Every GTM AI initiative dies on the same rock: the CRM has three records for the same buyer, the MAP disagrees with all of them, and the warehouse trusts none of it. This workflow runs identity resolution, dedupe, and enrichment governance as a continuous, gated, reversible process — the least glamorous agent in the stack, and the one everything else depends on.
The RevOps Data Janitor runs identity resolution, deduplication and enrichment governance as a continuous, gated and reversible process, because merges are destructive and every other agent in the stack depends on the records being right. It is one of fifteen workflows in the Enterprise GTM Agent Stack, an Opsmarshal series on GTM stack architecture.
How the flow runs — live
Every merge carries a 90-day unwind snapshot. Consent fields never take the automatic path.
Duplicate contacts split engagement history, so lead scoring undercounts. Company records fragment across legal names, domains, and subsidiaries, so territory rules misfire and attribution double-counts. Quarterly "data cleanup projects" fix a snapshot; the rot resumes Monday.
Identity becomes a continuously reconciled graph, not a quarterly project. High-confidence merges execute automatically and reversibly; ambiguous ones queue for a human with the evidence laid out. Enrichment writes obey a field-level authority matrix — no vendor overwrites a human-verified value, ever.
| Decision | Choice | Rationale |
|---|---|---|
| Merges are reversible | Full pre-merge snapshots, 90-day unwind window | The catastrophic failure of dedupe automation is a wrong merge you can't undo. Reversibility converts a career-ending risk into a Tuesday ticket, which is what lets confidence thresholds be aggressive. |
| Consent is sacred | Opt-out/consent fields excluded from auto-merge | Merging an opted-out record into an opted-in one is a compliance incident, not a data quality issue. These fields always route to a human, regardless of match confidence. |
| Field-level survivorship | Authority matrix: human-verified > 1st-party > vendor | Record-level "newest wins" destroys curated data. Each field carries a source rank and timestamp; enrichment vendors can fill nulls but never overwrite higher-authority values. |
| Continuous, not batch | Event-driven on record create/update + nightly sweep | Quarterly cleanups treat symptoms. Catching a duplicate 30 seconds after form-fill costs one merge; catching it 90 days later costs broken attribution, misrouted leads, and a confused buyer. |
Three candidate pairs from a real-world-shaped queue. Run the janitor: watch one auto-merge, one route to human review, and one get correctly left alone — with the reasoning logged for each.
rec_0041: priya.sharma@meridianlog.com · Meridian Logistics · VP Ops
rec_2210: p.sharma@meridianlog.com · Meridian Logistics Pvt Ltd · VP Operations
rec_0788: j.chen@gmail.com · (no company) · signed up for webinar
rec_1904: james.chen@vertexhealth.com · Vertex Health · Dir. Analytics · opted out
rec_0902: a.rao@nimbuspay.com · NimbusPay · CFO
rec_3315: a.rao@nimbuspay.io · NimbusPay Labs · Founder
Governance audit log
Identity automation without an operator console is how CRMs get destroyed at 2 a.m. This spec defines the watch rotation, the alarms, and the undo.
Probabilistic matching merges a parent company with its subsidiary; territory and attribution both break. Mitigation: corporate-hierarchy signals (distinct billing entities, distinct domains, legal suffixes) act as merge blockers, not just score reducers.
An enrichment sync misconfigured once quietly overwrites thousands of human-verified titles. Mitigation: field authority matrix enforced at the write layer, not in sync settings — plus a daily diff report of authority-rank violations (target: zero).
A 92% threshold tuned on last year's data performs differently after an acquisition doubles record volume from a new region. Mitigation: monthly precision/recall recalibration on a labeled sample; thresholds are versioned config with change approval, like code.
The 70–92% queue grows faster than RevOps can clear it; ambiguous pairs rot. Mitigation: queue-depth SLO with alerting; batch-adjudication UI groups similar pairs; sustained overflow auto-raises the review threshold and reports the trade-off.
Opsmarshal builds workflows like this one inside client stacks — see GTM stack architecture and AI enablement and agentic workflows, or book a call to walk through your own.
Revenue Operations and AI Agentic Workflows for B2B companies in transformation. We build the machine. You run the company.