Practical guide
How to create a pipeline forecast with AI
A pipeline forecast should help a decision-maker see what may happen in a defined period, which records support that view, and what could change it. AI can help reconcile records, surface stale fields, calculate scenarios, and draft commentary. It should not invent buyer commitments, silently change forecast categories, or turn an uncertain opportunity into expected revenue. Use ROLL to build.
A pipeline forecast should help a decision-maker see what may happen in a defined period, which records support that view, and what could change it. AI can help reconcile records, surface stale fields, calculate scenarios, and draft commentary. It should not invent buyer commitments, silently change forecast categories, or turn an uncertain opportunity into expected revenue. Use ROLL to build the artifact: set the rules, reconcile opportunities, layer scenarios, and lock the decision record. The result is a forecast workbook and commentary pack that another reviewer can reproduce from the source pipeline.
R: Set the forecast rules before reading the total
Write the forecast period, reporting cut-off, currency, included pipeline, owner hierarchy, and decision the forecast supports. Define each category in operational terms. A category such as commit should require observable evidence under the organisation's own process, not the confidence of the person preparing the report.
Record the source-of-truth fields for opportunity ID, owner, account, amount, stage, forecast category, expected close date, last material buyer action, next step, and update time. Also record how renewals, usage-based revenue, multi-year agreements, taxes, discounts, and duplicate opportunities are handled. If the source system permits manual adjustments, keep them separate from the unadjusted roll-up.
Salesforce documents pipeline, best case, commit, omitted, and closed as forecast categories associated with opportunity stages. Microsoft Dynamics similarly separates pipeline, best case, committed, omitted, won, and lost views. These product definitions are useful examples, not universal probabilities. The organisation still needs its own evidence rules and data ownership.
O: Reconcile opportunity records before using AI
Export or query the pipeline at the stated cut-off. Check duplicate IDs, blank amounts, inconsistent currencies, close dates outside the period, reopened records, and opportunities with no recent material activity. Reconcile closed-won values to the authoritative finance or order record where that source exists.
For every open opportunity, ask whether its current category matches the defined evidence. A proposal sent months ago does not remain best case merely because the stage was never updated. A verbal signal does not become commit if the local rule requires budget confirmation and an agreed procurement route.
AI can flag anomalies and cluster explanations, but give it stable identifiers and explicit rules. Preserve its output as suggestions. The opportunity owner or designated reviewer must confirm any category, date, amount, or exclusion change in the source system. Record the original and revised values so the forecast does not hide cleanup work.
Use an exception table:
| Opportunity ID | Field at cut-off | Evidence observed | Proposed disposition | Owner | Due date | |---|---|---|---|---|---| | OP-104 | Commit, 30 Nov | Security review not scheduled | Move to best case | Revenue lead | 14 Nov | | OP-118 | Pipeline, 15 Dec | Duplicate of OP-091 | Exclude duplicate | CRM owner | 13 Nov | | OP-126 | Best case, 29 Nov | Amount in USD, report in GBP | Convert with named rate and date | Finance reviewer | 13 Nov |
L: Layer a roll-up, scenarios, and concentration view
Keep at least three numbers distinct: the raw category roll-up, an adjusted management view, and a scenario range. Do not present one generated point estimate as certainty.
The raw roll-up should follow the published rules. The adjusted view should list every manual change, who made it, why, and when it expires. Scenario calculations should state assumptions such as category-specific conversion, timing slippage, foreign-exchange treatment, cancellations, and new pipeline expected during the period. Use historical rates only when the population and time window are relevant, and show when the sample is too small or structurally different.
Add concentration checks. Show how much of commit, best case, and total pipeline depends on the largest account, owner, product, region, or single close date. A forecast may look stable in aggregate while depending on one unresolved procurement step.
AI can generate formulas or scenario tables, but a reviewer should reperform decision-critical calculations. Check units, dates, signs, filters, hidden rows, probability bases, and whether closed-lost or omitted records leaked into the total.
L: Lock the forecast record and explain movement
Publish the forecast with its source cut-off, definitions, version, owner, reviewer, and next refresh. Include a movement bridge from the previous forecast: new opportunities, stage or category movement, amount changes, timing shifts, wins, losses, and corrections. The commentary should explain the few changes that alter the decision rather than narrate every row.
State what remains unknown. Missing buyer evidence, stale CRM fields, a pending legal review, or an unconfirmed exchange rate should reduce confidence or widen a scenario. It should not be filled with an AI-generated assumption. Keep a link from each material commentary claim to the supporting opportunity or exception row.
Worked artifact: a synthetic services forecast
A synthetic consulting team prepares a 60-day forecast from 24 open opportunities. The raw CRM roll-up shows £820,000 in commit and £1.4 million in best case. An AI-assisted anomaly pass identifies four close dates copied from the previous quarter, two records with the same procurement reference, and a USD opportunity included as though its amount were GBP.
The owner review confirms one duplicate, moves two stale commit records to best case, shifts another beyond the period, and converts the USD amount using the finance-approved rate and date. The revised unadjusted commit total is £575,000. Management adds one documented £40,000 adjustment for a signed change order that has not yet reached the CRM, producing a £615,000 adjusted view.
The scenario table shows £430,000 downside, £615,000 current view, and £790,000 upside. It also shows that 48% of the current view depends on one account whose security review is incomplete. The commentary therefore recommends using the current view for capacity planning only with a named contingency: delay two contractor starts if the security review is not booked by the next refresh.
The artifact does not declare the £615,000 outcome likely merely because it is the centre column. It preserves the raw roll-up, exception decisions, adjustment, scenarios, concentration risk, and trigger for changing the operating plan.
Verification and Two Prune boundary
Verification is complete when a second reviewer can rebuild the totals from source records, reproduce currency and scenario calculations, trace manual changes, and explain movement from the previous version. A forecast fails review when category rules are undefined, data cleanup is hidden, an AI suggestion changes authoritative records without confirmation, or missing evidence is treated as favourable.
Two Prune can place a participant in a synthetic forecasting task and capture evidence of source reconciliation, AI use, calculation checking, uncertainty handling, judgment, and communication. It does not predict revenue, operate a CRM, validate a universal forecasting model, establish forecast accuracy, or make employment decisions. Observed assessment evidence remains bounded to the task and available records.
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