opsmarshal
Workflow 07 — Deal Execution

Forecast Reconciliation Agent

Runs the forecast twice — once from rep commits, once from deal evidence — and makes the variance the artifact: every gap between what reps say and what signals show gets a written explanation a CRO can interrogate.

At a glance

The Forecast Reconciliation Agent produces two forecasts, one from rep commits and one from deal evidence, and treats the variance between them as the deliverable: every gap gets a written explanation attached to the number. It is one of fifteen workflows in the Enterprise GTM Agent Stack, an Opsmarshal series on revenue operations.

Owned by
RevOps
Writes back to
Forecast snapshots, Variance log, Audit log
Held to
100% — Variances with written drivers
Gate
Permission → Evidence → Approval → Audit
Writes to: forecast snapshots + variance logRead-mostly: never edits rep commitsCadence: weekly + on-demand
Business Pain

Why this exists

The status quo

Forecast calls run on assertion: reps defend numbers, managers apply gut haircuts, and the CRO triangulates vibes. When the quarter misses, nobody can reconstruct which assumption broke.

What this workflow changes

Two forecasts, one variance report. Each gap names its driver — “3 commit deals show no paper-process evidence” — and snapshots are versioned, so misses become diagnosable: either the model was wrong, or the pipeline was.

Architecture

Reference architecture

Inputs
Rep forecast categoriescommit / best case / pipeline, untouched
Deal inspection scoresfrom WF-06, with evidence
Stage velocity + historycohort-based conversion patterns
Agent Layer
Signal Modelevidence-weighted probability per deal, explainable
Reconcilercomputes variance and writes the driver prose
Snapshot Keeperversions everything for post-quarter diagnosis
Governance
Gate: permission → evidence → approval → auditthe agent never edits a rep number · leadership publish requires RevOps review · input freshness printed on every report
Systems of Record
Forecast snapshotsboth views, versioned weekly
Variance logdrivers with evidence links
Audit logpublish approvals + input freshness
DecisionChoiceRationale
Read-mostly by designThe agent never edits rep commitsA system that overwrites rep numbers gets gamed or ignored. Preserving both views keeps the tension visible — and the tension is the product.
Explanations over scoresEvery variance ships with prose and evidence linksA CRO won't act on “model says 0.62.” A CRO will act on “these 4 commit deals carry zero economic-buyer engagement.”
Live Demo

Run a scenario through the gate

Three curated scenarios — one that passes, one that flags for a human, one that the gate blocks. Watch the verdict, the evidence rows, and the audit log.

Forecast Reconciler

Governance audit log

[ready] agent idle — pick a scenario
Operator Control Panel

Someone has to run this thing

Agents don't remove operators — they change what operators watch. Console spec: who owns it, what they check, what pages them, and how they pull the plug.

Forecast Reconciliation Agent — Ops Console

Owner: RevOps

Daily monitors

  • Snapshot integrity — all pipelines captured, no partial snapshots

Weekly reviews

  • Variance report review before every leadership publish
  • Explanation quality sample — do the written drivers hold up?

Alert thresholds → who gets paged

  • Signal-model inputs stale >72h (inspection scores, transcripts) → RevOps
  • Snapshot failure before a forecast call → sev-1

Manual controls

  • Deal-level exclusions — known-anomalous deals removed from the model with a reason
  • Model-weight versioning — changes reviewed like code
  • Kill switch — snapshots continue; publishing halts
Eval Metrics

What this workflow is held to

Both
Model error vs. rep error
tracked per quarter — the honest scoreboard that earns authority
100%
Variances with written drivers
no naked numbers reach leadership
100%
Snapshot completeness
every pipeline, every week, versioned
Failure Modes

How it breaks, and what catches it

Model worship

Leadership treats the signal number as truth; reps stop maintaining CRM because “the model knows.” Mitigation: Both errors are published every quarter. The model earns authority publicly or loses it publicly.

Sandbagging arms race

Reps under-commit to beat the model and bank easy wins. Mitigation: Variance is reported in both directions; chronic under-commit patterns surface by rep.

Stale evidence cascade

A broken transcript sync silently degrades the signal model for weeks. Mitigation: Input freshness is printed on page one of every report — a forecast with stale inputs says so before it says anything else.

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