Governance of artificial intelligence
How BOUCLR governs AI: human oversight, traceability and minimisation, designed to meet the requirements of the European AI Act.
Version 1.0 · 17 July 2026
The agent proposes, the human validates
At BOUCLR, artificial intelligence does not decide in your place: it prepares, recommends and executes within the framework and thresholds you have set. Human control is not an option, it is the foundation of the platform.
1. A design principle, not an added layer
The European AI Act requires that AI systems remain under human oversight, traceable and proportionate. BOUCLR did not have to become compliant after the fact: the platform was built around these principles from the start. Human oversight, the logging of every action and data minimisation are written into the architecture, not bolted on to tick a box.
We deliberately use the wording “designed to meet the requirements of the AI Act”. The AI Act does not issue a label or a certificate for this type of tool: any promise of “certification” would be misleading. What we guarantee is concrete alignment, verifiable in how the platform works.
2. Human oversight: four modes, chosen by you
You decide, action type by action type, the degree of autonomy you entrust to BOUCLR. The dial goes from full control to supervised automation, and it is yours at all times.
- Manual: BOUCLR proposes, your team executes. No action takes place without an explicit human decision.
- Assisted: BOUCLR prepares the complete draft, your team reviews and triggers it.
- Semi-automatic: BOUCLR executes within a pre-approved perimeter, your team keeps control over out-of-scope cases.
- Automatic: BOUCLR executes and your team supervises, within the thresholds and rules you have defined.
- In every case, thresholds and business rules are entered by humans. No outgoing action towards your end customers escapes the framework you have set. You can raise the level of control or suspend everything at any time.
3. Traceability: every action is logged
Every agent execution is recorded in the platform: which agent, when, on which perimeter, with which result. This traceability makes the AI’s activity auditable end to end and allows accountability at any time.
- Per-agent activity log: timestamp, perimeter and result of each task.
- Shared inter-agent log: learnings are recorded and re-read, making the platform’s collective reasoning explicit rather than leaving it in a black box.
- Data mapping: an overview of the data used, its origin and its circulation.
- Same rules as the rest: these logs are subject to per-client isolation and to retention periods, then purged.
4. Data minimisation and protection
The AI receives only the strict minimum needed for the task. Your end customers’ data stays in the BOUCLR database, in the European Union; the model processes a one-off request and then keeps nothing of it.
- Minimal extract: only the elements useful for the task are sent to the model.
- Pseudonymisation: when the task allows, the real identity is replaced by an opaque reference; the mapping stays in the platform, in the EU.
- No training on your data: the model providers used via API contractually undertake not to use it to train their models.
- Purge: bounded retention periods, anonymised digests, cascading deletion at the end of the contract.
5. Transparency
Your contacts know when they are dealing with content produced with the help of AI. Generated content (campaign drafts, recommendations, analyses) is presented as such to your team, which remains the issuer of decisions. BOUCLR does not impersonate anyone and does not conceal the role of AI in producing content.
6. What BOUCLR does not do
Governance also means clear limits. BOUCLR does not use AI for purposes that the AI Act prohibits or heavily regulates.
- No social scoring of individuals.
- No automated individual decision with legal effect without human intervention.
- No categorisation based on sensitive data (health, opinions, affiliation) to target people.
- No manipulation or exploitation of recipients’ vulnerabilities.
- No outgoing action outside the framework and thresholds defined by the client.
7. Alignment with the AI Act, requirement by requirement
| AI Act requirement | How BOUCLR meets it |
|---|---|
| Human oversight | Four execution modes chosen by the client; thresholds and rules entered by humans; suspension possible at any time |
| Traceability and record-keeping | Per-agent activity log, shared inter-agent log, data mapping, all auditable |
| Transparency | AI-assisted content identified as such; the client remains the issuer of decisions |
| Data management and minimisation | Minimal extract, pseudonymisation, no training on the data, bounded purge |
| Robustness and security | AES-256-GCM encryption, per-client isolation, backups and replication (see “Data architecture and security”) |
| Prohibited practices | Excluded by design (see “What BOUCLR does not do”) |
8. A risk register anchored in the MIT AI Risk Repository
Governance is not declared, it is kept up to date. Our AI risk register is built on the MIT AI Risk Repository (airisk.mit.edu), the academic reference that catalogues and classifies the documented risks of artificial intelligence, published under a CC BY 4.0 licence. We track its published version and re-examine every update.
- An operational register, not a theoretical list: the risks we retain are rewritten in business language and attached to each phase of the platform.
- Every agent and every connector is attached to the register. If a new building block is added without being attached, the platform flags it: nothing escapes governance.
- Every risk carries a human decision (accepted, mitigated, monitored) and a review date: the register is re-examined at least every three months.
- A synchronisation with the MIT base never overwrites a human decision: it proposes, the human decides. Always the same principle.
9. Frequently asked questions
- Is BOUCLR “AI Act certified”? No, and no one can be for this type of tool: the AI Act does not issue such a label. BOUCLR is designed to meet its requirements, verifiable in how the platform works.
- Can the AI send a message to my customers without validation? Only if you chose an automatic mode for that type of action, within the framework and thresholds you have set. By default, the agent proposes and your team validates.
- Can I know what the AI did? Yes. Every action is logged (agent, time, perimeter, result) and available for review.
- Does the AI decide the thresholds on its own? No. Thresholds and business rules are entered by humans. The agent proposes, the human validates.
The questions we get before signing.
No. The model providers used through their APIs contractually commit not to use your data to train their models. The model receives a one-off request, processes it, and keeps nothing.
That is not something you declare, it is something you verify on precise points: where the data is hosted, what is actually sent to the model, how long it is kept, and what happens to it when the contract ends. For BOUCLR: hosting in the European Union, minimal extract sent, pseudonymisation wherever the task allows it, return on request and cascading deletion.
In the European Union, on Hetzner infrastructure in Germany and Finland, with strict per-client isolation and secrets encrypted with AES-256-GCM. Your end customers’ data stays in the BOUCLR database; the model stores none of it.
Three substantive requirements: genuine human oversight, traceability of actions, and proportionality of the data used. BOUCLR was designed against those requirements rather than retrofitted for compliance. One useful point when comparing offers: the AI Act issues no label or certificate for this kind of tool, so any promise of “AI Act certification” should put you on guard.
Every execution is logged: which agent, when, on what scope, with what result. On top of that sit a shared log between agents and a map of the data used, its origin and its circulation. Those logs follow the same rules as everything else: per-client isolation, bounded retention, then purge.
Only if you have explicitly allowed it for that type of action, and within the thresholds you entered. No outbound action towards your end customers escapes the frame you set, and you can tighten it at any time.
Your data is returned on request, and deleted in cascade at the end of the contract, logs included.
Yes, and they are written down: no social scoring of individuals, no automated individual decision with legal effect without human intervention, no categorisation based on sensitive data to target people, no manipulation of recipients, no outbound action outside the frame defined by the client.