Generative AI has entered business processes faster than regulations anticipated. Under the EU AI Act, organisations must trace what their AI systems produce, along with the prompts and datasets used to generate those outputs.
The prompt becomes critical data
A prompt is no longer a simple user query. It is a decision-making element that directly influences model outputs, and consequently, the downstream actions taken:
- drafting of contractual documents;
- financial analyses;
- assisted medical recommendations;
- legal positions;
- audit or compliance reports.
If the company cannot demonstrate which prompt was used, when, by whom, with which model and which data, it cannot defend its decisions.
What the EU AI Act requires
- the traceability of training data;
- the documentation of models used;
- the retention of usage logs;
- transparency regarding generated outputs;
- the ability to reconstruct a decision a posteriori.
This requirement directly aligns with those already applicable to regulated sectors: trace, prove, preserve.
The need for an AI memory
To meet these obligations, organisations must build a genuine memory of their interactions with AI:
- every prompt is logged;
- every output is preserved;
- every model version is referenced;
- every dataset is versioned;
- every access is traced.
Without this memory, it becomes impossible to explain a decision, reproduce a result, or defend an action taken by an AI agent.
KARL enabling traceable AI
KARL provides this temporal dimension natively. Every artefact linked to AI usage — prompt, embedding, dataset, output — can be stored, versioned, and retrieved at any given point in time.
Organisations thus build a body of evidence ready for audits, litigation, or simple internal analysis.
To learn more
The Challenges article introduces KARL and temporal storage:
Read the Challenges article on KARL →