures with a new dataset version instead of mixing outputs from two tool versions in one release.

For the outcome, confirm the join with the source system's own record, not a downstream report. A customer success dashboard that marks an account churned may lag, merge accounts, or count downgrades as cancellations.

The link breaks down when one call is part of a long journey, when the decision maker was never on the line, or when account identifiers in the phone system and billing system do not match cleanly. In those cases, narrow the claim: label the call with its conversational signals and mark the outcome as unlinked rather than forcing a match on name or phone number.

Stop and bring in specialist review when the audio's rights basis is unclear, when calls include third parties who were not told about recording, or when the dataset would let someone infer a sensitive fact about an identifiable caller. Whether a given recording can be reused for model training depends on the disclosures given and the laws where the parties were, and counsel reviews those facts. The linkage protocol does not settle consent; it assumes consent and rights were checked first.

Finding call archives that already connect to account outcomes

Most public speech datasets have audio and transcripts but no downstream account events, which is why outcome-aligned call data usually comes from a business's own phone, CRM and billing history. The hard part is the join, and it is easier when the business already keys calls to accounts.

Call archives keyed this way can be requested from SourceX, which sources them from U.S. businesses for each buyer request; the buyer sets out scope, volume, language, format, timing and licensing needs. Delivery waits until the supplying company has approved the scope and terms and signed, and SourceX de-identifies the records through approved processing routes, so account linkage can be kept through consistent pseudonymous keys while callers' identities are removed.

Frequently asked questions

What is business call outcome alignment?

It is the practice of linking a recorded business call to what later happened on the account, such as a renewal, a cancellation or a refund, through a defined time window and matching rule. The call's own labels and the later outcome are kept as separate, linked layers.

Should a call be relabeled as failed if the customer canceled afterward?

No. Keep the in-call result, such as issue resolved, and record the cancellation as a separate outcome with its timing and reason. Relabeling loses the information that the conversation went well locally and the account still left.

What is a call disposition, and is it an outcome?

A disposition is the label an agent or phone system logs at the end of a call, often about logistics such as connected, voicemail or no answer. It describes the call, not the customer's later decision, so it should not be used as the outcome label.

How long should the outcome window be?

Long enough for the outcome you care about to appear, and fixed before linking. Calls whose window has not closed should be labeled pending, and accounts whose data ends early should be labeled censored, rather than counted as retained.

Does linking a call to a cancellation show the call caused it?

No. The link shows timing and, with review, plausible relevance. Involuntary cancellations, decisions made by people who were not on the call and earlier events can all explain a cancellation, so causal claims need a separate study design.

Sources

  • Stripe's OpenAPI specification separates a subscription's cancellation request time (canceled_at) from its end (ended_at) and records a cancellation reason and customer feedback values. Source
  • The WebVTT specification gives each cue start and end timestamps and lets a voice span name the speaker. Source
  • pyannote.audio is an open-source speaker diarization toolkit that outputs speaker turns with start and stop times. Source
  • NIST's SCTK includes sclite, which aligns reference and hypothesis transcripts to score speech recognition output. Source
  • HubSpot's calls API records a call outcome such as connected, left voicemail, no answer or wrong number, set through internal identifiers. Source
  • Media Fragments URI 1.0 defines temporal fragments such as t=10,20 for the interval from second 10 to second 20. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies