Chief technology officers
We build AI with your team, controls built in
Logging, evaluation and human approval go in as code from the first commit, for better quality of work at lower cost.
Tijn van der Zant is your partner here, and he signs your briefing.
Eight controls, each answering a duty
We build each one with your team, as code in the running system.
| Duty | The control we build | Applies |
|---|---|---|
| The inventory and your role | A register of every place a model is called, with your role: provider or deployer. | High-risk dates |
| The supplier file | Each model provider as a subcontractor: where data is processed, the audit terms, the way out. | Since 17 January 2025, through financial clients |
| Evaluation | Tests on your own tasks before use, during use and after every change of model. | High-risk dates |
| Logging | Each decision logged: input, output, model, version and who approved. | High-risk dates |
| Human oversight | A fixed point where a named person must approve, with the authority to stop the system. | High-risk dates |
| Staff AI literacy | A record of which staff use which systems, and what each was taught. | Since 2 February 2025 |
| The incident route | A written route for a serious incident: who is told, and who stops the system. | High-risk dates; through financial clients since 17 January 2025 |
| Vulnerability reporting, if you ship software | A reporting route that meets the legal clocks for exploited vulnerabilities and severe incidents. | Since 11 September 2026 |
High-risk dates: from 2 December 2027 for uses listed in Annex III of the EU AI Act, and from 2 August 2028 for AI inside regulated products.
The rule and its source for each duty
- The inventory and your role. EU AI Act, Article 3(3) and (4): the two roles. Article 25: you become the provider of a high-risk system when you put your name on it or modify it substantially.Sources: Regulation (EU) 2024/1689, Article 3, Official Journal; EU AI Act, Article 25.
- The supplier file. DORA, Article 28 and Article 30: duties of the financial firm that reach its suppliers through the contract, among them the locations of the data, audit rights and an exit strategy.Sources: Regulation (EU) 2022/2554, Articles 28 and 30, Official Journal; Regulation (EU) 2022/2554, Article 64.
- Evaluation. EU AI Act, Article 26: deployers of high-risk systems use relevant, representative input data and monitor the system.Source: EU AI Act, Article 26.
- Logging. EU AI Act, Article 26: deployers of high-risk systems keep the logs for at least six months.Source: EU AI Act, Article 26.
- Human oversight. EU AI Act, Article 26: deployers of high-risk systems assign human oversight to people with the competence, training and authority for it.Source: EU AI Act, Article 26.
- Staff AI literacy. EU AI Act, Article 4: the duty to support the AI literacy of staff.Sources: EU AI Act, Article 4; Regulation (EU) 2026/1744, Article 1, point (40)(a), amending Article 113 of Regulation (EU) 2024/1689.
- The incident route. EU AI Act, Article 26: a deployer informs the provider and the authorities of a serious incident. DORA, Article 30(2)(f): a supplier helps its financial client during an incident.Sources: EU AI Act, Article 26; Regulation (EU) 2022/2554, Articles 28 and 30, Official Journal.
- Vulnerability reporting, if you ship software. Cyber Resilience Act, Article 14: an early warning within 24 hours of an actively exploited vulnerability or a severe incident, a notification within 72 hours, then a final report.Sources: Regulation (EU) 2024/2847, Articles 2, 3, 12 and 14 and recitals 11 and 12, Official Journal; Regulation (EU) 2024/2847, Article 71(2) and Article 69(3).
A working example: human approval no setting can switch off
Tijn van der Zant builds software with his own agents, whose source is public: four human approval gates that no configuration can switch off, and every dispatch logged. It is how the partners work, not a product for sale.
Source: The public repository on GitHub.
Where it runs: three kinds of deployment
The deployment changes where the data sits, not what has to be governed.
| Kind of deployment | What the provider covers | What stays with you | When it fits |
|---|---|---|---|
| Custom AI solutions | No training on your data; zero retention on request. | A European region where offered, retention in writing, your log. | Data that may leave under contract; the strongest hosted models. |
| Private deployments | Model maker and other customers excluded, under documented terms. | Regional deployment fixed in code; abuse monitoring off by application. | Processing in the EU needed; provider's jurisdiction accepted as risk. |
| Fully air-gapped systems | No provider; a licence, and a setting blocking outside calls. | Hardware, operation, controlled updates, access control, a file per model. | Data that never leaves; tasks measured on your own work. |
What the providers' documentation does not guarantee
Read on 5 October 2026. Providers change it, so read it again before you decide.
- Custom AI solutions. Zero data retention has exceptions: data may be kept where the law requires it or where automated safety systems flag it, and some features fall outside the arrangement. Without the arrangement, logs for abuse monitoring are kept for a period, in one provider's documentation up to 30 days. One provider documents no European processing option on its own interface. The providers read for this row are under United States jurisdiction.Sources: Anthropic, documentation on data retention; Anthropic, documentation on data residency; OpenAI, documentation on data controls; United States Department of Justice, resources on the Clarifying Lawful Overseas Use of Data Act.
- Private deployments. A deployment type labelled global may process in any geography. By default a sample of prompts may be stored for abuse monitoring and read by the provider's staff; switching that off takes an approved application. Your own key covers data at rest. A provider based in the United States stays under United States jurisdiction, whatever the region. A European provider of open-weight models carries the quality gap of those models, measured on your own work.Sources: Microsoft Azure, documentation on data, privacy and security for hosted models; United States Department of Justice, resources on the Clarifying Lawful Overseas Use of Data Act.
- Fully air-gapped systems. Nobody else operates, patches or monitors the system, and a licence says nothing about how good the model is on your work. A licence text written by the maker itself has to be read before use.Source: Hugging Face, an example of a licence text written by the maker of the model.
How we work
The free briefing costs nothing. Each paid step has its own fixed price, in writing before it starts.
- 1. Free briefing
- 2. Discovery
- 3. Design
- 4. Build
- 5. Run and improve
What you risk
The EU AI Act binds you, DORA reaches you through your financial clients, and the Cyber Resilience Act binds you if you ship software.
For breaking the obligations the Act lists, such as the transparency duties
€15 million or 3%
Source and caveat
Breaking the other duties the Act lists, such as the transparency duties: up to €15 million or 3% of worldwide annual turnover, whichever is higher.
European Union Artificial Intelligence Act (Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744), Article 99(4). Applies from 2 August 2026.
This is the most the law allows, not what is usually imposed. A fine can only follow a duty that applies: the transparency duties from 2 August 2026, the high-risk duties from 2 December 2027 and 2 August 2028.
Source: EU AI Act, Article 99(4), consolidated version of 27 July 2026, EUR-Lex, accessed 5 October 2026.
For a software maker that misses the cybersecurity requirements of the Cyber Resilience Act
€15 million or 2.5%
Source and caveat
Not meeting the essential cybersecurity requirements or the duties of manufacturers: up to €15 million or 2.5% of worldwide annual turnover, whichever is higher.
Cyber Resilience Act (Regulation (EU) 2024/2847), Article 64(2). Applies from 11 December 2027.
This is the most the law allows, not what is usually imposed. The reporting duty has applied since 11 September 2026; the fines apply from 11 December 2027.
Source: Regulation (EU) 2024/2847, Article 64(2), Official Journal text on EUR-Lex, accessed 5 October 2026.
The ceiling is the most the law allows, not what is usually imposed. All fines, with their sources.
Worked examples
These are worked examples. They do not describe real clients.
- The bank's questionnaire now asks about your model provider
Written for the chief technology officer of a software company.
- The ranking your recruiters asked a chat assistant for is on the high risk list
Written for the chief technology officer of a large employer.
- What the board of a water company approves when it approves artificial intelligence
Written for the chief technology officer of a drinking water company.
This page was last checked on 5 October 2026.