You automate MEDDIC by having an AI model read each call transcript and draft the six fields in your CRM, using a separate prompt for each field. The rep reviews the draft and corrects it. This works because the prospect usually says the qualifying information on the call, and the fields stay empty because nobody has time to type it up afterward.
Why is MEDDIC data usually incomplete?
MEDDIC (Metrics, Economic Buyer, Decision Criteria, Decision Process, Identify Pain, Champion) asks reps to turn a long, unstructured conversation into six structured fields after every call. That admin competes with the follow-up email and the prep for the next call, and it loses. As a rep carrying a quota, I filled in qualification fields last, often from memory at the end of the day, and by then the specifics the prospect had given me were vague.
The information exists. The prospect said what the problem costs them (Metrics), who has to approve the spend (Economic Buyer) and what the product has to do (Decision Criteria). All of it is in the transcript, and very little of it reaches the CRM.
Everything that depends on those fields then runs on partial data: deal scoring, forecasting, pipeline reviews and coaching.
How does AI extract each MEDDIC field?
Give each field its own agent, meaning its own prompt and its own output format. Every agent reads the same transcript and looks for different things.
| Field | What the agent looks for |
|---|---|
| Metrics | Quantified business impact: money, percentages, time lost, headcount affected. If the prospect gave no figures, it flags the gap so the rep can ask next time. |
| Economic Buyer | Who approves the budget, whose name comes up when the rep asks who else has to sign off, and whether that person has been on a call. |
| Decision Criteria | The requirements the prospect named, captured in their words. For example, “it has to integrate with Salesforce” or “it has to be SOC 2 compliant.” |
| Decision Process | The evaluation stages, who is involved at each one, the timeline, and any procurement, security or legal review. |
| Identify Pain | The problems the prospect described, in their own words. Features they liked don’t belong in this field. |
| Champion | Which contact is doing the internal selling: scheduling meetings with other stakeholders, sharing materials, pushing the timeline. |
Champion is the hardest field to automate. The agent can flag a likely champion from engagement patterns, and the rep should make the call. When I review drafts with reps, Champion is the field they correct most often, because the contact who replies fastest can turn out to have little influence inside their company.
The agents run as soon as a transcript is available. When the rep opens the deal after the call, the fields hold draft data to review.
How do you build the extraction workflow?
The workflow has five steps:
- Transcript source. Gong, Fathom and Fireflies each have an API that returns the transcript text when a call completes. Use the recorder you already have.
- Trigger. The recorder sends a webhook when a new transcript is ready. An automation tool such as n8n receives it and starts the workflow.
- Extraction. The automation tool sends the transcript to an AI model such as Claude or ChatGPT, once per field. The Metrics prompt, for example, asks for every quantified business impact the prospect mentioned, with the quote, what it measures and its size, returned in a structured format.
- Write-back. The structured output is mapped to custom properties in your CRM through its API. Only fields where the transcript added something new get updated.
- Notification. The rep gets a Slack message listing which fields changed, with a request to review and confirm.
If you haven’t built a workflow like this before, how to set up your first AI sales agent covers the trigger, context, decision and action structure in more detail.
What should happen when a call has nothing useful in it?
Plenty of calls contain no MEDDIC information. A short check-in won’t cover budget or decision process. Plan for three cases:
- Nothing found. Have each agent return an explicit result such as “No quantified metrics mentioned in this call.” That tells you the call was processed and the topic didn’t come up, which is different from a call the workflow missed.
- Very short calls. Filter these out at the trigger. Log that the call happened and skip extraction, which saves API calls and keeps the fields clean.
- Uncertain extractions. Poor audio, overlapping speakers and long technical tangents all produce unreliable output. Have the prompt mark anything it’s unsure of, and send those items to the rep for review before they’re written to the CRM. A wrong Metric in the CRM does more damage than an empty one. One mistake I made was letting extraction write straight to the CRM with no review step; figures a prospect had mentioned in passing were recorded as firm Metrics, and I added the review step after that.
What changes when the MEDDIC fields are filled in?
- Deal scoring has something to score. A risk model can tell an early deal apart from a late-stage deal that’s missing qualification data.
- Pipeline reviews get specific. A manager can point to the gap (“the Economic Buyer hasn’t been on a call and the deal is in Proposal”) and skip the status questions. That’s a large part of replacing the weekly pipeline review.
- Forecasts rest on qualification data. The forecast can account for what’s known about each deal as well as its stage.
- Coaching gets patterns to work from. A manager can see that one rep rarely captures Metrics and another finds the Economic Buyer late.
The usual caveat about CRM data quality applies here too. Duplicate contacts and deals with no associated calls will limit what the agents can write back.
How do you roll it out if your team barely uses MEDDIC today?
Skip the mandate to fill in the fields. Connect your call recorder to the extraction workflow, let it draft the fields for a couple of weeks, then show the team their own deals with the data in place.
Reviewing a draft built from the prospect’s words is a much smaller ask than writing six fields from memory. Reps start to trust the fields once they see that the drafts are accurate. Before I show a team anything, I listen back to a sample of calls and compare them with the drafts field by field.
Plan for upkeep. The prompts need adjusting when you add a product, change your deal stages or notice reps correcting the same field over and over. Someone should own the workflow and read a sample of its output regularly.