CRM activity signal scoring should answer one operational question: which accepted account event changes the next seller action? It should not convert every open, click, meeting, or field change into a claim that a buyer intends to purchase.
Direct answer: keep account fit, activity evidence, recency, identity confidence, repetition, and action risk visible as separate dimensions. Score only normalized eligible events, apply suppression before routing, and backtest the frozen rule on later outcomes.
Define the decision before the score
A score without an action contract becomes decoration. Choose whether the output creates a research task, raises an account in a daily queue, drafts a message for review, requests manager inspection, or does nothing. Define the owner, expiry, evidence displayed, and maximum queue size.
This page owns CRM activity as an event source. The broader signal scoring framework covers multi-source models, while CRM activity tracking defines the underlying events.
Separate fit, engagement, and timing
HubSpot documents fit, engagement, combined, and deal scores. Its current score-building guide describes configured property and event rules, group and total limits, and stored score properties. Those are useful mechanics; configured points are not observed purchase probabilities.
| Dimension | Question | Example |
|---|---|---|
| Fit | Should this account be eligible? | Segment, region, industry, business model |
| Event strength | How directly does the activity support the play? | Accepted meeting versus passive page view |
| Recency | Is the event still useful? | Observed-at time and type-specific expiry |
| Identity | Is the activity attached to the right person and account? | Authenticated contact versus unresolved domain |
| Saturation | Is this new evidence or repeated noise? | First qualified event versus duplicate retries |
| Risk | What is the consequence of acting incorrectly? | Research task versus buyer-facing message |
Normalize CRM activity evidence
Define an event contract before assigning points: source system, event type, actor, account, related opportunity, source ID, occurred-at time, ingested-at time, eligibility, deduplication key, retention, correction status, and permitted use.
Do not count a connector retry as another activity. Do not merge a marketing email open, seller-sent email, buyer reply, calendar hold, completed meeting, and manually logged note under one “engagement” label. Their evidence and failure modes differ.
Preserve unknown identity. An account-domain event may support account research, but it should not be presented as a named contact’s behavior.
Use a transparent scoring worksheet
A portable starting model is a decision worksheet, not a universal formula. Give each eligible event visible component values and a short reason:
Priority score = fit + event strength + recency + identity + corroboration − saturation − action risk.
Choose local scales and weights only after examining historical cases. Keep component caps so repeated weak activity cannot overpower one missing hard gate. Show the score explanation to the rep and store the rule version. If the model changes, do not compare old and new scores as if they have the same meaning.
A second, often safer representation is a matrix: fit level on one axis and accepted event class on the other. The matrix routes a fixed action and exposes uncertainty without false precision.
Add suppression and abstention rules
Suppression should run before score routing. Exclude customers under a separate motion, open opportunities owned by another rep, competitors, employees, test records, unsubscribed contacts, restricted regions, bounced addresses, recent outreach, and unresolved duplicates according to local policy.
Abstain when identity is ambiguous, the activity is stale, the source clock is missing, contradictory records exist, or the proposed action requires a claim the evidence cannot support. A valid “review required” state is better than forced ranking.
Backtest before routing reps
- Freeze event definitions, weights, expiry, exclusions, and action thresholds.
- Split historical time periods so tuning and evaluation do not use the same outcomes.
- Label identity, relevance, accepted action, meeting state, opportunity outcome, and unknowns.
- Report queue precision, relevant-event recall, acceptance, correction, duplicate, complaint, and action latency by event type.
- Run forward in read-only mode before creating rep tasks.
Use complete denominators and preserve unknown outcomes. A score can improve queue focus even if it cannot establish causal revenue impact; describe only what the design supports.
Where Gangly fits
Gangly repository facts describe recency- and strength-based ranking for supported connected signals while showing the triggering evidence. Treat that as first-party scope. Test Signal Detection on the same labeled CRM events, identity rules, suppression cases, and rep decisions. No universal threshold or outcome lift is asserted here.