A telephone number looks like an attractive key when a customer register contains repeated names and spelling variations. It is compact, familiar and often populated. The temptation is to treat a shared number as a sufficiently strong reason to combine records, particularly when the organisation wants fewer apparent duplicates. That reasoning skips a question about the circumstances in which the contact was supplied. The number may belong to a household, an assistant or an intermediary. It can explain how to reach a person without uniquely identifying that person.
I made that mistake in a 2019 proposal at Vezqerthi. As director of information integration, I supported automatic merging based too heavily on a common contact. A review of 920 candidate pairs supported 610 same-person relationships, identified 230 different-party pairs and left eighty unresolved. The value of the review was not a claim that telephone information was useless. It showed that the meaning I had assigned to the field was stronger than the evidence behind it.
Ask what the shared field represents
A field's label can make it appear more uniform than the process that produced it. One customer supplies a personal mobile number, another uses a family contact and a third receives assistance through a business office. The database may store each in the same column. That does not make the underlying relationships identical. A data leader needs to understand enough of the collection context to avoid converting storage uniformity into an assumption about people.
This does not mean every matching exercise must reconstruct a person's entire life. It means the proposed evidence should be appropriate to the conclusion. A shared contact can be a reason to investigate a connection. It may contribute to a match alongside other supported attributes. The difficulty arises when the presence of the field is treated as permission to collapse the records without examining plausible alternatives. Convenience is not another piece of identity evidence.
The cost of a mistaken merge is also different from the cost of leaving two records unlinked. An unlinked pair may require staff to search twice or ask for clarification. A wrong connection can place another person's information inside an account that later users assume is coherent. The actual consequences depend on the workflow, and a data review should not invent a breach or adverse insurance result merely because an error is possible. It should nevertheless explain why the two errors deserve separate attention.
Our delivery sponsor argued that the register would become easier to navigate immediately and exceptions could be repaired later. That was a serious operational concern, not an argument to dismiss as impatience. Staff were already spending time with an untidy account. I withdrew the automatic merge because the proposed improvement depended on treating a relationship as identity before the review supported it. The better response had to acknowledge the continuing navigation work instead of pretending that the safer choice was free.
An unresolved pair is still a useful finding
The eighty unresolved pairs were uncomfortable because they did not fit a simple success story. We could not say that they were the same person, but neither could we confidently say that they were different. Retaining that state meant further work and a less complete-looking result. It was still an informative outcome. The review had identified where the available evidence did not yet justify the decision the proposed system wanted to make.
I prefer a defined handling route for that uncertainty to a default hidden in code. What further evidence would be useful? Who is authorised to obtain or assess it? Which service can continue while the relationship is examined? Those questions connect the data state to operating responsibility. The answer should be proportionate to the purpose and should not encourage staff to request unnecessary personal material simply to make a register look settled.
The 610 supported links also required care in communication. They belonged to the reviewed candidate set and entered controlled stewardship. They were not proof that every existing customer relationship was correct or that every future pair sharing the same attributes could be merged safely. A review can support a decision for its actual scope without becoming a universal endorsement of the rule that produced the candidates. I want that distinction retained when results move from a technical discussion into a management slide.
We spent ₹62 lakh on correction and accepted a six-week delay. The expenditure did not recover a quantified loss, and the case does not establish that a live insurance decision had already gone wrong. We withdrew the proposed automatic merge before it entered the customer account. That timing matters. A responsible account can describe a prevented unsupported action without converting a potential consequence into an invented incident or a monetary saving that was never measured.
The merge must remain accountable to its evidence
A linking decision should be explainable after the person who designed the rule has moved on. I want the source records and the basis for a supported relationship to remain available to the appropriate reviewer. A combined view can be useful, but it should not destroy the account of why separate records were connected. If new evidence challenges the relationship, the organisation needs a way to revisit it without reconstructing the entire history from memory.
There is a practical limit to how much detail every user can see. A service colleague does not need the full methodology beside each record. The interface does need to avoid communicating greater certainty than the underlying process supports. An unresolved relationship, a confirmed link and a common contact should not become indistinguishable simply because the screen was designed around one preferred state. Technical simplicity should serve the decision rather than erase its relevant qualifications.
My role in this example included an error in the original recommendation. I cannot credibly ask others to preserve uncertainty while presenting my own judgement as consistently superior. The useful leadership account explains why the shortcut appealed, who challenged it and what changed in the work. Rina's review mattered because it questioned the meaning of the key, not merely the accuracy of its transcription. I had to change the premise I had approved.
The people and figures here are fictional, but the reasoning is available to a real data team. Before combining two party records, ask what relationship each shared attribute can support and what other explanation remains plausible. Then make the handling route reflect that evidence. A smaller register may be easier to present. A register that distinguishes a contact from a person is more likely to help the next colleague understand whom the organisation is actually dealing with.