What to measure, and what the published numbers already say

Three measures show whether AI CRM automation is working: how often each rep accepts or dismisses what it drafts, per field; how many fields get filled without a person and how many of those survive to close unedited; and whether the evidence-adjusted forecast lands closer to the actual quarter than the weighted roll-up. All three need a baseline taken before anything ships. The published industry numbers describe the problem, not the fix, and each one has a weakness worth knowing before it goes in a board deck.

What do the published numbers say?

• Selling time and quota (Salesforce, 2024). The sixth State of Sales report surveyed 5,500 sales professionals across North America, Latin America, Asia-Pacific and Europe between March 8 and April 18, 2024. Reps reported spending 70% of their time on non-selling tasks, 84% said they missed quota last year, and 67% said they did not expect to hit it this year. The weakness: it is a vendor-run survey, by the company that sells the CRM the admin happens in, and the time split is self-reported.
• Forecast deals won (CSO Insights, 2014 to 2016). CSO Insights put the win rate of forecast deals at 45.9% in 2014 and 46.2% in 2016, from its annual surveys; the 2015 edition drew on more than 1,000 companies. Roughly a coin flip, and flat for years. The weakness: the data is a decade old, and from 2015 CSO Insights belonged to MHI Global, the sales training and consulting firm later called Miller Heiman Group.
• Confidence in the forecast (Gartner, 2020). Gartner's State of Sales Operations Survey found that only 45% of sales leaders and sellers had high confidence in their organization's forecasting accuracy. It is the only figure here not run by a company selling sales software or training. The weakness: it is from February 2020, and no sample size was published.

Which widely quoted numbers did we leave out, and why?

Two figures circulate in almost every forecasting pitch: "79% miss the forecast by more than 10%" and "93% cannot forecast within 5%." Each is credited to a different research firm depending on who is quoting it. We could not retrieve a primary document for either. A number nobody can trace is the opposite of what this series argues for, so both stay out.

Why can't a forecast built on CRM fields see what the buyer said?

Because its inputs are fields a rep chose. Stage, amount, close date and activity counts go in, and the model is only ever as honest as the person filling them in. The Northwind deal shows the gap.

Paper process named and not started. Economic buyer never in a room. A competitor that costs nothing because the buyer already owns it. Scope cut on a call but still priced in the CRM. Each is a dated, quotable fact, and none of them is a feature a stage-field model is built to weigh. That is the argument of part 4, where the leader sees only the rows where the CRM contradicts the buyer.

The honest position is narrower than most vendors' pages. We have not found a platform that forecasts primarily from conversation content today. Gong comes closest: it says it uses 300+ signals to predict deal outcomes with "20% more precision than algorithms based on CRM data." That is a vendor claim against its own baseline, not an independent audit, and it should be quoted that way.

What should you instrument before anything ships?

Two of these come from earlier parts: the correction loop in part 7 and the proposal measures in part 8. Notice what is not on the list: a time-saved number. Self-reported time is the weakest kind of figure on this page, the same kind as the 70% above, and it is the easiest one to round up.

How do you take the baseline?

1. 1
   Sample the record you have

   Pull twenty recently closed deals and check each CRM field against the calls and emails. The share that match what was said is your starting accuracy.
2. 2
   Freeze the forecast

   Record the weighted forecast at the start of the quarter and the actual at the end. That pair is the line any adjusted forecast has to beat.
3. 3
   Time the paperwork

   Measure call to proposal sent from email timestamps for a month of hand-built proposals.
4. 4
   Compare like with like

   Same reps, same quarter length, one change at a time. Either the adjusted line is closer or it is not.

It is the same discipline the outbound series uses to trace a reply-rate drop against a matched control: change one layer, hold the rest still, and believe the comparison rather than the story.

Next in the series

Part 13, Automate the transcription. Keep the judgment., closes the series with the line between what runs unattended and what a person decides, and where a head of sales should start.