Context card
Three processing purposes when AI supports case handling
Which processing purposes have to be kept apart when an AI system supports the handling of individual cases?
The short answer
When an AI system supports case handling, at least three purposes belong in separate descriptions: building and testing the model on data, analyzing the procedure in aggregate, and supporting the handling of one individual case. Each purpose has its own data set, its own basis, its own retention and its own oversight arrangements. Evidence collected for one purpose does not transfer to another.
For: Project managers and technical writers in public-sector AI projects
Key points
- Development data serves the model; it does not decide a case.
- Procedure analysis describes throughput and bottlenecks, not one applicant.
- Case support produces a proposal that a responsible person has to be able to check.
- A match found in public sources is a lead, not an established fact.
- The three purposes need separate access rights, retention periods and records.
The context
Three purposes, three data sets
Development is the first purpose: data is used to build a model and to run the training, validation and test steps that show whether it behaves as intended. The relevant properties here are the origin, representativeness and data quality of the set, plus the question of whether it may contain personal data at all, and under which conditions.
Procedure analysis is the second: aggregated figures on volumes, durations, bottlenecks and error rates. Such figures describe a process, not a person, and they lose their value when they are quietly reused to characterize an individual file.
Support in a single case is the third: a suggestion, a cross-check, a flag in one file. Here the responsible person must be able to see what the output is based on, how firm it is and what remains to be checked. As of October 2026, the EU AI Act contains separate requirements for high-risk systems on data governance over training, validation and testing data sets and on human oversight during use; they address different objects and cannot substitute for one another.
A search hit is a lead, not a finding
Comparing statements in an application with publicly available internet data produces matches. A match raises three open questions: is it the same person, how reliable is the source, and does the content have any legal weight for this decision. Until those are answered, the result is an indication to be examined, and the system should present it that way rather than as a confirmed inconsistency.
What the file has to show
For each processing operation, technical documentation should state the purpose, the data origin, who has access, how long the data is kept, and at which point a person actually decides. Keeping the purposes apart does not by itself settle whether a given operation is lawful; that assessment stays with the responsible body and its legal review. What the documentation can show is that the three purposes are not silently merged — the most common way for development data to end up influencing a decision nobody meant it to influence.
Questions readers ask next
- Why can development data not simply be reused for case support?
- It was assembled to measure model behavior, with its own selection, labeling and retention. Using it in a live case changes the purpose and the people affected, which requires its own basis and its own description.
- Who is responsible when a proposal turns out to be wrong?
- The decision stays with the person or body that makes it. A system can make that practicable — showing sources, uncertainty and the pending action — or it can make it hard, through time pressure and unexplained outputs.
- Does aggregated monitoring need less documentation?
- It needs different documentation. Definitions of the figures, the aggregation level and the re-identification risk matter more than individual explanations, because small groups can make aggregated data personal again.
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text — Publications Office of the European Union (EUR-Lex), July 27, 2026
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 — NIST, January 26, 2023
Review log and changes
Every context card is checked against its sources before it is published, and again whenever it changes; the date under the byline is the last review. Corrections (something was wrong) and additions (something was missing) are logged below with date and time (Berlin time). Typos, formatting and link fixes are not listed.
Reviewed
No corrections or additions since publication.