Build a Lead-Scoring Rubric from Lead Data

Workflow area: Sales · Tier: Advanced · Hands-on time: ~45 minutes

Most lead-scoring rubrics are folklore: someone senior once said enterprise leads from webinars are good, so webinars get 20 points. Working from your own outcome data replaces folklore with evidence — and the validation step tells you honestly whether the rubric beats the coin-flip you are currently using.

When to use — and when not to

• Use it when you have at least 150–200 closed leads with outcomeswon, lost, or disqualified, plus a handful of attributes (source, company size, industry, engagement). Below that, patterns are mostly noise; write a judgment-based rubric and revisit in two quarters.
• Use it when leads outnumber follow-up capacityscoring only matters if someone must choose which leads to work first.
• Skip the AI for the statistics if you have a data teamthis playbook is the non-analyst's version; a real logistic regression beats prompt-assisted pattern-finding. Use this to prototype, then hand over.
• Skip fully automated routing on day onescore leads and show the score to humans for a month before any automation acts on it.

The recipe

1. 1.

   Prepare an anonymized export. From your CRM, export closed leads from the last 12 months with: lead source, industry, company size band, job seniority of contact, country/region, engagement signals (demo requested, pricing page, event attended), days to first response, and the outcome. Delete names, emails, phone numbers and company names before the file goes anywhere near an AI tool — the pattern analysis needs attributes, not identities. Split it: the oldest 11 months as the training set, the most recent full month held out for validation.

2. 2.

   Ask for patterns, not a rubric, first. Upload the training set to ChatGPT (with data analysis) or Claude:

   Copy

   This CSV contains anonymized closed sales leads with attributes and
   outcomes (won / lost / disqualified).
   
   Using actual computation on the data (not estimation):
   1. For each attribute, show the win rate per value with the count
      of leads in each group (e.g. "source=webinar: 18% win, n=112").
   2. Flag any group with n < 20 as "insufficient data".
   3. List the 5 attributes where win rates differ most across values.
   4. List anything surprising — attributes with NO relationship to
      outcome that people might assume matter.
   
   Do not build a scoring model yet. Show the tables.
   

   Spot-check two or three of the computed rates yourself with a pivot table or COUNTIFS in Excel. If the model's numbers don't reconcile with yours, stop — it is estimating instead of computing.

3. 3.

   Draft the rubric from the verified patterns.

   Copy

   Based only on the win-rate tables we verified above, propose a
   lead-scoring rubric:
   
   - 4 to 6 criteria maximum, each tied to an attribute with a real
     win-rate difference and n >= 20.
   - Points per criterion proportional to the strength of the signal.
   - Total score 0-100, with suggested bands: A (route immediately),
     B (standard queue), C (nurture).
   - For each criterion, cite the win-rate evidence it is based on.
   - Explicitly list attributes you EXCLUDED and why.
   

   A good rubric line reads like evidence, not opinion:

   Copy

   Demo requested: +30 pts (win rate 34% vs 9% baseline, n=87)
   Company size 200-1000: +15 pts (22% vs 11%, n=140)
   Source = purchased list: -10 pts (3% win, n=95)
   

4. 4.

   Screen for proxy bias before anything else. Ask: "Could any criterion act as a proxy for a protected characteristic or unfairly penalize a legitimate segment?" Region and seniority scores can encode bias; a criterion like "responded within one day" may just measure your team's behavior, not the lead's quality. Cut anything you cannot defend out loud to the sales team.

5. 5.

   Validate against the holdout month. Score every lead in the held-out month with the rubric — a simple Excel formula, no AI needed — then compare: what share of that month's actual wins scored in band A? What share of band-A leads actually won versus band C? A useful rubric concentrates wins meaningfully in A. If A and C win at similar rates, the rubric is decoration; go back to step 2 or accept that your current attributes don't predict outcomes.

6. 6.

   Implement in Excel with the rules visible. One sheet: criteria, points, evidence note, and the SUMIFS-style scoring formula. Reviewers should be able to see why a lead scored 75 — a rubric no one can inspect is a rubric no one will trust or correct.

Make it yours

• Recalibrate quarterly. Rerun steps 2 and 5 with a rolling window every quarter; sources decay and markets shift. Log rubric versions with dates so pipeline comparisons stay honest.
• Negative scoring for disqualifiers. If certain segments are consistently disqualified (wrong geography, students, competitors), encode them as instant band-C rules rather than point deductions — cleaner and easier to audit.
• CRM formula field. Once the rubric survives a month of human-visible scoring, implement it as a formula field in your CRM so scores appear at lead creation — still routing to humans, just with the score attached.

Quality check

• Every criterion cites a verified win-rate difference with n >= 20no folklore points.
• You reconciled at least three computed rates manuallythe model computed, you confirmed.
• The holdout month shows real separation between bandsmeasured, not assumed.
• No criterion survives that you couldn't defend as fairproxy-bias screen done and documented.
• The rubric is inspectableanyone in sales can see the criteria, points and evidence in one sheet.