Every operator eventually inherits the instance: years of imports nobody deduplicated, fields that mean different things per team, contacts who changed jobs twice since anyone checked, and reporting nobody trusts enough to argue with. The cleanup has been scoped three times and cancelled three times, because as a human project it is a quarter of soul-destroying work that decays the moment it ends. This walkthrough is the agentic version of that quarter - the data hygiene mission set pointed at a neglected instance, with the arc it actually follows and the standing state it hands over to.
The instance everyone inherits
The neglected CRM is not anyone's failure - it is continuous decay meeting episodic attention, compounded by every tool that ever wrote to it. Its costs are mostly invisible until listed: lead routing that misfires on duplicates, sends that burn the domain on dead addresses, segments that miss half their members, reports whose numbers lose to anyone's spreadsheet, and - the quiet one - every downstream automation project stalling on "well, first we'd have to clean the data". That last cost is why the cleanup quarter often precedes everything else in the lifecycle marketing program: it is the foundation the rest stands on.
Month one: archaeology
The quarter opens with the damage audit: the agent quantifies duplicate candidates by confidence band, undeliverable share by segment, field drift (how many formats "country" currently has), staleness distribution, and provenance coverage - producing the baseline every later claim of progress will be measured against. Then the first sweeps start working the backlog: high-confidence merges execute automatically with both prior states logged; ambiguous pairs queue for human review per the gate pattern. Expectation-setting matters here: the review queue is genuinely busy in month one - years of accumulated ambiguity surfacing at once - and review capacity, not agent capacity, is the constraint. The practical answer is a daily twenty-minute queue session and ruthless confidence-threshold conservatism, because the fastest way to lose the team is one bad auto-merge a rep discovers.
Month two: backlogs and thresholds
| Mission | Month-two state |
|---|---|
| Dedupe sweep | Backlog clearing; thresholds tuned against observed precision |
| Enrichment pass | Sparse records filling, sourced and dated, conflicts flagged |
| Verification refresh | Active segments re-verified; dead addresses suppressed |
| Standardization | Canonical schemes applied; integrations patched to stop re-polluting |
| Staleness audit | First archive recommendations queued for decision |
Month two is where the system tunes itself to the instance: merge thresholds adjust against the review queue's verdicts (if humans approve 95% of a band, the band promotes toward automatic), enrichment fills with per-field provenance under the observed-beats-inferred rule, and standardization does its unglamorous work - including the part manual cleanups always skip: patching the integrations that were writing the drift, so the pollution source closes rather than refills. Missions that accumulate clean track records start climbing the autonomy ladder, and the review queue visibly shortens.
Month three: steady state
What clean actually buys
The quarter's real return shows up downstream. Routing and scoring run on records instead of noise. Sends inherit a verified base, which the email deliverability math converts directly into inbox placement. The winback program and every lifecycle motion get segments that actually contain their members. And RevOps reporting stops being an argument, because the contract binds clean inputs. Teams scoping the quarter alongside the broader revenue operations configuration usually sequence it first for exactly that reason - it is the unlock the rest of the roadmap was waiting on.
Frequently asked questions
How long does an agentic CRM cleanup take?
Roughly a quarter on a neglected instance: month one quantifies the damage and starts the sweeps, month two clears backlogs and tunes thresholds against human verdicts, month three reaches the steady state and hands over to standing weekly missions.
What does the human do during the cleanup?
Judgment at the gates: a daily short pass on the merge-review queue (busy in month one, short by month three), decisions on archive recommendations, and approval of threshold promotions as missions earn autonomy.
How is progress measured?
Against the opening audit’s baseline: duplicate rate, undeliverable share, field drift and staleness distribution, re-measured monthly. The quarter ends when the numbers reach background levels and the diffs turn small.
What stops the CRM from decaying again?
The handover: the same missions continue on weekly and monthly cadences indefinitely, and the integrations that caused the drift get patched during the quarter. A cleanup without standing missions is a snapshot; this one ends as infrastructure.
Sources
- Google - Email sender guidelines (why verification is part of hygiene)
- Gartner - data quality research
Every playbook on this blog ships as a runnable mission.
Open a workspace and the playbook library is waiting - describe the outcome and the agents carry it end to end, on your plan's monthly credits.