Decide what the score is for
A scoring system is useful only when it changes a decision: which conversations deserve research, which accounts need a follow-up, or which records should be discarded. Define that decision before assigning points. Keep eligibility rules separate from priority. A duplicate, a prohibited contact, or a conversation unrelated to the product should not become actionable simply because it matches a keyword.
Start with four dimensions that a reviewer can explain: fit, specificity of need, timing, and evidence confidence. Give each a short definition and positive and negative examples. A buyer who states a concrete migration problem may be more useful than someone at a perfect-fit company who shared a general industry article. Unknown information should remain unknown rather than default to a favorable score.
| Dimension | Strong evidence | Weak evidence |
|---|---|---|
| Fit | Workflow and organization match the offer | Only a broad industry match |
| Need | Specific pain or required capability | Category name with no problem |
| Timing | Active project or stated deadline | No sign of an evaluation |
| Confidence | Direct source and clear context | Inferred intent from a fragment |
Choose weights from outcomes, not intuition alone
An initial rubric can use ordinal labels rather than a 100-point formula. Once reviewers have labeled a sample and outcomes have matured, compare scores with qualified opportunities and false positives. Adjust criteria that fail to discriminate. Keep a holdout sample for evaluation, and recheck results by source and segment; a model that works for one community may perform poorly elsewhere.
Record why a record received its priority. A short evidence note should include the source, date, relevant quote, assumptions, and reviewer. This allows sales to choose an appropriate response and makes errors correctable. If automated classification is used, sample its decisions regularly and let reviewers override them with a reason.
Connect thresholds to responsible actions
A high-priority record may trigger human review, not automatic email. Medium priority might become an account research task. Low confidence might need more evidence or no action. Define suppression and privacy rules alongside the threshold. Monitor precision among reviewed records, reviewer agreement, opportunity quality, and negative feedback from the channel. These measures are more meaningful than an unverified “99% precision” headline.
Calibrate the rubric with labeled examples
Gather a sample of past signals that reached an actual review decision. Include successful and unsuccessful examples from each source rather than choosing only obvious wins. Two reviewers should read the full source context and label fit, stated need, timing, and confidence independently. Where they disagree, refine the definitions and keep the disputed examples. This process exposes whether “strong intent” is a shared concept or merely a persuasive phrase.
Only then consider numerical weights. Start with a transparent baseline, such as equal weights, and compare its routing decisions with outcomes. A high score that mainly reflects company fit may route too many curious but inactive accounts. A timing weight that depends on a single phrase may miss slower enterprise evaluations. Adjust one rule at a time and keep a holdout sample so improvements are not judged on the same examples used to tune the rubric.
Define actions for uncertainty
Priority is not the same as permission. A high score should first tell a person which evidence to review. A medium score might prompt account research or a helpful public response. An unclear source should remain in research, even if other dimensions look promising. Add a suppression rule for prohibited channels, duplicate records, and people who have declined contact.
Revisit the rubric every reporting cycle. Compare reviewer agreement, the fraction of routed records that become qualified, missed opportunities found later, and negative feedback. Slice results by channel because the meaning of a signal differs between a technical forum and a professional network. Report the sample size and labeling method with any performance figure. An explainable system earns trust by making mistakes visible and correctable, not by claiming near-perfect precision.
Questions to settle before a score routes a person
What does a score of 80 actually mean?
Nothing by itself. It means something only when the dimensions, weights, source quality, and routing action are defined. In the interactive lab, 80 is arithmetic from four equal-weight sliders. It is deliberately not a probability of purchase or a tested precision measure. A real team should compare scores with independently labeled outcomes and inspect false positives by channel before setting operational thresholds.
Should one weak dimension veto the total?
Sometimes. Evidence confidence is a good example: an apparently urgent statement quoted out of context should trigger verification, not outreach. Eligibility and channel rules can also be hard gates. Define these overrides explicitly so a high aggregate cannot conceal a prohibited action or an unknown identity. Reviewers should see the original source and the reason for every override.
Copy this scorecard specification
Write decision being routed, eligible sources, dimensions and definitions, example evidence for each level, weights, unknown-value handling, hard exclusions, human review step, and feedback fields. Save the first labeled sample before changing weights. On each calibration cycle, report the sample size, reviewer agreement, false positives, qualified outcomes, and what changed in the rubric. Keep the score as a tool for allocating attention rather than a substitute for judgment.