CRM 90 to 12: AI cannot fix dirty data

In one of my corporate roles I was handed a tidy-sounding task: merge and map more than 90 Salesforce instances down to 12.

The number we were aiming for was not one. It was 12.

The single 360-degree view of the customer is the thing every CRM is sold on, yet even the ambitious version of success here meant a dozen competing versions of the truth.

What you discover when you actually try to merge them is that the data was never one thing to begin with. The same customer sitting in five instances with five sets of details. Fields that meant one thing in one system and something else in the next. Records last touched years ago that nobody would admit to owning.

Volume was never the problem. We had data spilling out of every corner of the business. What we lacked was the discipline to keep any of it worth trusting, because maintaining data is slow and unglamorous, and the slow, unglamorous work is always the first thing to be dropped.

Which is what makes the current pitch around AI worth pausing on. AI is often sold as the tool that finally lets you drink from the firehose. Cut through the noise, ask your data a question, get a clean answer back with minimal need for data maintenance.

But point a confident model at 90 unreconciled instances and, unless the system is built to expose conflicts, it will not warn you that they disagree.

It will choose, or it will blend, and hand back something fluent and authoritative built on records that contradict one another. The fragmentation does not go away. It simply stops being visible.

AI does not clean the water in the data lake. It does make drowning look like swimming.

Before you ask a model to find the signal in your data, a plainer question comes first.

Do you actually know what is in there?

The lesson: you can’t outsource data reconciliation.