How AI Deal Scoring Works (and Why Gut Feel Forecasts Fail)

AI deal scoring is a health score generated from what actually happens in a deal, not from a stage a rep clicked in the CRM. It's built from signals in sales calls, emails, and account activity: who's engaged, what's been said about competitors, whether the timeline is moving or drifting. It exists because gut-feel forecasting, the kind built from a rep's confidence level and a manager's adjustment on top, is wrong often enough that most sales leaders don't trust it.

What Is AI Deal Scoring?

AI deal scoring assigns a deal a health score, usually alongside a win-probability estimate, based on evidence rather than a rep's self-report. Instead of asking a rep how they feel about a deal, it looks at what was actually said and done: has the economic buyer spoken on a call, has a champion been identified and is that person acting like one, has a competitor come up in the last three conversations, has the expected close date moved twice already.

This is different from the stage-based forecasting most CRMs run by default. Stage-based forecasting assumes every deal in "Proposal" is roughly as far along as every other deal in "Proposal." It isn't. One of those deals has a signed mutual action plan and an economic buyer who's asked for a contract redline. The other has a rep who moved the stage forward because the quarter was ending and the deal felt close. Stage tells you where a rep says a deal is. A deal health score tells you what's actually happened in it.

Why Do Gut-Feel Forecasts Fail?

Gut-feel forecasting fails for a structural reason, not a discipline problem: everyone who touches the number has an incentive to distort it. Reps sandbag, deliberately understating a deal's likelihood so a later "surprise" win looks good against a lower bar. Reps also suffer from what sales ops teams call happy ears, reading a buyer's polite interest as commitment. Managers then apply their own adjustment on top, usually a gut discount based on how the rep's last few quarters went, which has nothing to do with this specific deal.

None of this is hypothetical. In a 2020 Gartner survey, fewer than half of sales leaders and sellers said they had high confidence in their organization's forecasting accuracy. That's not a stat about a badly run sales team. It's the median. Benchmarking compiled across sales-ops research since then generally puts manually submitted, gut-feel forecasts in the range of 30 to 40 percent variance from actual results, against roughly 8 to 15 percent for forecasts built from signal-based or AI-derived scoring. Treat those specific figures as an industry-compiled estimate rather than a single peer-reviewed study, but the direction holds across every source we checked: forecasts built from evidence beat forecasts built from feeling, and the gap isn't small.

What Signals Actually Go Into a Deal Health Score?

A deal health score is only as good as the signals feeding it, and the useful ones map closely to what MEDDPICC already asks a rep to qualify by hand:

None of this is a new idea. It's the same criteria a good sales manager has always coached reps to track by hand. The difference is that a score built from call transcripts and account activity catches drift in real time, deal by deal, instead of once a quarter when a manager finally sits in on a call.

Why Can't a Manually-Entered CRM Stage Do This?

Because a manually entered field only reflects what a rep chose to type in, after the call ended, when they remembered to update it. Proponent's own analysis of customer CRMs found the average CRM record is only 35% complete. That's not an external research figure. It's what we've seen looking at our own customers' data, so treat it as an observation about the accounts we work with rather than a universal industry statistic. The mechanism behind it isn't specific to us, though: a field a human has to remember to fill in, after the fact, with no evidence attached, degrades every quarter a team runs on it. A deal health score built from the call itself doesn't have that failure mode, because there's no step where a rep has to remember to log it.

How Is a Deal Health Score Actually Built?

The honest version: a good deal health score is built by scoring the call and account activity directly, not by asking the rep to summarize it afterward. That means analyzing the transcript for the signals above, who spoke, what was said about the timeline, whether a competitor's name came up, combining that with CRM and email activity, and producing a score that updates as new calls happen rather than sitting frozen at whatever a rep last typed. (For a look at how this plays out against a specific competitor's approach, see our Gong vs Proponent comparison.)

This is also why a deal health score should come with the evidence behind it, not just a number. A score with no explanation is just a fancier gut feel. A score that says Economic Buyer engagement dropped because no one from the champion's leadership has been on a call in three weeks gives a manager something to act on, not just something to distrust.

If you want to see what a deal health score built from your own calls looks like, book a demo.