How to forecast a sales pipeline with AI

Start with a stage-evidence contract

Define the evidence required to enter and leave every stage before inspecting the current pipeline. For example, “proposal” might require a named buyer, a confirmed problem, a dated decision process, and a proposal actually delivered. A label without its exit evidence is not a reliable forecasting input.

Export opportunities with owner, amount, stage, stage-entered date, expected close date, last meaningful activity, next step, and next-step date. Add historical outcomes by stage and age where the sample is relevant. Ask AI to flag missing fields and contradictions, but keep the source records beside every flag.

Use a hygiene table:

| Check | Rule | Forecast treatment | |---|---|---| | Stage evidence | Required artifacts are present | Keep current stage | | Stale activity | No meaningful activity within the agreed window | Move to downside or exclude | | Close date | Dated buyer event supports timing | Keep in period | | Amount | Scope and commercial basis are recorded | Keep amount; otherwise mark uncertain | | Duplicate | Account, need, and buying event match another record | Resolve before aggregation |

Build opportunity-level scenarios

Do not ask a model to invent a probability from the opportunity description. Use relevant historical ranges when available, and show when history is sparse. Create base, downside, and upside cases by changing explicit assumptions: stage conversion, timing slippage, and amount at risk.

This synthetic example illustrates the method:

| Opportunity | Evidence condition | Amount | Base treatment | Downside treatment | |---|---|---:|---|---| | North | Buyer confirmed a 30 September decision | $40k | In-period at observed late-stage range | Slip one month | | Harbor | Proposal sent, but no next meeting | $25k | Apply stale-record discount | Exclude | | Cedar | Discovery complete; scope unresolved | $60k | Outside committed view | Exclude | | Vale | Contract edits active with dated owner | $30k | In-period at contract-stage range | Reduce for timing risk |

The output should preserve opportunity-level reasoning. A $95k base case is only useful if a reviewer can see which rows and assumptions produce it. Separate “commit,” “base,” and “upside” if the organization uses those terms, and define them rather than assuming shared meaning.

Test timing and sensitivity

Run at least three checks. First, re-age every opportunity using the date it entered its current stage. Second, shift close dates for opportunities without a buyer-controlled milestone. Third, vary the conversion range for the largest opportunities and show concentration risk.

The UK government’s financial-model guidance emphasizes explicit assumptions, scenarios, sensitivity analysis, and traceable model logic. Its forecasting handbook also describes forecasts as iterative views that should incorporate current information rather than static annual numbers. Those principles support transparent scenario construction here; they do not supply Two Prune with sales benchmarks.

A weak forecast says, “The model predicts $120k because most deals look positive.” A stronger report says, “The base range is $80k–$100k; $30k depends on one contract-stage record, and two proposal-stage records lack dated buyer events.” The second can be challenged and updated.

Verify the deliverable

Before release, confirm:

The evidence artifact is the forecast workbook or table plus a hygiene log, assumption register, and scenario bridge. Preserve these together so the next forecast can explain what changed.

• every included record meets the stated stage-evidence rule;
• stale, duplicated, and undated records have an explicit disposition;
• probabilities come from a named historical cohort or are labeled assumptions;
• scenario changes are visible at opportunity level;
• amounts and timing reconcile to the source export;
• large-deal sensitivity is shown; and
• AI-generated summaries contain no facts absent from the records.

Sources

Financial model essentials

• Budget holder forecasting handbook

Sources: [1] [2]