Resolve duplicates against a stable identity rather than against a fuzzy name match, with a dry run before anything merges.
Duplicates are not a tidiness problem. They are a reporting problem and a routing problem: four records for one person means four entries in a campaign, a lead score split four ways, and an account owner who is not sure which record is real. Most CRMs accumulate them structurally, through form fills with personal addresses, imports, and the same person appearing under two employers.
The usual fix is a fuzzy match on name and email, which works until it does not and then merges two different people. This workflow does something different: it resolves each record to an external identity first, and merges on that. The match is only as good as the confidence behind it, which is exactly why the confidence score is the thing the merge rule reads.
A merge is one of very few CRM operations that is genuinely hard to undo. Run the whole workflow in report-only mode first, export the proposed merges, and have somebody who knows the accounts read a hundred of them. Every bad dedupe story starts with somebody trusting a match rate instead of reading the output.
5 steps. The ones in blue are the data layer; everything else is plumbing you already own or can stand up in an afternoon.
Group records that might be the same person: same email domain and surname, same profile URL, same normalised name at the same company. This is the cheap, local step and it narrows what you pay to resolve.
Reverse Contact turns an email, a name and company, or a profile URL into a canonical record. Two CRM rows resolving to the same identity is a much stronger signal than two rows with similar names.
A miss returns null with a confidence score rather than a guess, and the precision threshold can be raised on the request. Below your threshold, nothing merges: the pair goes to a review queue instead.
Report-only mode. Export the proposed merges with both records side by side and the confidence attached. A hundred rows read by somebody who knows the accounts will find the pattern your rules got wrong.
Which record survives, which fields win, what happens to conflicting owners and to the activity history. Decide all of it before the first merge, because this is the step you cannot walk back.
This workflow ships as a skill file: the API surface of every tool in the chain, the build order, the limits that break it, and how to check it worked. Point an agent at it and it has what it needs without reading 3 sets of vendor docs first.
--- name: crm-identity-dedupe description: Build a CRM deduplication pass that resolves records to an external identi ty rather than fuzzy-matching names, merges only above a confidence threshol... --- # Deduplication and identity resolution ## What you are building ## Prerequisites ## The tools, and how to call them → Reverse Contact, HubSpot, n8n ## Build it → 6 steps ## Gotchas ## What breaks, from the vendors' own docs ## How to verify it works → 5 checks
mkdir -p .claude/skills/crm-identity-dedupe curl -o .claude/skills/crm-identity-dedupe/SKILL.md https://www.reversecontact.com/integrations-hub/skills/crm-deduplication.md
For Codex, Cursor, or anything else that reads a file or fetches a URL, hand it https://www.reversecontact.com/integrations-hub/skills/crm-deduplication.md directly. It is plain markdown with YAML frontmatter and no tool-specific syntax, so nothing in it depends on which agent is reading it.
3 of the joins in this chain have a full guide behind them, HubSpot and Reverse Contact and HubSpot and n8n among them. The chain crosses 3 of the seven layers a stack splits into, from the CRM layer outward, which is why no single vendor owns it.
Documented by one of the vendors in the chain, and easier to design around than to discover.
All five are documented by a vendor rather than observed by us, which is the useful part: a limit you can read is a limit you can design around before it costs you a week.
What each component bills on, and which line actually scales with your volume.
| Component | Billing | What drives it |
|---|---|---|
| HubSpot | Any tier | Reads and merges through the API, which is free on every tier. The orchestration sits outside HubSpot. |
| Reverse Contact | Per credit | One call per candidate record, not per CRM record. Clustering first is what keeps this affordable. |
| n8n | Free self-hosted | The right home for this: a long-running job with a dry-run mode, a review queue and an audit trail you control. |
| Human review | An afternoon | The line nobody budgets and the one that prevents the expensive mistake. |
The lever on this bill is the clustering step. Resolving every record in the database is the naive version and costs an order of magnitude more than resolving only the records that already look like candidates.
Reverse Contact resolves an email, a domain or a profile URL into a verified people and company record, with a confidence score your logic can branch on. The free trial takes no card and runs until the 500 credits a new account starts with are spent, which is enough to run this end to end before anyone signs anything.
Free trial, no card required, and 500 credits to start.
Use it as well. It catches the obvious cases cheaply. What it cannot do is tell you that two records with different emails, different name spellings and different companies are the same person, because it only sees what is already in your CRM. Resolving to an external identity is what adds that.
It turns the merge decision into a threshold rather than a judgement call. Reverse Contact returns null with a confidence score instead of a best guess, and the precision threshold can be raised on the request, so you can set a bar above which merges are automatic and below which a human looks. Fuzzy matching gives you a similarity number with no such contract behind it.
That is a policy question and you should answer it before you build. Usually they are the same identity and two legitimate records: one carries the history at the old company, one is the current relationship. Merging them destroys the history. Linking them is almost always the better answer.
Days, and most of that is not code. The clustering and the API calls are an afternoon. The dry run, reading the output, and agreeing the merge rules with whoever owns the accounts is the rest, and skipping it is how this becomes an incident rather than a project.
Tier requirements, sync behaviour and limits read from each vendor's own documentation in September, October and November 2026. Vendors change these without notice.
Reverse Contact appears in this directory as a data provider and publishes it. Every other tool is here on merit and nobody pays for placement.