Win/Loss Prediction for B2B Deals

Stage percentages are a polite fiction

Most B2B CRMs assign probability by stage: discovery 10%, proposal 40%, negotiation 70%. Multiply each deal by its percentage, add them up and you have a pipeline forecast. Everybody in the sales meeting knows it is wrong.

A deal that has sat at 'proposal' for five months is not the same as one that reached proposal last week after three meetings with the finance director. A deal the rep moved to negotiation the day before the quarterly review is not 70% likely either. Stage-based probability ignores the signals salespeople already use in their heads, and invites optimistic stage-hopping at month end.

What a win prediction model looks at

The history of changes matters more than the current snapshot. A model that only sees today's stage and amount is barely better than the stage table. The stage history, close-date changes and activity pattern over time are where the signal lives, which means you need a CRM that keeps field history, or a nightly snapshot of the pipeline.

The data honesty problem

Win/loss prediction exposes CRM habits in a way few other projects do.

• Deals that were lost but left open inflate the pipeline and confuse the model
• Activity only logged by some reps makes 'low activity' mean 'rep does not log'
• Lost reasons picked from a dropdown at random, usually 'price'
• Opportunities created retrospectively after the contract was signed, which look like instant wins
• Stage definitions that changed when a new sales director arrived

None of these block a project outright, but each has to be found and handled. The fifth is especially common: if the meaning of 'proposal' changed eighteen months ago, older data may need mapping or excluding. Our post on AI integration with your CRM covers the plumbing that makes this easier.

An illustrative example

Consider an industrial software vendor with twelve account executives, 400 opportunities closed per year and a sales cycle of four to nine months. Quarterly forecasts built from stage percentages overshoot most quarters, and leadership responds by applying a blanket haircut.

A model trained on two years of closed opportunities, with stage history and activity, might find that deals with only one engaged contact after the demo stage close at a fraction of the rate of multi-threaded ones, and that any deal with two or more close-date pushes rarely closes in the quarter at all. Neither finding is a surprise to good salespeople. What the model adds is applying those lessons consistently to every open deal, every week, without anyone having to argue about it. The details are invented, but these two patterns appear in almost every B2B pipeline we look at.

Using predictions without breaking the sales culture

1. Show the model probability next to the rep's own view, never replacing it, and discuss the gaps in pipeline reviews
2. Explain each score with its main drivers: 'single-threaded, close date pushed twice, no reply in 21 days'
3. Use it for coaching on at-risk deals rather than for judging reps
4. Forecast the quarter from model probabilities alongside the sales team's commit, and track which is closer
5. Keep compensation well away from the score, or reps will learn to feed it what it wants

That last point matters. If logging a meeting raises a deal's score and scores affect reviews, meetings will be logged whether or not they happened. A model that becomes a target stops being a measurement.

When win/loss prediction is not worth it

Businesses closing a few dozen large deals a year have too little data, and each deal deserves a proper qualification conversation instead. Very short transactional sales cycles, where most opportunities close within days, do not need deal-level prediction; a volume forecast is enough. And if CRM usage is patchy, spend the budget on making the CRM worth using first.

Some CRMs now include AI deal scoring. For standard sales processes with good data, they are worth trying before a custom build. A custom model is better when you need data from outside the CRM, such as product trial usage, or want to understand and control the drivers.

How we build one

At SpiderHunts, we start by reconstructing pipeline history: what every open deal looked like each week, so the model trains on the information available at the time rather than hindsight. Then a baseline from stage percentages, a gradient-boosted model, back-testing on recent quarters, and scores with reasons written back into the CRM. Our machine learning team typically delivers a first version in six to eight weeks. It sits naturally alongside machine learning lead scoring, which handles the earlier question of which leads deserve a deal at all.

Frequently asked questions