Four records, one company, and the two years of revenue in between.

A machine tool builder we scoped had a customer called Meridian Fasteners. SAP knew them as MERIDIAN FASTENERS PVT LTD. The CRM had Meridian Fastners, a typo from 2019 that nobody ever fixed. The dealer portal had M/s Meridian. The service application filed them under a plant address in Rajkot with no company name at all.

Four systems, four customers. Sales saw an account that had not ordered in two years. Service saw fourteen machines under contract and kept visiting. Finance saw a receivable. Nobody saw that the same company had let an AMC lapse, taken a competitor's quote and was about to replace eleven machines.

Why it does not solve itself

Fuzzy matching on names gets you most of the way and the last stretch is where the money is. Two similarly named entities in adjacent districts may be one company or two. A subsidiary may need to resolve to its parent for credit and stay separate for service. A dealer may be a customer in one line and a channel in another.

None of that is decidable by an algorithm alone, which is why the review queue exists and why your own operations people label the ambiguous pairs. It costs them a few hours in total and it produces the golden dataset that every subsequent evaluation is scored against.

Lineage, or nobody believes the result

Every resolved field carries where it came from, when it was read and what transformed it. The first time a number looks odd to someone senior, they will ask. If the answer takes a week, the number stops being trusted permanently, and after that it does not matter whether it was right.