## TL;DR
Duplicates across CRMs survive single-key matching because each system holds different identifiers. Match in tiers (exact domain plus name, billing id, website, address), score candidate pairs, auto-link only the certain tier, and send the ambiguous tier to humans with the evidence side by side. The agent's job is detection with proof, not silent merging.

```text
agent failed to detect duplicate accounts across two CRMs
```

## Use this when
- The same companies appear in both CRMs with different ids and spellings
- Reports double-count accounts and pipeline
- One CRM came from an acquisition or a team split

## Not for this skill when
- Overlap is tiny and a manual spreadsheet pass suffices
- A single shared identifier already exists and just needs enforcing

## Steps
1. Profile both account sets: counts, and which identifiers each actually populates. Verify: the usable matching keys per system are known
2. Run tiered matching: strong keys first (domain, billing id), then fuzzy name plus address. Verify: candidate duplicate pairs are produced with a match tier per pair
3. Measure precision on a labelled sample of pairs per tier. Verify: each tier's trust level is evidenced, not assumed
4. Auto-link the high-precision tier, and queue the ambiguous tier for human decision with evidence shown. Verify: certain duplicates are linked and uncertain ones get human judgement
5. Add duplicate prevention at creation: search-before-create checks in both systems for new accounts. Verify: the duplicate population stops growing while the backlog drains

## Variant phrasings
### duplicate accounts across two CRMs
A related phrasing for the same underlying problem.
### cross system account matching
A related phrasing for the same underlying problem.
### same company two CRM records
A related phrasing for the same underlying problem.

Compatibility: Compatibility: dual CRM estates (Salesforce plus HubSpot) with warehouse or agent-based matching, current versions.

## Why it happens
Cross-system identity has no master key, so matching is evidence-weighing. Tiered keys with measured precision let automation handle the obvious 80 percent safely while humans spend their judgement only where evidence genuinely conflicts.

## Edge cases / pitfalls
- Domains lie for subsidiaries and franchises. Treat domain as strong evidence, not proof, below the top tier.
- Never merge records as a side effect of detection. Linking preserves both systems' history; merging is a governed business decision.
- Keep the match evidence with the link. When a link is later disputed, the evidence is the whole conversation.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_ShGteAgLK9AgQdTFYWqSXQ