The short version
- A RevOps checklist for judging an AI co-worker's dedupe work: what merges alone, what waits for a human, and the log fields to demand before connecting HubSpot.
Contents
- 1Which merges can run unattended in a CRM dedupe pass?
- 2Which merges should wait for approval?
- 3What must the execution log show before you connect HubSpot?
- 4How do you pilot a CRM dedupe pass before the live portal?
- 5How does rollback work when a native merge cannot be reverted?
- 6Where does Coco's approval model fit this checklist?
- 7What is the decision rule?
An AI co-worker's CRM dedupe pass is worth connecting to HubSpot only if three things hold: exact-key merges with nothing attached run unattended, every other pair waits for a human, and the execution log lets you rebuild any merge without the tool's help. If a vendor cannot show you all three before connection, the pass is a risk, not a feature.
A note on provenance: this note examines the mechanism from outside. No platform version was run directly, and no HubSpot release was observed, so treat the thresholds below as a checklist to test, not as measured behaviour.
Which merges can run unattended in a CRM dedupe pass?
Only merges where the match is deterministic and the loser record carries nothing a human would grieve. In practice that is a narrow class: identical lowercase email on two Contact objects, no open Deal association on either, no owner conflict, and no enrollment in an active workflow.
The survivor rule must be stated, because routing disputes are ties in disguise. Pick one and write it down: oldest createdate survives, or most recent activity survives. A CRM dedupe pass that cannot name its tie-break is guessing.
A record that matches no rule is left alone and logged as "no match". It is never merged on a hunch, and it is never deleted.
Which merges should wait for approval?
Everything with a fuzzy score, and a few things even a perfect score should not touch. Three groups go to the review queue regardless of confidence:
- Company objects with parent-child links, since merging collapses a hierarchy someone built on purpose.
- Any pair where either side has an open Deal, because pipeline totals and forecast snapshots shift.
- Pairs with conflicting
hubspot_owner_id, because someone is about to lose an account.
Published guides disagree on the cutoff, with examples running from roughly 75% to 95% confidence. The number matters less than what sits above it. A 99% score on a parent company is still an approval item, and a CRM dedupe pass should route it to a person every time.
What must the execution log show before you connect HubSpot?
Each log entry should let a stranger reconstruct the merge at 03:00 without asking anyone. The minimum fields:
- Both record IDs and which one survived, plus the tie-break rule that chose it.
- The match reason (which key fields agreed) and the confidence score.
- Full before-values of every property on the loser, and the values of any property the survivor overwrote.
- Associations on the loser at merge time: deals, companies, workflow enrollments, owner.
- Who or what approved it: an approver identity for queued pairs, or the named unattended rule for automatic ones.
- A timestamp and the run ID, so one pass can be filtered as a unit.
The before-values are the field most logs skip. Without them there is no rollback, only a record that something happened. Ask for this log from the first dry run, not after the first live merge.
How do you pilot a CRM dedupe pass before the live portal?
Use a developer sandbox or an exported sample, and run the pass in dry-run mode first: it proposes merges, writes the log, and changes nothing. Then read the proposals, not the summary.
A worked example, stated as an assumption: take 400 proposed pairs, hand-review a random 60, and set the pass bar at zero wrong merges in the auto-merge tier. In the approval tier, an error rate you can live with is whatever your reviewers can clear before the queue goes stale. The unattended tier gets a zero-tolerance bar, because its mistakes arrive without a reviewer.
How does rollback work when a native merge cannot be reverted?
It does not revert. It reconstructs. Native HubSpot merges are not reversible, so rollback means recreating the loser from its logged before-values, restoring overwritten properties on the survivor, and re-pointing associations.
Then verify three things: the recreated record's workflow enrollment, its hubspot_owner_id, and its deal associations. This is slow by design, which is the argument for keeping the unattended tier tiny. Rehearse it once on sandbox data before any live CRM dedupe pass runs.
Where does Coco's approval model fit this checklist?
Coco is pre-beta, so this note makes no claim about results. What its published facts describe maps onto the checklist: autonomous steps are permission-gated and confidence-scored, they leave a logged execution trail, and the user grants the bounds in which a strategy may run. Skills and strategies are versioned with rollback, and customer-facing actions sit behind a hard approval floor. The Coco product page lists HubSpot among its integrations.
Those are statements about control and auditability, not proof of reliability. The test is the same as for any vendor: ask to see the log fields above on a dry run of its CRM dedupe pass.
Other vendors state their positions in their own words. Clarify prices AI work in credits, so ask what happens to a half-finished dedupe pass when credits run out. Artisan now pitches running alongside your stack, so ask what write access that mode holds in the CRM. Neither page, as read on 2026-11-08, says how merges are approved or logged, so those remain discovery questions.
What is the decision rule?
Connect only when the unattended tier is small enough to read in one sitting, the approval queue shows the match reason next to each pair, and a rehearsed rollback has restored one record end to end. Correction policy for this note: if a platform version is later observed to behave differently, the change will be dated here.
Sources
- Coco: the approval-first AI co-worker for GTM teams — Coco, 2026-11-10
- Clarify pricing — Clarify, 2026-11-08
- Artisan homepage — Artisan, 2026-11-08