Everything you know about a customer, in one place
What is scattered across systems, read and unified without moving it
The problem
A customer fills in a survey and writes three lines explaining why they gave a six. Another opens a ticket and describes a problem in detail. Another replies to an email saying not now, to talk again in September. A salesperson notes after a meeting that the contact changed and the budget moved to a different department.
Each of those things lives in a different system and none of them knows the others exist. Whoever serves that customer tomorrow sees one part, and which part depends on where they enter from.
Breaking silos has been sold for fifteen years as an integration project: unify systems, migrate everything to one place, two years and a budget that almost never gets approved. And even if it is approved, it solves half the problem, because it unifies what is structured and leaves out everything written in text — which is precisely where the why lives.
The solution
Read the sources where they are, without moving them, and unify what they say about each customer.
The result comes in two forms, and both are needed.
Data.
Fields that can be used like any other: reason for the complaint, topic mentioned, stated intention, when they asked to be contacted again, product they enquired about. They go to the CRM or warehouse alongside everything else, and serve for segmenting, filtering, triggering an action, or feeding a model.
Context in text.
A profile someone can read in thirty seconds before taking a call or entering a meeting: what happened, what they requested, what is pending, in what order. A timeline, not a data dump.
The two consumers are different. Data feeds systems; the profile feeds people — and also any system that later has to make a decision about that customer using the complete picture, not just the transactions.
Every data point saves where it came from, with the exact phrase. This is not a detail: it is what allows someone to question a datum and verify it in one click, rather than having to trust it.
And it works backwards. The history that already exists is processed once, and suddenly there are two years of context where before there was only a purchase table.
The extraction system reaches agreement with the human expert's criterion above 75% (κ > 0.75 on the Cohen scale) — a threshold equivalent to what would be demanded of a quality analyst. Fields that do not reach that threshold are not published until the system learns them.
Which signals are worth extracting, and which are not
This is where it is decided whether this works or is an expensive exercise.
Almost anything can be extracted from text, and that ease is the trap. A field filled in 3% of cases is not useful. One that depends on how the question was worded is not either. And one that sounds interesting but that no one will look at is worse than not having it, because it takes up space and gives a false sense of coverage.
The ones worth extracting meet two conditions: they appear frequently enough in your sources, and there is someone who would do something different if they had them. That list is assembled with your team before building anything, and it is the part of the work that most determines the result.
Distinguishing them is what is learned by taking data projects to production, where a field no one uses is not an aesthetic detail but something that needs to be maintained for years.
How it integrates
Sources are read where they are. Fields appear in the CRM or in the warehouse, and the profile is accessible from the customer record, updating as new information arrives.
The same system, other applications
Unifying what is scattered about the same entity applies wherever the entity is not the customer.
Sales follow-up.
Meeting notes and recordings converted into lead status, pending items and decision-maker, without anyone updating the CRM manually.
Open survey responses.
The text alongside the NPS classified by topic and quantified — where the reason the numeric score does not capture lives.
Internal project progress.
Minutes, emails and reports converted into status, blockers and owners.
Supplier or product profiles.
The same mechanism when what needs to be filled in is not a customer.