Practical guide
AI pipeline forecast verification checklist
A pipeline forecast is ready for use only when a reviewer can rebuild its totals, trace material classifications to source records, see every manual adjustment, and understand what would change the planning decision. This checklist verifies a completed forecast. It does not replace the work of defining forecast rules, maintaining the source pipeline, or exercising commercial judgment. Use AUDIT.
A pipeline forecast is ready for use only when a reviewer can rebuild its totals, trace material classifications to source records, see every manual adjustment, and understand what would change the planning decision. This checklist verifies a completed forecast. It does not replace the work of defining forecast rules, maintaining the source pipeline, or exercising commercial judgment. Use AUDIT after the forecast has been prepared: anchor the reporting boundary, uncover record defects, duplicate the calculations, inspect the narrative and approvals, and track residual uncertainty. Copy the checklist into the review record and mark every item pass, fail, not applicable, or unknown.
A: Anchor the reporting boundary and definitions
Confirm the forecast period, source cut-off, currency, included pipelines, owner hierarchy, and intended decision. Obtain the category definitions that were in force at the cut-off. The labels pipeline, best case, commit, omitted, won, and lost are not self-explanatory evidence rules.
Salesforce and Microsoft both document category-based roll-ups, but organisations can configure stages, categories, hierarchies, and adjustments differently. Verify the local configuration rather than assuming a product default applies.
Fail this section if the total cannot be tied to a fixed population. A live dashboard screenshot without a saved cut-off can change while the review is under way.
- [ ] The forecast period, time zone, cut-off, version, owner, and reviewer are recorded.
- [ ] The planning decision and intended readers are named.
- [ ] Included and excluded pipelines, entities, products, and regions are explicit.
- [ ] Category definitions state observable evidence and responsible owners.
- [ ] Currency, tax, discount, renewal, and multi-year treatment are defined.
- [ ] The source export or query can be reproduced at the stated cut-off.
U: Uncover missing, stale, duplicated, and contradictory records
Profile the source population before sampling. Count blank amounts, missing owners, past-due close dates, stale activity, records outside the period, reopened deals, unusual stage-category combinations, and duplicate identifiers or procurement references. Reconcile won records to an authoritative order or finance source where one exists.
Use AI to suggest anomalies only if the reviewer can inspect the rules and source rows. Do not allow an AI summary to overwrite the CRM or silently remove an opportunity.
Record the full denominator. Ten checked rows out of twelve is different from ten checked rows out of twelve thousand.
- [ ] Every included row has an identifier, owner, amount, currency, category, and expected close date.
- [ ] Duplicate and linked records have documented dispositions.
- [ ] Stale dates and activities are reviewed against current buyer evidence.
- [ ] Category and stage combinations follow the approved mapping or carry an exception.
- [ ] Won and lost records reconcile to the relevant authoritative source.
- [ ] AI-flagged anomalies were confirmed or rejected by a named reviewer.
- [ ] Original and corrected values are preserved in the exception log.
D: Duplicate the roll-up, adjustments, and scenarios
Reperform the calculations independently. Start from the saved source population and rebuild category totals. Then reproduce currency conversions, probability or conversion assumptions, timing logic, manual adjustments, and scenario outputs. Check hidden rows, filters, signs, units, rounding, and cross-period leakage.
Do not accept a generated formula because its result looks plausible. Trace its range and operators, then test at least one edge case such as a negative adjustment, multi-currency record, or date exactly at the period boundary.
- [ ] Raw category totals reconcile to source rows.
- [ ] Every conversion rate has a source, population, period, and limitation.
- [ ] Exchange rates have a named source and effective date.
- [ ] Manual adjustments are separated from the raw roll-up.
- [ ] Each adjustment has an owner, reason, evidence, amount, and expiry.
- [ ] Downside, current, and upside scenarios use explicit assumptions.
- [ ] Concentration by account, owner, product, region, and close date is visible.
- [ ] A second calculation reaches the same decision-critical totals.
I: Inspect commentary, changes, and decision conditions
Trace every material sentence in the forecast commentary to a source row, exception, calculation, or documented judgment. Check the movement bridge from the prior version. New pipeline, category movement, amount changes, timing shifts, wins, losses, corrections, and adjustments should explain the change in totals without double counting.
If the narrative says risk fell while the exception log grew, require a reconciliation. If one account drives the decision, name the unresolved event and the contingency rather than hiding it in a general confidence statement.
- [ ] The commentary distinguishes facts, calculations, assumptions, and judgments.
- [ ] Material claims link to the relevant records or analysis.
- [ ] Movement from the prior forecast reconciles to the new total.
- [ ] Buyer commitments are not inferred from seller confidence alone.
- [ ] Missing evidence widens uncertainty rather than being filled optimistically.
- [ ] Planning actions include owners, dates, dependencies, and reversal triggers.
- [ ] The distributed commentary matches the reviewed workbook version.
T: Track approvals, failures, and residual uncertainty
Package the reviewed source extract, definitions, exception log, calculation checks, adjustment ledger, scenario table, movement bridge, commentary, approvals, and next refresh. Preserve failed checks and their repairs rather than rewriting the record to look clean.
The review outcome should be one of four states: release, release with named conditions, repair and re-review, or do not use for the stated decision. State which unknowns remain and who accepts them.
Downloadable checklist artifact
| Check ID | Test | Evidence | Result | Owner | Repair or condition | Retest date | |---|---|---|---|---|---|---| | A1 | Population fixed at cut-off | Saved query and row count | | | | | | U1 | Duplicate records resolved | Duplicate report and dispositions | | | | | | D1 | Raw roll-up reproduced | Independent calculation | | | | | | D2 | Adjustments reconciled | Adjustment ledger | | | | | | I1 | Commentary claims traced | Claim-to-record links | | | | | | T1 | Residual uncertainty accepted | Decision record | | | | |
Add rows for local rules. Do not delete failed rows after repair; append the new result and preserve the earlier disposition.
Worked example: reviewing a synthetic quarterly forecast
A synthetic services team submits a quarterly forecast showing £920,000 committed and £1.6 million best case across 31 open opportunities. The workbook appears complete, and the commentary says the quarter is on plan.
AUDIT finds four decision-changing issues. The source cut-off is not stated, so the reviewer saves a fixed export. Two opportunities share one procurement reference and one is a duplicate. A €180,000 opportunity is treated as GBP. Three committed opportunities have close dates copied from the prior quarter and no buyer action in 45 days.
The independent roll-up and corrected conversion reduce raw commit to £640,000. A separate £55,000 signed change order is documented as a temporary management adjustment, producing a £695,000 adjusted view. The movement bridge shows that most of the reduction is cleanup rather than a new deterioration in buyer activity.
The review outcome is release with conditions. Capacity planning may use the adjusted view, but two contractor starts depend on a named security review being booked by the next refresh. The decision record retains the duplicate, conversion, stale-date, and adjustment evidence so the next reviewer can verify whether the conditions were resolved.
Verification and Two Prune boundary
The checklist passes only when all items have a disposition, decision-critical totals are reproducible, failed checks affect the release decision, and residual uncertainty has a named owner. A sample of favourable rows, an untraceable AI summary, or a commentary-only review is insufficient.
Two Prune can place a participant in a synthetic forecast-review task and capture evidence of source inspection, calculation checking, AI use, contradiction handling, judgment, and communication. It does not forecast revenue, operate a CRM, certify an organisation's forecast process, establish later workplace performance, or make employment decisions.
Primary sources
Salesforce, Pipeline Forecasting Implementation Guide
- Microsoft Learn, Forecast revenue accurately with Dynamics 365 Sales
- Microsoft Learn, View your forecast: check pipeline health and track quota progress
Sources: [1] [2] [3]
Sources
- 1.Salesforce, Pipeline Forecasting Implementation Guide · resources.docs.salesforce.com
- 2.Microsoft Learn, Forecast revenue accurately with Dynamics 365 Sales · learn.microsoft.com
- 3.Microsoft Learn, View your forecast: check pipeline health and track quota progress · learn.microsoft.com