AI links records: how to avoid merging two customers into one person
How to evaluate AI-assisted record linkage: false merges, missed duplicates, groups, field provenance and repair.
Article language: English

In this article
There are two Anna Nowaks in a database. Their email addresses are similar, they use the same company domain and both ordered materials for the same event. The system suggests a merge. A person sees two similar rows and clicks "approve", only to discover that one Anna organises the event and the other represents a supplier. Sharing a surname does not mean sharing one case folder.
AI-assisted record linkage can tidy a mess, but it can also create a very consistent record of a person who does not exist. The problem starts when finding similar records flows straight into deciding they have the same identity. An audit should separate those steps and examine what happens to data after a merge.
A controlled synthetic database is the easiest place to evaluate the process. When preparing one, check consistency of relationships and events in synthetic data (in Polish), especially the roles and history assigned to each person. You can deliberately include namesakes, shared inboxes and changed names without exposing real customer relationships. The expected answer is known in advance, so both a false link and a missed duplicate can be checked.
First establish what a record represents
One row may describe a person, a company, a branch, a delivery location or a person's relationship with an organisation. Two records of the same person should not always become one row if they hold different roles. Before evaluating matching, document the database unit and the relationships allowed between units.
In a fictional contact catalogue, Anna may appear as a company employee and as the contact for a particular order. Changing employers does not mean an old order should receive a new company. If the data model does not distinguish a person from their relationship, a merge tool cannot fix that ambiguity merely by guessing the surname correctly.
A similar issue applies to organisations. A brand, an entity and a branch may share words in their names but play different roles in a process. Do not automatically assume a longer name is a more complete version of a shorter one. The reviewer needs an identity criterion suited to how the database will be used.
In deduplicating a company database before moving to a CRM (in Polish), field cleaning is part of data preparation. Here the key moment is when a tool proposes a link and the system begins treating two sources as one object. That moment needs its own control and decision trail.
At this stage, set out three possible outcomes: the same object, different objects, and insufficient basis to decide. The last answer is necessary. If an application always requires "yes" or "no", some uncertain cases will enter one category simply because the interface has no other route.
A candidate for review is not an instruction to merge
OpenRefine's description of cell clustering explains grouping similar text values and user approval of proposed changes. There, spelling similarity helps tidy values. Finding similar names alone does not establish the identity of people or organisations.
Evaluate your process in two stages. First check whether the mechanism finds pairs worth examining. Then examine whether the decision rule distinguishes a duplicate from a misleading neighbour. Good candidate selection can coexist with overly aggressive automatic merging. One overall "accuracy" score hides that distinction.
Prepare easy and difficult pairs. The same synthetic person might have a misspelled surname, an old address or a number written with different spacing. Different people might share a full name, a switchboard number or a departmental inbox. Each pair should have an explanation of the expected outcome recorded before the tool runs.
Do not treat a shared inbox as several independent pieces of evidence. A domain, organisation name and office address may describe the same connection to a company. In the report, show whether the team understands what the signals concern. Mechanically adding similarity scores can look convincing precisely where they all come from one shared feature.
Add a case of data changing over time. An old and a new address may belong to the same person, but may also belong to two successive users of a role-based inbox. You need dates and field provenance to assess it. Replacing the old value with the new one without history makes a later error harder to explain.
In visual product search (in Polish), we distinguish a similar appearance from the same model. A similar name plays an analogous supporting role in records. Instead of showing only a similarity percentage, the review interface should reveal agreements, differences and the data sources needed for a decision.
Count both directions of error
A false link joins different objects. A missed duplicate leaves records of the same object separate. Their consequences differ. The first can mix case histories; the second can maintain repeated contact. Do not decide their relative importance solely by how easy a spreadsheet is to clean.
Splink's evaluation documentation describes checking false matches and missed links against labelled examples at a chosen threshold. It is a useful reference: a known outcome and an explicit decision boundary are needed. The number of groups created alone does not measure quality.
On a synthetic sample, show the number of pairs that should be linked and those that should remain separate. Record the tool's results alongside them. If you change the threshold, compare the same cases and both error types. Fewer duplicates are not enough if more customers are incorrectly merged.

Test the sample after shuffling input order too. If the same synthetic record group ends with different selected data, find out why. The rule may keep the first entry encountered. That may be an intentional configuration, but it needs to be named and assessed against the task rather than credited to the quality of person recognition.
Set aside a group for manual decisions and check its size. A solution can avoid automatic errors by sending nearly everything to a person. That need not be a bad decision, but the actual work must be counted. Record the time required to review a case and the information the reviewer lacked.
In human approval of AI actions (in Polish), an approval button alone does not guarantee meaningful control. Before a merge is approved, show the effect: which fields will remain, which links will move to the record and whether both source histories will still be available. Otherwise the person approves the operation's name rather than its result.
Test rejection of a suggestion. After the two Annas are marked as different people, does the tool soon suggest the same pair again? Does the decision have a reason and scope? Rejecting that specific link should not prevent checking a real duplicate of either person elsewhere in the database.
Inspect the group before its separate elements disappear
Pairs do not yet describe the entire group. Record A may resemble B because of a shared phone number, and B may resemble C because of a surname. It does not automatically follow that A and C describe the same person. Add such a chain to the test and inspect the final group. This reveals effects not visible when judging individual suggestions.
Walk through the rule for selecting values after a merge. Which address becomes primary, what happens to an empty field, and how is a contradiction resolved? "The newer record" is not a complete rule if one source was automatically updated yesterday and another holds confirmed information from a week ago. The basis for selection must suit the field.
Keep identifiers of the input records and the provenance of retained values. This helps trace why a phone number appeared on a contact. It does not mean copying the entire database into the report. The data used in a check should fit the task, a point to connect with an AI data-flow register (in Polish).
For fields being merged, show at least one conflicting value and one missing value. Adding a number to an empty cell needs a different assessment from replacing an existing number. If the report calls both operations "contact completed", it is hard to see that some correct values were overwritten. Separate additions, replacements and unchanged data in the change preview.
After the merge, open related objects too: cases, orders and assigned roles, if the system has them. The contact row can look correct while an old order has acquired the wrong organisation. Accepting only the row misses the effects for which the database is being cleaned in the first place.
Re-import one source record. Check whether the system recognises its existing link or creates a new customer and starts the merge again. This distinguishes one-time file cleaning from a process continuously receiving data. The record of links must help during the next database feed too.
For a many-to-one relationship, check that all source identifiers still point to the right object. The fictional Anna may have three valid records from different systems. Linking two and leaving the third without a recognisable relationship does not finish the task. At the same time, do not add the other Anna's similar record merely to create one elegant group.
Carry out a controlled separation of an incorrect group. Can the source records and assignments be reconstructed? What happens to data added after the merge? If the tool has no simple undo, establish the repair method and scope of automatic changes before running the process. A copy of the input helps, but it does not by itself decide who owns a later note.
An operational review with Aeneia Audit can start with a synthetic sample, a description of the rules and a view of the final group. This lets you test the process logic before cleanup changes the history of real contacts.
Review the use of the cleaned database as well
If the database feeds aggregate analysis, inspect the result after merging. A false link can change the number of customers, responses or contacts with a case. In analysing open survey answers (in Polish), it is especially important to know whether you count statements, people or events. "Duplicates removed" does not replace that definition.
Finally, record the scope in which automatic merging was not allowed. This might include shared departmental inboxes or records without a source identifier. The process owner should know where separate checks are still needed, without inferring that the entire database is ready from a good result on easier pairs.
Decide which changes trigger another sample review: a new source type, changed field selection rules or a different grouping mechanism. Comparing the number of remaining rows is not enough. Return to difficult pairs and chains, inspect decisions and test repair.
Finally, ask a database user to find both synthetic Annas and explain their latest cases. If they can reconstruct sources, roles and the reasons for linking or separating the records, cleanup has produced a useful result. Fewer rows are then the consequence of understandable decisions, and both Annas still have their own contact history.










