The Transcript Says $500,000. Who Said It?

Anna with Oppy

A human reviewer examines redacted call notes beside a transcript excerpt, with a restrained overlay connecting a dollar amount to its speaker.

A buyer consultation contains this exchange:

Agent: “Would $500,000 be comfortable?”
Buyer: “No. We need to stay below $450,000.”

Now imagine the follow-up brief says: Budget: $500,000.

The number appeared in the call. It just belonged to a question, not the buyer's answer. A transcript can preserve the words while a summary assigns them the wrong authority. Apparently, even a question can get promoted.

This is a fictional example, but it identifies a specific review task: before treating a statement as customer information, establish who supplied it and what they were doing when they said it.

The economics make precision worth discussing. In its November 17 mortgage-production report, the Mortgage Bankers Association reported a $30 average pretax production loss per originated loan in Q3 2026, versus a $973 profit in Q2. The production sample covered 335 reporting companies, mostly independent mortgage companies. This is production economics, not total company profitability, the cost of attribution mistakes or a promised AI saving. It is a reason to measure useful work, not admire tidy summaries.

Searchable does not mean attributable

Oppy's Meeting Intelligence walkthrough describes searchable meeting transcripts saved in Conversations. Its HubSpot update describes extracting budget, timeline and preferences from Oppy conversations, alongside contact updates and transcript attachment.

Those capabilities make a useful question urgent: when a person relies on an extracted budget, can they trace it to the customer's actual statement?

The public documentation does not establish speaker-diarization accuracy or a native attribution gate. This article proposes a separate, human-reviewed Attribution Clerk role. It is not a shipped feature, a transcript correction service or a mechanism for intercepting CRM synchronization. CRM means customer relationship management software.

The clerk's job is narrower: produce a draft evidence brief from an authorized excerpt. A human decides what, if anything, to change elsewhere.

The input frontier is expanding. Google's November 18 Gemini API release notes announced direct Microsoft Office and OpenDocument processing alongside controls for selected document portions. New formats do not settle who made a statement. Nor does file selection establish privacy compliance. This workflow assumes no Oppy connection to that API.

Review the speaker, then the statement

A speaker label is not proof of identity. “Speaker 2” needs a human-confirmed mapping before the clerk treats that person's words as the buyer's.

Even a correctly identified buyer can describe someone else's position: “My brother thinks we should spend $500,000.” That is a third-party report, not necessarily the buyer's budget. “If we sold our condo first, maybe $500,000” is conditional. Neither should become an unqualified customer assertion.

Keep two questions separate: Who said it? What kind of statement was it? Only then write the draft claim. This is about attribution, not deciding whether a buyer approved a search amendment.

Set up a draft-only review role

Oppy positions itself as a platform for launching and managing dedicated AI agents, called oppies. Its creation walkthrough documents role templates and uploaded knowledge. Use those building blocks for approved review instructions, not as evidence of an automatic routing workflow.

Start with a human preparing a redacted excerpt with stable turn IDs. Include the relevant question, answer, qualification and correction, not just the sentence containing the number. Supply a speaker map confirmed by someone who can establish the identities. Do not ask the clerk to infer identity from names, voices or context.

Confirm appropriate access controls before supplying call material. Oppy's Functions guide documents authorized email and phone entries for Admin-only actions. Those invocation permissions do not make an entire conversation private. A prompt is also not an access control.

In an AI-security video clip published by NVIDIA on November 18, Amanda Grady said:

“Big believer that you cannot have security through obscurity.”

That is a security observation, not a transcript-accuracy result. Here, make the access boundary and the evidence inspectable. Keep personal data out of the review unless its use is specifically authorized.

Keep the review separate from existing integrations. The human supplies the excerpt and receives the draft. Any later system update needs human authorization and a verified, configured connection or manual entry.

Copy-ready Attribution Clerk prompt

ROLE
You are an internal Attribution Clerk. Review supplied text;
return a draft evidence brief, not verified customer facts.

REQUIRED INPUTS
case_id: pseudonymous case identifier
source_id: identifier for this excerpt
source_type: transcript or human-authored notes
review_owner: named human responsible for resolution
speaker_map: turn labels, roles, and human confirmation status
turns: ordered turn_id, speaker_label, exact supplied text
target_fields: fields to review, such as budget or timeline

ALLOWED ACTIONS
Read only supplied inputs. Identify candidate claims about
customers. Classify speakers and statement types. Flag gaps.
Draft a brief for the review_owner. No external actions.

FOR EACH CANDIDATE RETURN
field; candidate_claim; source_id; supporting_turn_ids;
exact_quote; relevant_context_turn_ids; speaker_label;
identity_status: human-confirmed or unresolved;
statement_type: assertion, question, hypothetical,
third-party report, negation, correction, or ambiguous;
attribution_status: customer-attributed or unresolved;
qualification_or_contradiction; proposed_draft_wording;
human_review_needed; precise_question_for_review_owner.

EVIDENCE RULES
Every candidate must cite exact supplied text and turn IDs.
Customer-attributed requires a human-confirmed customer mapping.
It does not establish truth or transcript accuracy.
Do not turn staff questions into customer preferences.
Preserve negation and conditions. Do not merge different speakers.
Retain an earlier statement beside a correction, identifying
which wording the correction addresses. Never silently erase it.
Do not invent timestamps, confidence scores or missing words.

STOPPING AND ESCALATION
Stop review if authorization is absent or inputs contain
unapproved sensitive data. Request a redacted, authorized input.
If IDs or a review_owner are missing, request them first.
For uncertain identity, missing context or conflicting turns,
leave the affected claim unresolved and ask the review_owner
one specific question. Continue only with unaffected claims.
Never guess an answer to clear an exception.

PROHIBITIONS
No CRM writes, search changes, customer messages, third-party
sharing, credit decisions, legal conclusions or final clearance.

ACCEPTANCE TESTS
1. Staff asks $500,000; buyer says below $450,000:
   do not attribute $500,000 as the buyer's budget.
2. Buyer reports brother's $500,000 suggestion:
   preserve third-party-report classification.
3. Speaker mapping unconfirmed: attribution stays unresolved.
4. Buyer states a condition: preserve it in draft wording.
5. Missing evidence: flag the gap; never manufacture a quote.

Test the attribution, not the confidence

Before live use, have a human label a small redacted test set containing these traps. Record the number of candidate claims reviewed, how many were assigned to the wrong speaker and how many lacked supporting turns. Report the counts with their denominator. Separately inspect unresolved claims so caution does not become a substitute for useful work.

Passing these tests is a local acceptance check, not proof of universal accuracy. The review owner still checks the source before relying on a consequential claim.

The clerk should finish with evidence a person can inspect, not a more confident version of the original summary. A number is easy to extract. Knowing whose number it is takes a better job description.

References

  1. Oppy Meeting Intelligence walkthrough
  2. Oppy HubSpot product update
  3. Oppy platform overview
  4. Oppy creation walkthrough
  5. Oppy Functions and permissions guide
  6. Mortgage Bankers Association Q3 2026 production results, November 17
  7. Google Gemini API release notes, November 18 entry
  8. NVIDIA's November 18 AI-security clip featuring Amanda Grady