A score can organize a sales queue, but it cannot decide on its own whether an opportunity is worthwhile. Missing data, new markets, and unusual requests can score poorly even when they deserve a conversation.
Short answer
Automate validation, permitted enrichment, calculation, and routing. Do not automate final rejection. Keep fit, urgency, and completeness as separate dimensions, and send uncertain cases to a human queue with a deadline.
Minimum lead contract
{
"lead_id": "immutable",
"received_at": "date and time zone",
"source": "declared origin",
"contact_permission": "verifiable state",
"need": "problem in the person's words",
"fit": "0|1|2|unknown",
"urgency": "0|1|2|unknown",
"completeness": "0|1|2",
"owner": "queue|person",
"status": "new|review|contacted|closed"
}
Google Forms supports required fields and format validation; that improves input but does not prove truth. HubSpot documents, in a beta feature requiring opt-in and eligible Professional or Enterprise plans, scoring based on criteria and engagement signals. Salesforce describes qualification through need, budget, authority, and fit. Treat these as capabilities and prompts, not a universal formula.
Routing matrix
| Fit | Urgency | Completeness | Next action |
|---|---|---|---|
| high | high | sufficient | owner plus short deadline |
| high | low | sufficient | follow-up queue |
| low | high | any | human review; do not auto-reject |
| unknown | any | low | ask only for essential information |
| any | unknown | sufficient | exploratory contact or review |
Collapsing everything into one number hides the reason. Preserve all three dimensions and show which answers supported each classification.
Flow without silent rejection
form received
→ validate format, source, and contact permission
→ deduplicate without deleting the original record
→ preserve the need in the person's own words
→ apply versioned criteria
→ fit, urgency, or data unknown?
→ assign queue and deadline
→ human confirms contact, nurture, or closure
→ record reason and outcome
A new rule must not silently rewrite history. Store the criteria version and rescore only for a stated purpose.
AI as a suggestion, not a verdict
A model may summarize text and suggest a category. It must expose the original message, support unknown, and never invent budget, authority, or timing. Stop if requests in another language, short messages, or a new sector are systematically marked weak.
Pilot and evaluation
Create synthetic cases for strong fit, low urgency, missing fields, duplicates, a new segment, another language, and ambiguous text. Include examples that break easy shortcuts. Compare routing with human review without using real prospects in the initial test.
Track time to assignment, leads without owners, human disagreement by dimension, unnecessary information requests, reopened cases, and rejected opportunities later recovered. Conversion also depends on offer, price, channel, and execution; this workflow does not guarantee it.
Connect qualified work to proposal and follow-up automation and use spreadsheet versus CRM to decide where history belongs. Useful qualification explains the next action; it does not reduce people to an opaque score.