Most of what your CRM asks a rep to type after a call was already said during it: the next step, who was in the room, the buyer's timeline, the competitor they mentioned, the pain in their own words. With a transcript, software can pull those into structured fields and ask the rep to confirm them in a minute, instead of making the rep reconstruct the call from memory at 6pm. Here's how to set it up.
Start with what a conversation can actually tell you
A sales call reliably contains some fields and never contains others. Sort your CRM fields into three buckets before you touch any tooling.
- Fields the conversation states outright. Next step and its date. Who attended and their titles. The timeline the buyer gave. Competitors named. The current tool or process being replaced. These are near-certain to be extractable.
- Fields the conversation implies. Pain, quantified impact, decision criteria, the economic buyer, the decision process. These are the MEDDPICC and SPICED fields, and a good extraction can fill them when the rep asked the question. When the rep didn't ask, the correct value is "not discussed", which is useful data on its own.
- Fields the conversation never contains. Deal amount before pricing has been discussed, close date before a paper process exists, forecast category, lead source, territory. Keep these manual or system-derived. Asking software to guess them produces confident nonsense.
Most CRM hygiene problems come from the third bucket being forced on reps as if it were the first. Sort honestly, and half the fields that are always blank turn out to be fields nobody could have known after a first call.
How extraction works, in plain terms
The transcript is the input. During or right after the call, a model reads it against a schema you define: field name, what it means, allowed values, and an instruction for what to do when the answer isn't there. The output is a set of proposed values with the supporting quote from the transcript attached to each one. The rep sees the proposals, edits or approves them, and the confirmed values write to the CRM record.
Two design choices matter more than the model. First, every proposed value should carry its evidence. "Timeline: end of Q3" with the buyer's sentence next to it is something a rep can check in two seconds and a manager can trust. A bare value is something everyone learns to ignore. Second, the schema should distinguish between "the buyer said there is no budget" and "budget wasn't discussed". Collapse those two and your pipeline reports become fiction again.
Say a rep runs twelve discovery calls a week. Confirming twelve sets of proposals at a minute each is twelve minutes. Typing twelve sets of notes into eight fields each is closer to two hours, done late and done badly. That gap is the business case.
Step 1: cut the schema before you automate it
Automation makes a bad schema worse, faster. Before you extract anything, pull a report of field fill rates over the last two quarters. Any field that is filled on fewer than one record in five is either impossible to know or nobody uses it. Delete it or take it off the required list.
What survives should map to a decision someone makes. Next step drives the pipeline review. Stakeholders drive multithreading. Pain and impact drive the business case. Decision process and timeline drive the forecast. If you can't name the meeting where a field gets looked at, it doesn't belong in the extraction schema.
Step 2: write field definitions a stranger could apply
The model will apply your definitions literally, so write them the way you'd brief a new sales ops hire. "Champion" is not enough. "Champion: a named person at the buyer who has personally stated they want this to happen and has access to the economic buyer. Quote the statement. If nobody meets both conditions, return 'none identified'." Do this for every field.
Picklists need the same care. If the competitor field allows five values and the buyer names a sixth, the rule has to say what happens: add "Other" with the free-text name, or flag it for review. Decide once in the schema rather than letting each rep decide differently.
Step 3: keep the rep in the loop, briefly
Fully automatic writes to the CRM feel efficient and fail in two ways. Reps stop reading the values, so errors compound. And reps lose ownership of the record, which is the thing you were relying on for accuracy in the first place. The better pattern is propose, confirm, write. The confirmation screen shows every field with its evidence, defaults to accept, and takes under a minute.
Make the confirmation happen while the call is fresh. Right after the call ends, or in the fifteen minutes before the next one, the rep still knows whether "end of Q3" meant fiscal or calendar. By Friday they don't.
Step 4: set guardrails for the fields that move money
Some fields change forecasts and compensation: stage, amount, close date. Even if extraction can propose them, don't let it write them without a human. Route proposed changes to those fields through the rep, and if the stage moves more than one step, through the manager too. The same applies to anything that triggers downstream automation, like a handoff to implementation.
Log every automated write with its source quote and timestamp. When a forecast call turns into "why did this deal jump to Commit", the answer should be one click away rather than an argument.
Step 5: inspect five fields, not fifty
Once the pipeline is fed from calls, change how you inspect it. Pick the five fields that predict outcomes for your motion. For most B2B teams that is next step with a date, economic buyer identified, decision process, timeline, and quantified impact. Review those weekly for every deal that moved. Ignore the rest unless something looks wrong.
This is also how you measure whether the automation is working. Fill rate on the five fields should climb within the first month. Accuracy is checked by sampling: pull ten records a week, read the quotes against the values, and count the misses. Set a field-specific accuracy threshold with sales ops, and rewrite the definition for any field that misses it.
Where this goes wrong
- Automating the existing schema. Sixty fields fed by a model produce sixty fields of low-confidence guesses. Cut first.
- Values without evidence. If the rep can't see the quote, they can't check it, and within a month nobody trusts the record.
- Treating "not discussed" as blank. A field left empty because the question wasn't asked is a coaching signal. Store it as its own value.
- Letting extraction write stage and amount. Those fields need a human every time.
- Confirming on Friday. The benefit depends on confirmation happening while the rep still remembers the call.
Where Aircover fits
Aircover extracts structured fields from live transcripts during the call, with recording off by default, and syncs the extracted fields to Salesforce, HubSpot, or ServiceNow along with the notes and next steps. It scores calls against MEDDPICC, SPICED, or a custom rubric, so the qualification fields arrive with their evidence attached. The AI CRM sync page shows the field mapping in detail.