Keep spoken amounts and reference numbers accurate by extracting them into separate fields, preserving their source wording and checking them before generating the call summary. Use explicit read-backs during the conversation, attach audio timestamps where available, and leave uncertainty visible. A correctly formatted number can still be wrong.
For a business considering AI chatbots, the useful design decision is which spoken values need confirmation before entering a customer record. This proposed AI automation workflow includes a reusable numeric review runbook.
Define what each number means before extracting it
Start with the records staff actually use. A call might contain an invoice amount, a proposed instalment, an account reference and a requested delivery date. Giving all four a field called “number” leaves their meaning unclear.
For amounts, capture value, currency and purpose separately. A quoted price and an outstanding balance remain different fields even when they match. Do not infer whether tax is included from the total alone.
For references, preserve the complete sequence as text. Leading zeroes, letters and separators may matter. An identifier should never be rounded, added or reformatted as money.
For dates, retain the spoken expression alongside any resolved calendar date. “Next Friday” needs the call date and the speaker’s intended meaning. “The sixth” needs a month and year before it becomes complete.
Use four handling statuses: captured, confirmed, unresolved and superseded. Under this proposed rule, confirmed requires explicit speaker agreement to the complete value or verification against an identified authoritative record. Label the method. An audio check establishes what was spoken; record it separately and leave an otherwise clear, unconfirmed value captured.
Separate transcription from numeric verification
A transcript supplies candidate evidence. Extraction assigns it to fields. Review checks what was said, confirmation establishes agreement or a record match, and the summary communicates the result. Keeping these steps separate helps locate errors.
OpenAI’s speech-to-text documentation describes completed-recording transcription and specialised options for speaker labels or timestamps. Select output that supports your review process; do not assume every transcription result includes timing information. Source: OpenAI speech-to-text
Keep the recording reference and an audio location with each consequential field. A reviewer should find the statement, surrounding context and later corrections. If recordings are split, locations must lead back to the original call. Never invent precise timestamps from an untimed transcript.
Structured Outputs constrains responses to a supplied schema, supporting consistent amount, currency and status fields. That is a formatting safeguard; it does not establish that an extracted value matches the audio. Source: OpenAI Structured Outputs
Generate the narrative from reviewed fields. Otherwise, a summary may reintroduce an earlier value or drop an unresolved warning.
Confirm the parts that are easy to mishear
Use focused read-backs instead of asking whether “everything is correct”. For amounts, state currency, whole units and cents. For references, read characters individually or in agreed groups, explicitly naming zeroes.
In a hypothetical call, ask: “I have one thousand five hundred rand and fifty cents. Is that the amount you mean?” This distinguishes R1 500.50 from shorthand with unclear decimal meaning.
For a hypothetical reference, ask: “Two zeroes, then seven, four, one, nine. Is that correct?” Preserve 007419. Do not replace it with 7419 because the numbers seem equivalent.
For a date, repeat day, month and year, then confirm its role: requested delivery, promised callback or payment intention. Agreement to a date does not establish business acceptance of the requested action.
After a correction, repeat the replacement and seek agreement again. Silence or an interrupted reply is not confirmation. A vague “fine” remains unresolved if its meaning is unclear. For completed recordings, absent confirmation may require follow-up; it cannot be supplied retrospectively by the summariser.
Use this numeric review runbook
This proposed operating rule prepares evidence for a person. It does not authorise payments or account changes. Adapt the field names to your records while retaining the distinction between audio review and confirmation.
Numeric call-summary review runbook
- Identify the call: record the call identifier, recording reference and call date. State any unavailable evidence.
- Capture each field: record its name, purpose, exact spoken wording, candidate value and speaker where identifiable.
- Preserve its format: store references as text; separate amount and currency; retain spoken dates beside resolved dates.
- Locate the evidence: record audio start and end positions for the statement, read-back and corrections. Mark unavailable positions explicitly.
- Check completeness: flag missing currency, unclear cents, incomplete dates, uncertain characters and conflicting values. Never guess missing information.
- Record audio review separately: note who checked the recording, what they heard and whether the candidate matches. Audio review alone does not confer confirmed status.
- Establish confirmation: record explicit speaker agreement to the complete value or verification against an identified authoritative record. Label the method and supporting evidence; otherwise record no confirmation.
- Resolve repetitions and assign status: group mentions of the same field. Mark replaced values superseded, clear unconfirmed values captured, qualifying values confirmed, and missing or conflicting values unresolved. Give unresolved fields a named reviewer and follow-up question.
- Review consequential use: have the responsible person assess evidence and applicable records before approving a consequential change. Confirmation is not action authorisation.
- Write and reconcile: generate the summary from reviewed fields, preserve confirmation methods and unresolved warnings, and check saved records against approved values.
Work through ordinary, missing and conflicting cases
All figures, identifiers and timestamps below are hypothetical. They illustrate expected handling rather than measured performance.
Ordinary case: at 02:10, a caller states an invoice amount of “two thousand three hundred rand and forty cents”. At 02:22, the agent repeats it and the caller explicitly agrees. Record R2 300.40, invoice purpose, both evidence locations and confirmed status with the speaker-agreement method. The summary says the caller confirmed the amount; it does not say payment occurred.
Clear but unconfirmed case: a reviewer hears that same amount clearly, but there is no read-back or authoritative-record check. Record the audio review and retain captured status. The summary says “Caller stated R2 300.40; not confirmed.” The responsible person decides whether further confirmation is needed for the intended use.
Missing case: the recording cuts out before the reference. Leave it empty and unresolved. Staff can retain the conversation summary, but must obtain the identifier before linking the amount to an invoice. A reference from another call is not a substitute.
Ambiguous case: a caller says “fifteen fifty” without specifying currency or cents. Preserve the wording and audio location. Ask whether they mean R1 550, R15.50 or something else. If the call ended, assign follow-up rather than choosing whichever interpretation fits the balance.
Duplicate case: reference 007419 appears three times. Group the mentions under one candidate with separate evidence locations. Repetition is not confirmation. If the recording is processed twice, use the call identifier to detect a potential duplicate note, then have a person check before merging or discarding records.
Correction case: the caller states 007419, then says, “Sorry, the last digit is eight.” Preserve the original as superseded and capture 007418 as the replacement. Review the context and seek complete-value confirmation if absent. Two differing references without an explicit correction remain unresolved.
Keep the saved record faithful to the evidence
Distinguish what the caller stated, what was confirmed and what needs attention. “Caller reported an amount; currency needs confirmation” is useful. Removing the warning makes the record misleading.
Treat saving to a CRM as a separate controlled action. OpenAI function calling allows a model to request application functions; application code executes those requests. Your application must supply the numeric checks and approval rules described here. Source: OpenAI function calling
A custom AI agent can be designed around this process. The custom AI agent workflow guide explains the broader approach. AI agents versus automation helps distinguish interpretation from fixed validation rules.
Before deployment, check current product access, account, plan and regional eligibility. Test representative recordings against human-checked expected fields. Include zeroes, corrections, duplicate processing and missing audio. Count wrong values separately from unresolved ones; completed summaries alone do not demonstrate numeric accuracy.
FAQ: spoken numbers in call summaries
Can a reviewer’s audio check make a number confirmed?
Under this proposed rule, no. It verifies what was spoken. Confirmed status requires explicit speaker agreement to the complete value or verification against an identified authoritative record. Record the method because these establish different things.
What if the caller switches languages while giving a reference?
Preserve the wording and have an appropriate reviewer check the audio. Ask the caller to repeat uncertain characters in a language they understand. Vocabulary hints can support transcription, but must not introduce an expected reference that was never spoken.
Should an unresolved reference stop the entire summary?
Retain useful content and flag the affected field. Hold actions depending on that reference until a person resolves it. An uncertain invoice reference need not prevent recording an unrelated callback request with its own evidence.
If your business needs help designing call summaries with clear numeric evidence and human review, get in touch to discuss the fields and exceptions your staff handle.