Post-call summary and CRM auto-update

The useful CRM write is a short summary plus a patch of fields, each tied to a line in the transcript. A Notes paragraph that restates the call is not that.

Gong, Read.ai, and Avoma sit in the same class: they turn a recording into a recap. Salesforce is the system of record. The design choice is whether the recap becomes a suggested patch the AE confirms, or a silent write that treats empty as permission to guess.

What belongs in the patch:

• Call summary. A few sentences for forecast: who was there, what they want to change, what is still unknown. Not the transcript.
• Field updates with cites. Only fields your schema actually has, and only when the transcript supports a value: a pain the buyer named, a competitor they named, a process step they described. Quote enough that a skeptic can find the line.
• Next step. A date and a person, with a cite. Not "follow up."

What does not belong in the write:

• Filling blanks the buyer never answered.
• Copying the whole transcript into Next Steps, Description, or Comments.
• Overwriting a field the AE already typed or corrected.

Speed here is that the AE reviews a draft instead of starting from a blank opportunity. It is not a logging-time number, and it is not a faster guess.

MEDDIC auto-extraction is a separate pass for champion, metrics, economic buyer, and decision criteria. This page is the post-call admin: summary, next step, and the few fields that actually changed on this call.

If the transcript does not support a value, leave the Salesforce field blank. Empty is honest. A guessed close date is not.

The failure mode is silent overwrite. The model sees Close Date empty, hears "we would like this in place this quarter," and writes the last day of the quarter. Pipeline then treats a preference as a date. The AE has to notice the write and undo it, if they notice at all.

Hold three rules:

1. Empty stays empty. No close date, amount, stage move, or next-step date unless the buyer named it, or named a calendar event you can cite.
2. The AE confirms before any overwrite. Suggested values sit in a review queue, a card, or a Salesforce task. They do not land until the AE accepts, field by field.
3. Never overwrite a field the AE already corrected. If the AE changed Close Date after a manager challenge, today's recap must not put the model's date back. Treat AE-owned fields as locked unless the AE re-opens them on this review.

If you cannot cite a line for Amount or Close Date, those fields stay as they were, including blank. "This quarter," "before peak," and "soon" are not dates. "It depends on headcount" is not an amount.

"Follow up" is not a next step. It is a phrase that fills the box so the opportunity looks alive.

A next step needs three parts:

• Who on their side, or a named AE action (send the security packet to Priya Sharma).
• When, as a calendar date the buyer agreed, or as a send-by they named. If they did not name a day, the date stays empty.
• The cite, a short quote or timestamp from the transcript.

If the buyer said "send me the one-pager and I will share it with finance," the next step is "Send one-pager to Jordan Chen; Jordan shares with finance." If they never named a day, do not invent Thursday.

If the AE already typed a next step, the recap may suggest an edit. It does not replace their text until they accept.

Dumping the transcript into Next Steps is the other failure. Managers stop reading. Keep the field to one or two lines.

Buyer promises (they will introduce you, they will run a pilot, they will send the RFP) belong in promised-commitments extractor, not as a blob in the same field.

If the next meeting is with a new function, update buying-committee mapping as a separate write, with a cite. Do not stuff "also map procurement" into next steps.

Run this as draft-then-confirm after every recorded discovery call. Do not start with a silent Salesforce write.

1. Use the transcript as source. Recording consent is a legal and policy question you settle before this workflow, not inside the model.
2. Draft a short summary. Audience is the AE's manager and the AE on Friday. Three to six sentences.
3. Draft suggested Salesforce fields. Only fields in your schema that this call could change. For each, show current value, suggested value, and cite. If current is blank and the transcript is silent, suggested is blank.
4. Lock AE-corrected fields. If the AE edited a field since the last accepted recap, the draft may show the model's suggestion as a note. It must not mark that field as a pending overwrite.
5. AE confirms. Accept, edit, or reject per field. Then write. A rejected suggestion does not come back as an overwrite on the next call unless the transcript has a new cite.
6. Hand the empties forward. Accepted summary and still-blank fields feed pre-call account briefing. Questions this recap shows were never asked belong to discovery gap analyzer, not to a longer Notes dump.

Start with fields that are usually empty after first discovery: next step, next-step date, primary pain, competitor named. Leave Amount, Close Date, and Stage out of the first patch unless your process already requires a human to set them.

Judge the workflow by whether accepted recaps leave those fields either cited or still blank, versus filled with "follow up" or a date nobody said. Do not judge it by minutes the AE used to spend typing.

The following is a made-up first discovery call, not a measured program and not reported results.

An AE is selling a customer-success platform to Maya Patel, VP of Customer Success at a mid-market payments company. Maya says renewals are slipping because health scoring lives in a spreadsheet, they looked at two other tools last year, and they would "love to have something in front of the CRO before the Q4 board." They ask for a one-pager. They do not name a budget, a committee, or a meeting date. The AE types nothing in Salesforce because another call starts in eight minutes.

The recap tool writes without review:

• Notes: the full transcript, pasted.
• Next Steps: "Follow up on one-pager and keep the conversation going."
• Close Date: 15 November (inferred from "Q4 board").
• Amount: the AE's typical deal size for that segment.
• Stage: Propose, because a one-pager was requested.

That is silent overwrite, an invented close date, and a transcript dump in one write. Pipeline now shows a dated, sized proposal for a first conversation.

What the draft should have been:

• Summary. First discovery with Maya Patel, VP of Customer Success. Pain named: renewal risk while health scoring lives in a spreadsheet. They requested a one-pager for the CRO. Budget, decision process, and a dated next meeting were not discussed. Other tools came up as unnamed vendors from last year.
• Next step. Send one-pager to Maya Patel for the CRO. Person: Maya. Date: empty unless they booked a send-by.
• Close Date. Leave empty. "Before the Q4 board" is not a date.
• Amount. Leave empty.
• Stage. Leave as Discover, or whatever the AE already had. A one-pager request is not Propose.
• Competitor. Empty, with a note that unnamed vendors came up. Do not invent names.

The AE confirms that patch between calls. Fields they already corrected on this opportunity stay put. Empty stays empty. The next step is a person and an artifact, not "follow up."

Gong, Read.ai, Avoma, and Salesforce will only be as disciplined as that review step. The speed win is a confirmed patch the AE would have skipped typing. It is not a faster guess.