Direct answer: use the number that fits the decision, not a universal winner. Preserve the AI baseline or distribution unchanged for planning and learning. Use a human override for an operating commit only when it has an owner, reason, evidence, expiration, and accountable decision right. Show both beside actuals. If the adjustment silently replaces the baseline, the company cannot learn whether the model failed, sales supplied new information, or the target pressured the number.
This is an operating-contract problem, not a philosophical trust contest. The point-in-time history in https://makinai.co/insights/en/ai-pipeline-forecast-crm-snapshots-temporal-leakage prevents evaluation with information that arrived later. Stage evidence in https://makinai.co/insights/en/ai-opportunity-stage-evidence-crm-pipeline separates observable progress from opinion. The missing layer is to preserve human intervention as its own event. MAKINAI calls it the Forecast Override Ledger.
Separate five numbers before the meeting
- Observed pipeline — the eligible opportunity rollup at cutoff, with no hidden judgment.
- AI baseline — value, interval, model version, features, and as-of time; never overwritten.
- Seller submission — commit, best case, or category submitted by the rep, with timestamp.
- Manager override — a delta applied with reason, evidence, scope, authority, and expiration.
- Executive scenario — the number used for cash, capacity, or spend, with the selection rule attached.
The interface may emphasize one number for today's decision, but the database should not collapse the tracks. Salesforce documents that an adjustment does not change the gross rollup. HubSpot separates categories and manual submissions. Microsoft allows adjustments for factors not yet reflected in opportunities. Those capabilities show that baseline and judgment coexist; they do not guarantee organizational learning.
Build the Forecast Override Ledger
- Identity — forecast, model or rule version, as-of cutoff, horizon, region, segment, currency, and hierarchy level.
- Change — original value, adjusted value, absolute and percentage delta, direction, and affected category.
- Authority — author, role, approver, and applicable decision right.
- Reason — controlled code plus short note, such as verbal commitment, legal delay, implementation risk, or macro exception.
- Evidence — dated opportunity, logged email, meeting, document, or signal; never “gut feel” alone.
- Validity — start, expiration, and reversal condition.
- Use — the decision that consumed the override: hiring, cash, quota, campaign, or capacity.
- Outcome — actual, baseline error, adjusted error, and post-close learning.
Do not rely on free text as the taxonomy. Start with eight to twelve reason codes, allow a supporting note, and review the set quarterly. An override without expiration becomes permanent truth; without a consuming decision it becomes bureaucracy; without an actual it becomes storytelling. The product-market-period grain in https://makinai.co/insights/en/opportunity-product-market-period-revenue-grain-ai-forecast still matters. Comparing a global forecast with regional actuals produces a false scorecard.
Decide when an override may change the commit
Require an override when a manager has material, verifiable information that is not yet observable in the CRM and the decision cannot wait. Make it optional when evidence is incomplete but testable. Prohibit it when the rationale is hitting quota, copying last quarter, compensating for an unpopular model, or retroactively changing a cutoff already used. For cash, hiring, or capacity, expose a range and sensitivity; a single point does not represent uncertainty.
Take the $8M versus $11M example. The baseline uses Friday's known state. On Monday, the VP of Sales produces two letters of intent that are not yet logged and adds $3M to best case for ten days. The executive commit may remain $8M until legal review passes, while capacity planning reserves part of the upside. If the letters lapse, the override expires. That is a traceable decision—not an arbitrary average of the numbers.
Run an eight-week shadow pilot
- Week 1 — select one unit, horizon, and executive decision; freeze baseline, categories, and the definition of actual.
- Week 2 — map submissions, current adjustments, hierarchy, currencies, timestamps, and permissions.
- Week 3 — implement an append-only ledger and eight to twelve reason codes without changing the official process.
- Weeks 4–5 — capture baseline, submission, and override in parallel; backtest at least four prior cutoffs.
- Weeks 6–7 — expose the scorecard by author, reason, horizon, and segment; review exceptions without punishing transparency.
- Week 8 — set decision rights, thresholds, evidence SLA, expiration, and criteria to promote or stop the pilot.
Deliverables are a forecast contract, field dictionary, reason taxonomy, versioned table, approval trail, baseline-versus-override dashboard, RACI, runbook, tests, retention policy, and recorded decision. Cost drivers include CRM count, hierarchies, currencies, submission cadence, history, model versioning, integrations, identity, permissions, attachments, audit, and meeting operations. License price is usually smaller than semantic reconciliation.
Measure whether judgment added value
- Error, bias, and calibration by horizon, segment, and probability band.
- Override value-add: baseline error minus adjusted error, retaining sign and magnitude.
- Adjustment frequency, direction, and size by author and reason code.
- Share of overrides with evidence, expiration, and reconciled outcome.
- Time from new evidence to submission, decision, and CRM update.
- Reversed decisions, idle-capacity cost, and cash exposure—not only statistical accuracy.
- Recurrence: if a reason repeats, convert it into data or a feature instead of depending on permanent manual intervention.
Fildes and colleagues study the common practice of applying judgment to a computerized initial forecast. Goodwin recommends limiting adjustments, requiring written justification, and assessing intervention outcomes. That does not prove every override helps or every model wins. Compare the frozen baseline and adjusted value against the same actual over multiple periods without selecting favorable examples.
Sales owns commit, upside, and supporting evidence. RevOps governs taxonomy, permissions, cutoffs, and versions. Finance defines actuals and economic decisions. Data owns the baseline, interval, quality, and evaluation. Marketing records influence and constraints without claiming revenue alone. Leadership chooses which track funds cash, capacity, and guidance. AI can summarize evidence, suggest reason codes, and detect patterns; it should not rewrite the baseline, promote an override, or optimize a number to target without human approval.
If your board receives one AI forecast, a second number from sales, and a third from finance, bring MAKINAI at https://makinai.co/services/en/data-content-intelligence-systems an anonymized cutoff, the fields, cadence, and one decision that depends on the number. We can scope the ledger, pilot, and decision rights before disagreement becomes another platform purchase.