Direct answer: before training lead scoring or feeding “lead quality” back to media, build a minimum layer that preserves six separate facts: intended work, executed attempt, technical delivery, human connection, confirmed disposition, and matured commercial outcome. CRM should not replace every channel; it should receive reconcilable identifiers, timestamps, states, and evidence. Start with one offer, one queue, and no more than three channels. If the system cannot distinguish unserved from served-with-no-interest, keep the outcome unknown.
The failure begins when a completed task becomes proof of contact. A rep marks “called”; telephony says no-answer. Email was submitted but not delivered. A message was sent without the appropriate permission. A meeting existed but nobody attended. Collapsing those records into no response teaches AI logging behavior, channel coverage, and staffing as customer propensity. https://makinai.co/insights/en/ai-lead-scoring-sales-contact-selection-bias explains the selection problem; this article defines the implementation that creates the missing denominator.
Write a Minimum Sales-Service Evidence Contract
- Intent — lead, offer, owner, planned action, deadline, permitted channel, and cadence version.
- Channel event — native ID, event time, direction, raw state, source, and sequence.
- Normalized state — requested, attempted, delivered, connected, completed, or failed; retain the raw value.
- Human confirmation — useful conversation, right person, subject, next step, and supporting field or authorized evidence.
- Disposition — interested, later, no fit, wrong contact, opt-out, duplicate, or unknown; short and versioned.
- Outcome — opportunity, sale, value, loss, or cancellation from the authoritative system after maturity.
- Correction — later event that revises a state while preserving actor, reason, and history.
- Permission — purpose, opt-in, opt-out, and channel restriction; observability never authorizes outreach.
Keep immutable events separate from derived state. Raw webhooks support audit and replay; normalized mappings can change when a provider changes its vocabulary. Do not copy complete message bodies or transcripts by default. Store the minimum evidence, an authorized pointer, and the disposition. For low volume, deterministic integration and a well-designed dropdown may outperform AI classification.
Technical delivery is not a buyer conversation
Twilio distinguishes queued, ringing, in-progress, completed, busy, failed, and no-answer, and notes that completed can include an IVR or voicemail. HubSpot exposes call lifecycle events and engagement creation; Salesforce stores Tasks, call duration, and disposition; Dynamics records activity type, creator, and timestamp. These mechanisms reconstruct service, but none alone proves the correct buyer discussed a need.
Create a channel translation table. An answered phone call is a technical connection until disposition establishes a useful conversation. For email, submitted, delivered, bounced, and replied remain distinct. For messaging, keep sent, delivered, read, inbound reply, and opt-out separate; do not infer intent from a read receipt. For meetings, scheduled, attended, no-show, and completed are different. Never normalize all of them to completed activity.
WhatsApp policy requires the number, opt-in, and respect for opt-out. The evidence layer must block or flag unpermitted work, not merely measure it. https://makinai.co/insights/en/chat-messaging-conversation-crm-next-action-ai covers conversation context and next action; this project focuses on proving that the promised service happened.
Run a six-week pilot
- Week 1 — choose one offer, queue, owner, and decision; inventory CRM, phone, and at most two other channels.
- Week 2 — sample events, activities, and outcomes; define IDs, clocks, states, and evidence gaps.
- Week 3 — implement idempotent capture, lead or account association, translation rules, and an exception queue.
- Week 4 — reconcile channel facts with rep activity; review false contacts, duplicates, out-of-order events, and ambiguous identities.
- Week 5 — operate in shadow mode, measure coverage, and request corrections without changing score, queue, or media.
- Week 6 — release normalized state to one report or transparent rule; only then decide whether it is training data.
The first deliverable is not another model. It is a queryable event ledger, state dictionary, merge rules, retention policy, exception reasons, coverage dashboard, runbook, and named owner for each failure. Apply the monitoring discipline at https://makinai.co/insights/en/monitor-ai-marketing-workflow-crm-api-changes and preserve call-to-revenue linkage from https://makinai.co/insights/en/campaign-phone-calls-ai-crm-verified-revenue.
Budget the integration, not only the license
Cost rises with providers, channels, countries, identities, webhook quality, history and backfill, volume, real-time requirements, retention, human review, seller training, test environments, and support. Compare native CRM capabilities, a light normalization integration, and a custom event ledger with replay. Custom infrastructure pays only when multiple systems, corrections, and auditability justify ongoing operation.
Measure eligible-lead event coverage, activities with native channel IDs, event-to-CRM lag, duplicates, unmatched events, completed tasks without evidence, useful conversations per attempt, disposition completeness, unknown outcomes, late corrections, rep burden, and opt-outs. Then compare score and conversion conditional on observable service. Better coverage does not prove lift; it prevents missing records from becoming false evidence.
Marketing owns the decision and channel; Sales defines useful conversation and disposition; RevOps owns identity and workflow; Technology captures, reprocesses, and monitors; Privacy owns purpose and retention; Finance or the commercial system confirms revenue. An agency or vendor should not independently approve the evidence used to prove its own performance.
If CRM says no response but cannot prove attempt, delivery, and connection, bring MAKINAI at https://makinai.co/services/en/crm-ecommerce-commerce-transformation one offer, an anonymized activity sample, and channel states. We can scope the minimum layer, pilot, and investment before another model purchase. The aggregate October 10–November 6 snapshot recorded 181 sessions, 11 organic clicks, and zero lead events; it cannot attribute demand or revenue to this content.