Field Notes - Jan 05, '26

Executive Signals

• CRM is the new data warehouse: source of truth beats sidecar spreadsheets and brittle exports
• Inboxes are systems, not pipes: search, labels, and replays create resilience and auditability
• SLOs hate late: tag violations, don't inflate failure rates
• Burn-in before broadcast: contain first-run noise until error stabilizes
• Stage mirrors prod: synthetic jobs catch breaks without spamming customers

Customer Success

Make CRM the Source of Truth for Compliance CCs

Routing lists drifting between a sheet and automation will fail quietly, then loudly. Treat compliance CCs as a first-class CRM concept with a formal contact role and one read view for jobs to consume. Keep the sheet as a temporary guardrail with edit warnings, then retire it.

• Create a “Compliance-CC” contact role and publish one read-only view for automation
• Migrate sheet to CRM within one sprint; enable edit warnings until cutover
• Kill CSV exports; point all jobs to read from CRM only

Engineering

Inbox Agent Over Forwarders for Workflow Email

Phase two depends on email threads with reference numbers and approvals. A pure forwarder or webhook path is brittle and opaque. A shared-inbox agent (IMAP/SMTP) gives search, labels for workflow state, and safe resubmits with an audit trail. Secure the agent and keep a controlled fallback.

• Store credentials in Secret Manager; scope to read and label-only writes
• Pilot on a shared inbox for 30 days; track false matches and time-to-resolve
• Keep forwarding behind a feature flag as a fallback

Post-Cutoff Policy That Preserves SLOs

Late-intake items retry every 15 minutes and distort failure rates. Decide once: tag and ignore after first detection while preserving visibility in CRM, or auto-cancel per rules. Separate “late-intake volume” from job failures so SLOs reflect system health, not user timing.

• Add a “Post-Cutoff” tag and stop retries after first evaluation
• Dashboard late-intake as its own KPI, separate from failure rate
• Gate behavior behind a feature flag for fast reversal

Quarantine First-Run Noise

Do not broadcast a noisy first production run to every stakeholder. Contain logs to a burn-in channel, send digests, and only expand audiences after stability thresholds are met. Page only on actionable signals.

• Use a burn-in channel until error rate is under 5% for 24 hours
• Send hourly digests for noncritical logs; page on action thresholds only
• Publish go/no-go windows with a named on-call owner

Stage Mirrors Prod; Tests Don’t Hit Live CRM

If cases are identical across environments, stage should catch breaks without touching production. Use stage for synthetic and replay jobs with fixtures, not live CRM polling. Promote dev to stage to prod with a checklist and a defined change window.

• Maintain fixtures for variants, including with and without on-screen reference numbers
• Disable CRM polling in stage; run synthetic and replay jobs instead
• Enforce a promotion checklist with an explicit production change window