Your company records what customers say all day. Survey verbatims, support tickets, call transcripts, chat logs, cancellation interviews, app reviews.
Almost none of it gets read. Not because nobody cares. Reading it at scale means putting it somewhere, and putting it somewhere is where the conversation with legal stops. So teams report the score and guess at the reason.
The score is not the reason
Measurement platforms are very good at one job. They tell you what happened. The score fell four points. Satisfaction dropped in one region. Churn rose in one segment.
None of them tell you why. The why is written in plain language, in the text those same platforms collect. So the quarterly review becomes a room full of senior people interpreting a number, with no evidence in front of them.
Signs you are reading the score, not the reason:
- Your readout opens with a number and closes with a theory.
- Nobody in the room has read a customer's own words this quarter.
- The action list would be identical if the number had moved the other way.
Three zones: automate, protect, decide
- Automate. Reading, tagging, clustering, translating, summarizing, counting. The work no team has the hours for, on data that has already passed your gates.
- Protect. Direct identifiers, special category data, and the key that reverses your masking. This never leaves your boundary, whatever a vendor promises in a sales call.
- Decide, case by case. Whether a theme is real or an artifact of the prompt. Whether it becomes a decision. Whether any individual customer gets contacted. A human reviews, every time.
What "anonymized" actually means now
In July 2026 the European Data Protection Board adopted Guidelines 02/2026 on Anonymisation, replacing guidance written in 2014. The public consultation closed on 30 October 2026. Three criteria decide whether data is anonymous:
- No record isolation. No unique combination of values points to a single individual.
- No linkage. No record can be joined to another dataset about the same person by means reasonably likely to be used.
- No inference. No specific, meaningful inference can be drawn. The EDPB notes that inferences produced by querying or prompting an AI model can break this one.
Pass all three and the data falls outside the GDPR. Fail one and it does not. Anonymity is not a property of a file: the same dataset can be personal data in one team's hands and anonymous in another's.
Removing the names is not enough
Most teams strip names and email addresses, then treat what is left as safe to use anywhere. It is not. Pseudonymized data remains personal data under the GDPR. It is a security measure, not a legal basis.
Removing the name changes how the file looks. It does not change what the file is.
The four gates before data reaches a model
1. Minimize
Decide what question you are answering, then bring only the fields that answer it. Most exports carry columns nobody asked for and nobody will use.
2. Mask
Replace direct identifiers with stable tokens, so a thread still reads as one conversation without naming anyone in it.
3. Aggregate
Cut resolution where the answer does not need it. A region instead of a site. A month instead of a timestamp. Resolution is what makes a row traceable.
4. Isolate
Keep the key that reverses the masking in a separate system, under separate access. If one team holds both, you have separated nothing.
There is a trade nobody names. Over-scrub the text and the insight dies with the detail. A complaint stripped of product, channel and timing is no longer a complaint. Decide per question how much resolution the answer needs, and write that decision down next to the dataset.
What the dashboard never showed you
Once the text is usable, you stop sampling and start reading all of it. Then put it next to your operational data: same week, same site, same queue, same release. The question stops being whether the score is down and becomes whether this is the customer or the process.
- Promoters and detractors have different reasons, not mirrored ones.
- Velocity beats volume: a theme that went from five mentions to forty in a quarter matters more than the biggest theme.
- Some themes have no field in your taxonomy, because nobody thought to create one.
Two wrong answers
"No customer data goes anywhere near a model."
Slower, more expensive, and you still cannot say why the number moved. The risk you avoided was never the risk that was hurting you.
"We removed the names, we are fine."
A false sense of safety. Pseudonymized text is still personal data, and a model can still draw inferences from it.
This is not legal advice. Run your own assessment with your data protection counsel, and run it again later: anonymization is expected to be reassessed as re-identification techniques improve.
This article is based on The Human-in-the-Loop Customer Data Guide to Claude 2026, with the full decision matrix and the prompts we use to run this with Claude.
This article reflects Engin Comakci's perspectives, prepared with AI writing assistance.
Sitting on customer text nobody reads?
Redesign helps CX teams turn verbatims, tickets and transcripts into decisions, with the data gates and human review that keep it safe.