The short version
Rules-based scoring (points for title, company size, engagement) is a reasonable starting point with no historical data. Predictive scoring, where a model learns from your actual won and lost deals which signals really correlate with closing, is meaningfully more accurate, but it needs enough historical examples to learn from, typically several hundred closed opportunities with consistent CRM data behind them.
Teams that try predictive scoring too early end up with a model trained on noise, which is worse than simple rules because it looks precise while being wrong. Start with transparent scoring, keep your CRM data clean enough to use later, and graduate once you have the volume and history to make it trustworthy.
Building a rules-based score you can test
Choose three to five signals tied to your ICP, such as employee range, target industry, relevant job title, a recent hiring or funding trigger, and a high-intent action like a pricing-page visit. Weight them in whole numbers, for example title 30, size 25, trigger 20, intent 15, industry 10. Simple weights are easier to defend and easier to adjust.
Define tiers with cut-offs, such as A at 70 and above, B at 45 to 69, and C below. Write what each tier triggers: A gets a call within one business day, B enters a sequence, C stays in marketing nurture. A score without a consequence is a number nobody acts on.
Testing the score against outcomes
Store the score and the tier at the moment of lead creation, not as a live value, so you can later ask what scored what at the time. After a quarter, compare conversion from lead to opportunity by tier. If A leads convert at three times the rate of C leads, the model works. If rates are flat, the weights are noise.
Look at each signal separately too. Compute win rate for deals with and without each signal. A signal that does not separate wins from losses should lose weight or go. Do this with a spreadsheet or a Power BI report; no data science is required, only discipline about logging the inputs consistently.
Knowing when to graduate to a model
Three conditions should hold: at least a few hundred closed opportunities, consistent field definitions across those records, and a rules-based score that has stopped improving. Before training anything, check class balance and leakage. Fields populated after the outcome, such as close reason, must never be features, or the model will look brilliant and fail live.
When you graduate, keep the rules score as a baseline and compare the model against it on a holdout period. Adopt the model only if it beats the rules meaningfully, for instance in conversion of its top decile. Require explanations, such as the top three contributing signals per lead, so a manager can still justify a ranking.
Worked example (illustrative)
A hypothetical team scores 1,200 new leads in a quarter with a five-signal rules model and reviews outcomes by tier.
What changed: The tier gap shows the signals carry information, and the team can now shift effort to the leads that convert. After the first review, one signal that did not separate outcomes was dropped.
Do it: the steps
Tick them off as you go. Your progress stays on this device.
Traps to avoid
- ×
Skipping straight to a black-box score because it sounds modern, then being unable to explain a single ranking to a sales manager.
- ×
Never recalibrating. Markets and offers change, so weights from two years ago can quietly punish your best new segment.
- ×
Counting activity as fit. Email opens and clicks inflate scores for people who are simply curious or are bots.
What to track
- Tier conversion lift
- conversion rate of tier A divided by tier C; a gap of 3x or more suggests a useful score
- Score coverage
- leads with all scoring inputs populated divided by total leads; aim above 90%
- Model calibration
- share of high-score leads that become opportunities compared quarter to quarter; investigate drift beyond a few points
Benchmarks are typical planning ranges, not guarantees. Calibrate to your own data.
Questions people ask
How many closed deals do I really need for predictive scoring?
There is no fixed number, but several hundred closed opportunities with consistent data is a common minimum, and more is better when you have many features. With fewer, a model overfits. Treat the figure as a planning floor and validate on a holdout set.
Should intent signals count as much as fit?
Usually fit sets the ceiling and intent sets the timing. A perfect-fit account with no activity is a target for outbound, whereas a poor-fit account with high activity is a distraction. Many teams use two separate scores, fit and engagement, rather than blending them.