opsmarshal
Workflow 06 — Deal Execution

Deal Inspection Agent

Scores every open opportunity against MEDDPICC or BANT — your choice of framework — using call transcripts, email threads, and CRM field hygiene as evidence. Flags at-risk deals before pipeline review, and never updates a record without passing the gate.

At a glance

The Deal Inspection Agent scores every open opportunity against MEDDPICC or BANT using call transcripts, email threads and CRM field hygiene as evidence, and never updates a record without passing the governance gate. It is one of fifteen workflows in the Enterprise GTM Agent Stack, an Opsmarshal series on revenue operations.

Owned by
Sales Ops + Frontline Managers
Writes back to
CRM opportunity, Audit log, Pipeline review doc
Held to
≥ 85% — Criterion-evidence precision
Gate
Permission → Evidence → Approval → Audit
Framework: MEDDPICC / BANT Writes to: CRM Opportunity Approval: Sales Manager
Business Pain

Pipeline reviews run on rep memory, not evidence

The status quo

Deal qualification lives in a rep's head and a stale CRM picklist. Managers inspect 40 deals in a 60-minute call, relying on whoever tells the best story. Slipped deals were "surprises" that had been visible in the transcript for six weeks.

What this workflow changes

Every deal carries a framework score computed from primary evidence — what the buyer actually said, what fields actually changed, who is actually multithreaded. Pipeline review starts from the flags, not the anecdotes. Reps argue with evidence, which is the point.

Architecture

Three agents, one gate, one system of record

Evidence Sources
Call transcriptsconversation intelligence
Email threadssales engagement / inbox sync
CRM opportunityfields, stage history, activity
Contact rolesmultithreading graph
Agent Layer
Evidence Extractormaps transcript spans → framework criteria
Framework ScorerMEDDPICC or BANT rubric, per-criterion confidence
Risk Narratorwrites the one-paragraph "why flagged" a manager reads
Governance
Gate: permission → evidence → approval → auditscore updates auto-pass · stage-change suggestions require manager approval
Systems of Record
CRM opportunityscore fields + evidence links written back
Audit logimmutable: actor, evidence hash, verdict
Pipeline review docauto-assembled flag list, ranked by risk × value
DecisionChoiceRationale
Framework is configurable MEDDPICC (default) or BANT Enterprise teams run MEDDPICC; velocity and mid-market teams still run BANT. The rubric is a config object, not hardcoded logic — the same evidence extractor feeds both.
Evidence granularity Transcript spans, not summaries A score of "Champion: weak" links to the exact 40-second span where the champion hedged. Summaries hallucinate; spans are checkable.
Write-back scope Score fields only; stage changes are suggestions The agent never moves a deal stage. It proposes; a human disposes. This is the single design choice that gets enterprise security to yes.
Model requirement Any reasoning model, ≥100k context Vendor-neutral. Transcript batches fit in context; no fine-tuning required. Swap providers without touching the rubric.
Live Demo

Inspect a deal

Pick a framework, pick a sample deal, run the inspection. Watch the gate rail: the score write auto-passes; the stage-change suggestion routes to approval. Every event lands in the audit log.

Deal Inspector

Governance audit log

[ready] agent idle — awaiting inspection run
Eval Metrics

What this workflow is held to

≥ 85%
Criterion-evidence precision
sampled weekly: does the cited span actually support the score?
−30–50%
Slipped-deal surprises
deals that slip without a prior flag, quarter over quarter
< 24h
Score freshness
max lag between new evidence and updated score
100%
Gate coverage
no write reaches CRM without a logged verdict
Operator Control Panel

Someone has to run this thing

The scorer is only trusted while its precision is monitored. This console spec defines who watches what, on what cadence, and where the manual overrides live.

Deal Inspection — Ops Console

Owner: Sales Ops + Frontline Managers

Daily monitors

  • Score freshness — deals with new evidence but stale scores (>24h lag)
  • Pending suggestions — stage-change proposals awaiting manager verdict
  • Low-confidence queue — deals the narrator declined to score, needing manual review

Weekly reviews

  • Precision sample — 20 criterion scores audited against their cited spans (target ≥85%)
  • Manager engagement — flags acted on vs. ignored; an ignored flag stream means the rubric lost trust
  • Framework fit check — deals where size/cycle disagrees with the assigned framework (BANT on enterprise, etc.)

Alert thresholds → who gets paged

  • Unapproved stage write detected → sev-1, RevOps lead, immediately
  • Precision sample <85% two weeks running → rubric owner; scoring paused for that segment
  • Evidence source disconnected (transcript/email sync) → Sales Ops within 1 hour

Manual controls

  • Per-deal exclusion — sensitive deals (legal, exec-sponsored) opt out of scoring
  • Rubric weight editor — versioned config, change-approved like code
  • Kill switch — freezes all write-backs; scores keep computing in shadow mode for later comparison
Failure Modes

How it breaks, and what catches it

Transcript over-trust

A buyer says the right words on a call; the deal still dies. Verbal signals inflate scores. Mitigation: behavioral evidence (stakeholders added, security review started, redlines returned) is weighted above verbal evidence in the rubric config.

Rep gaming

Reps learn what the extractor listens for and coach the language on calls. Mitigation: precision sampling audits scores against outcomes quarterly; criteria weights are recalibrated against actual win/loss, not call content.

Framework mismatch

BANT scoring on a 9-month enterprise cycle produces confident nonsense — budget rarely exists at Stage 1. Mitigation: framework is set per pipeline segment, not per org; the config warns when deal size and framework disagree.

Score without narrative

Managers ignore a number they can't interrogate. Mitigation: no score ships without the Risk Narrator paragraph and clickable evidence spans. If the narrator's confidence is low, the deal is flagged for manual review instead of scored.

Opsmarshal builds workflows like this one inside client stacks — see revenue operations and GTM stack architecture, or book a call to walk through your own.

opsmarshal

Revenue Operations and AI Agentic Workflows for B2B companies in transformation. We build the machine. You run the company.

Services

Company

© 2026 Opsmarshal. All rights reserved. Privacy  ·  Terms