Conversation Intelligence Integration: A Reconciliation Test

Conversation intelligence integration is the controlled path from an eligible call or meeting to the correct conversation artifact, CRM records, relationship edges, fields, workflows, and reports. A connector badge proves availability, not identity accuracy, safe write behavior, recovery, or fitness for your sales process.

This canonical owns integration validation, not product ranking or transcript benchmarking. Use the call-coaching software guide to select a category and the Gong accuracy test for a vendor-specific transcript and summary evaluation. Public documentation cited here establishes described capabilities only. Gangly did not operate or benchmark those products.

Define the capture-to-CRM boundary

Draw the pipeline before testing it: calendar or telephony event → capture eligibility → participant identity → recording and transcript → speaker and artifact extraction → account/contact/opportunity association → CRM activity or custom object → field proposal or write → workflow/reporting consumption. Each arrow has a separate owner and failure mode.

Freeze vendor and plan, CRM and edition, conferencing and telephony sources, calendar/mailbox connection, workspaces, capture rules, integration user, OAuth scopes, object schema, duplicate and validation rules, association configuration, fields, prompts or trackers, retention, region, timestamps, and observation window. Hash or export configuration where possible. A changed mapping, permission, model, or CRM automation creates a new test version.

Document the actual representation. Gong’s current CRM documentation, for example, describes Salesforce tasks/events and an optional custom Conversation object, while Dynamics uses CRM activities. That is a useful warning against assuming the same artifact or behavior across CRMs.

Write the identity, artifact, and authority contract

Authority belongs to a field or event, not to one universal system. Create one matrix row per item with source, stable ID, destination, allowed read/create/append/replace/clear/delete action, trigger, prerequisite, association, reviewer, dedupe key, retry, correction, rollback, audit evidence, retention, and owner.

CRM identity, owner, consent and core commercial state should remain protected unless a named process delegates a bounded action. Begin read-only. Promote a generated value from draft to human-approved write only after evidence and rollback pass. The broader CRM integration guide explains field-level authority design.

Build golden calls, meetings, and CRM records

Use consented test endpoints and synthetic CRM data for risky actions. Cover scheduled and ad-hoc meetings, inbound and outbound calls, answered and missed calls, recurring meetings, reschedules, forwarded invitations, external and internal-only meetings, recording disabled, late joiners, multiple external companies, shared domains, personal email, aliases, and a participant who requests no recording.

Seed CRM states with unique and duplicate contacts, lead plus contact, similar company names, parent/subsidiary accounts, several open opportunities, no opportunity, inactive owner, queue owner, missing contact role, merged record, deleted record, required field, invalid picklist, restricted record, and a protected suppression field. Give every source event, participant, object, edge, expected artifact, field, and prohibited side effect a test ID.

For extracted facts, label speaker, proposition, polarity, amount, currency, date, action, owner, due date, next step, uncertainty, and timestamp. Keep a blind holdout set. Transcript quality and CRM integrity are separate: perfect words attached to the wrong opportunity still fail.

Test account, contact, and opportunity association

Measure every relationship edge explicitly. For each call, list eligible participants and the expected contact, company/account, opportunity/deal, owner, and primary relationship. Include an expected “no safe match” outcome. Test exact email, alternate email, renamed domain, subsidiary domain, several contacts with one address, several active deals, contact without a deal role, owner change, and manual override.

Gong’s official entity-association documentation shows configurable account and deal logic, simulation, and reassociation. Its documented behavior is not a result for your data. Freeze the chosen rule and test it. HubSpot’s Associations API likewise makes relationships and labels explicit; a connector must preserve the intended IDs and association types.

Association precision = correct created edges ÷ all created edges. Association recall = correct created edges ÷ eligible expected edges. Report by object and ambiguity stratum. A wrong deal edge is not rescued by a correct contact edge. Quarantine ambiguity rather than choosing confidently without an approved rule.

Test fields, evidence, append, and overwrite rules

For every mapped field store the source transcript/version, supporting timestamps, proposed value, existing CRM value, data type, allowed values, null behavior, append/overwrite rule, reviewer, final value, write event ID, and before/after state. Test blank, correct, stale, contradictory, manually edited, locked, invalid, too long, multi-select, formula, and deleted fields.

Append and overwrite are business semantics, not formatting preferences. A call note may append as a versioned activity. A next-step proposal may replace a blank draft after approval. Stage, amount, consent, suppression, owner, and close date should not be silently overwritten by an unreviewed summary. If an output lacks evidence or is ambiguous, abstain.

Gong’s data-import documentation illustrates that selected fields, integration-user access, history configuration, and visible import status affect CRM context. Verify equivalent prerequisites for the exact product. Use CRM data quality controls to repair source defects before blaming the connector.

Inject duplicate, retry, order, and lifecycle failures

Prove the final business state after failure. Replay the same source event; duplicate a calendar event; lose acknowledgement before and after a write; deliver transcript after summary; deliver child before parent; return validation and permission errors; throttle requests; expire a token; interrupt a batch; change schema; and recover from an outage. Record attempts, backoff, error/dead-letter visibility, alert, operator action, duplicate effects, recovery, and residual mismatch.

Then reschedule and cancel a meeting, change organizer and owner, deactivate a rep, revoke a scope, merge contacts and accounts, convert a lead, reparent a contact, merge opportunities where supported, delete and restore records, and change association rules. Confirm queued work cannot execute under a deprovisioned identity and that later legitimate edits are not destroyed by recovery.

Gong’s current HubSpot export documentation warns that exporting calls through overlapping paths can create duplicate meetings and describes activity-update boundaries. Treat this as a documented product-specific case, then test all candidates for equivalent overlap. Idempotency means replay produces the same intended business state as one delivery—not merely a successful HTTP response.

Gate recording, permissions, retention, and deletion

Map notices and consent or other approved authority by jurisdiction, channel, participant, purpose, and later use. Test eligible capture, no-record path, pause/resume, private meeting exclusion, bot rejection, recording/transcript access, sharing, download/export, model use, subprocessors, retention, legal hold, deletion, backup/derived artifacts, and audit. This is an operating checklist, not legal advice; use the recording-governance guide with qualified privacy, employment, security, and legal owners.

The UK ICO’s worker-monitoring guidance emphasizes necessity and proportionality for monitoring and discusses workers’ access rights. Apply the rules that govern your facts. Test least privilege for rep, manager, admin, integration user, contractor, and departed employee. A scorecard or transcript should not become hidden employment evidence outside the approved purpose.

Run full-outer reconciliation

Export or query expected and actual manifests at one cutoff. Join by stable source and destination IDs, then classify source-only, destination-only, matched/equal, matched/different, duplicate, missing edge, extra edge, unauthorized value, stale version, late, and unknown. Reconcile artifacts, fields, relationships and history separately; totals can match while every meeting is attached to the wrong deal.

• Artifact reconciliation = correct accepted artifacts ÷ eligible expected artifacts × 100.
• Field fidelity = correct accepted field values ÷ eligible expected field values × 100.
• Association precision = correct created edges ÷ all created edges × 100.
• Association recall = correct created edges ÷ eligible expected edges × 100.
• Duplicate rate = unintended extra artifacts or writes ÷ eligible source events × 100.
• Recovery completeness = affected cases correct after recovery ÷ affected cases × 100.

Report raw counts, strata, exclusions, observation method, source commit time, first observation, accepted time, and error class. Reconcile conversation-derived reports to the same approved event population and cutoff; the reporting automation guide owns metric contracts and dashboard certification.

Worked example—not a vendor result: a fictional run expects 200 artifacts, 700 fields, and 300 relationship edges. It accepts 198 correct artifacts, 686 correct fields, and creates 302 edges, of which 294 are correct. Artifact reconciliation is 198 ÷ 200 = 99.0%; field fidelity is 686 ÷ 700 = 98.0%; association precision is 294 ÷ 302 = 97.35%; association recall is 294 ÷ 300 = 98.0%. If two prohibited writes occurred, the run fails regardless of those percentages.

Set hard acceptance gates

Pre-register gates before vendors or operators see results. Hard-stop on an unapproved recording, lost suppression, private-data exposure, wrong-account critical write, protected-field overwrite, execution under revoked authority, unbounded duplicate creation, invisible failure, unrecoverable deletion, or failed rollback. Set numerical thresholds for ordinary artifact, field, association, latency, recovery, correction effort, and incident measures based on business harm—not a universal benchmark.

Require a decision packet: frozen configuration, authority matrix, golden manifest, raw outputs, reconciliation queries, formulas, failures, exceptions, privacy/security approvals, gate results, rollback evidence, limitations, and named owner. A pass applies only to the tested configuration, population, permissions, period, and workflow.

Pilot, roll back, and calculate TCO

After synthetic tests pass, canary one team, limited call types, one CRM authority, and bounded fields. Keep consequential outputs draft-only. Run daily independent reconciliation, preserve a last-safe timestamp and change journal, monitor corrections and incidents, and include normal owner changes plus one controlled failure. Expand one dimension at a time.

Rollback means stop capture or writes in an approved order, preserve evidence, export the canary delta, compare-and-set reverse only integration-created changes, restore prior authority, validate suppression and relationships, revoke credentials where required, communicate, and monitor. Do not restore a whole CRM snapshot over legitimate later work.

Term TCO = licenses and minimums + capture/transcription/storage + CRM/API/connectors + implementation + identity cleanup + security/privacy/legal + golden labeling/testing + administration/monitoring/reconciliation + rep/manager correction time + incidents/remediation + parallel run + support + export/decommission. Compare cost per correctly reconciled eligible interaction, not cost per recorded hour. Use current quotes and low/base/high volumes; this guide publishes no vendor price or ROI claim.

Print the integration test checklist

A trustworthy conversation-intelligence integration is not the one that writes the most data. It is the one that captures only eligible interactions, attaches evidence to the correct records, respects field authority, exposes failure, reconciles completely, and can be stopped and reversed without damaging legitimate CRM work.