Deduplicating a customer list with AI: confirm identity before merging
Settle identity evidence and the history to keep, then sort AI duplicate candidates into merge, separate or hold.
Key points
- AI duplicates are a list to review
- Log identity evidence apart from names
- Pick history, then merge, separate, hold
Overview diagram in preparation
Contents
AI duplicate candidates are not merge instructions
A&A perspective
Deduplicating a customer list means combining the records that describe the same person or company into one record, when a list holds several rows for them. Treat the duplicate candidates the AI lines up as a list to verify before anything is merged.
A&A perspective
Give the AI business cards from a trade show, an inquiry form and a sales rep's spreadsheet, and it will find similar rows and set them side by side. That part is fast and saves effort. But rows that look alike are one thing, and knowing they are the same person is another.
From the sources
HubSpot, a customer management (CRM) tool, has a feature that suggests duplicates. HubSpot's help says it compares values such as name, email, phone number and company name for contacts. A person then looks at the two records side by side and chooses either to merge them or to reject the pair as a candidate.
From the sources
HubSpot's help says this duplicates tool needs a Professional or Enterprise plan and edit permissions. Bulk actions and custom rules need Data Hub Professional or Enterprise. Merging two records you choose yourself is available on every plan.
From the sources
The help also says that before merging in bulk, you should verify that the records really are duplicates and review the associated records to make sure the right records are kept.
A&A perspective
In other words, the mechanism that proposes candidates is separate from the action that decides the merge. Automate & Augment (A&A), which writes this article, reads that split as a line for how much to hand to AI. AI proposes candidates and gathers the material for comparing them. The person who knows the customer is the one who presses merge.
Hypothetical example
Suppose, hypothetically, that the AI proposes 40 pairs from a list. The owner does not review all of them. Each sales rep checks only the pairs involving their own customers. If the AI's list includes a column showing which fields matched, the pairs are easier to check.
A&A perspective
When there are many candidates, it is tempting to work down the list and merge in order. But what the list holds is rows waiting for confirmation. Whether a pair may be merged is decided only after the evidence described in the next section is in place.
Record name matches and identity evidence in separate columns
A&A perspective
On the check sheet for each candidate pair, write the fact that the names match and the evidence of identity in separate columns. Two people can share a name and still be different, and one person can change company or email and still be the same.
From the sources
HubSpot gives each candidate pair a score for how similar the records are. The help adds that when you build your own rule, using fields such as phone number or a custom ID increases detection confidence.
A&A perspective
A&A reads this score as a degree of similarity, not as proof that two records are the same person. A main phone line or a shared email address can be used by many people at one company. A matching name and company still does not settle it on its own.
Hypothetical example
The next two cases show how to use the sheet. Both are hypothetical examples, not real customers.
Hypothetical example
In the first, a company has two contacts with the same full name. Company name and main phone number match, but the email and department differ. The deal records show they are discussing different projects. These two rows are not merged; they are treated as different people.
Hypothetical example
In the second, a contact changed jobs, so the company name and email changed. The name matches, but the company and email do not. If the sales rep heard about the move from the person directly, that is the evidence of identity. If there is no record of that contact, the pair is put on hold until it can be confirmed.
A&A perspective
What to ask the AI for is only filling in the columns for matching and conflicting information. The identity evidence column is written by the sales rep who spoke with the customer. A row whose evidence column is still empty is not eligible for merging.
| Check sheet column | Same name, two people (hypothetical) | One person who changed jobs (hypothetical) |
|---|---|---|
| Matching information | Name, company name, main phone number | Name |
| Conflicting information | Email, department, deal project | Company name, email |
| Identity evidence | None (records show they are different people) | The person's own notice of the move |
| History to keep | Each record stays as it is | Decide which record keeps the former company's deals |
| Verdict | Separate | Consider merging if there is evidence; hold if not |
Merges cannot be undone, so decide what history to keep first
A&A perspective
Before pressing merge, decide which record to keep, and check the activity history and the related deals that will end up in one record.
From the sources
When you merge in HubSpot, the activity history of both records and their associated companies and deals are combined into one record. For property values, the values from the record you keep are used in principle. The other record's email remains as an additional email.
From the sources
There is no operation that cancels a merge. Instead, the help describes a workaround: delete the additional email and use that email to create a new contact. It also says that when a merge updates a property value, the date and time shown for that value is the time of the merge, not the date and time on the original record.
A&A perspective
So the displayed date and time alone cannot tell you when that value was originally set.
From the sources
By contrast, the help says that a duplicate pair you rejected and removed from the candidates can be undone within 14 days. However, if you change property values or other details on the records during those 14 days, the pair no longer appears in the undo list.
A&A perspective
A&A reads the check sheet's separate verdict as corresponding to this kind of rejection. Separate and hold can be reconsidered later. Even after 14 days, you can still merge two chosen records by hand. A merge cannot be reversed. A row you are unsure about is not merged; it is put on hold, with a named person to check it and a due date.
Hypothetical example
If you choose merge in the job-change example, the deals from the former company and the deals from the new company line up in one record. When someone who takes over the former company's account looks through the history, it becomes hard to tell which company a given item belongs to.
A&A perspective
That is why the history to keep is decided first. If you want the former company's deals to stay on the former-company contact record, one option is to keep the new-company contact as a separate record instead of merging. This is A&A's design idea, and which choice is right depends on how a company uses its records.
A&A perspective
To trial AI-assisted deduplication, you can take a few candidates that people judged in the past and compare the AI's check sheets with the human verdicts. If you keep the rows where they disagreed and the effort it took to correct them, that gives you material for deciding how much to hand to the AI.
A&A perspective
For a company with a small list, where a wrong merge would affect few people, this check sheet can be more than it needs. Even then, rows with separate deals get a human look before they are merged.
A&A perspective
The question of how far to let AI carry out actions is covered in the related article "Letting AI take action: decide by whether you can undo it, not the amount".
Summary
A&A perspective
Treat AI duplicate candidates as a list to verify before deciding whether to merge.
A&A perspective
Write identity evidence in a column separate from the name match.
A&A perspective
Because a merge cannot be undone, decide what history to keep first, then choose merge, separate or hold.
Treat AI duplicate candidates as a list to verify before deciding whether to merge. Write identity evidence in a column separate from the name match. Because a merge cannot be undone, decide what history to keep first, then choose merge, separate or hold.
FAQ
Can I merge a candidate as-is when its similarity score is high?
Not on the score alone. The score shows how similar two records are; it does not prove they are the same person. Only rows whose identity evidence column is filled in become eligible for merging.
Is it better to avoid the bulk merge feature?
Even on a plan that has it, use it only for candidates checked against the same evidence. HubSpot's help also says to verify that records are duplicates and review associated records, so the right ones are kept, before merging in bulk. A list that includes unchecked candidates is not a good fit for it.
How do I fix a merge made by mistake?
In HubSpot, a merge cannot be reverted. The help describes creating a new contact from the additional email, but A&A does not treat that as restoring the state before the merge. That is why a row you are unsure about is put on hold rather than merged.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- Review and manage duplicate records
HubSpot · Updated 2026-10-23
Accessed 2026-11-04 - Merge records
HubSpot · Updated 2026-09-05
Accessed 2026-11-04