AI Governance Framework
Operator: Urbano DX (EU / Poland) Contact: urbano@urbanodx.com Last updated: 2026-07-14 Review cadence: at least every 6 months, and on any material change to how we use AI
This document states how Urbano DX governs the use of AI when it designs, builds, and operates software for clients. It maps to the NIST AI Risk Management Framework and supports our obligations under the GDPR and the EU AI Act.
We do not currently hold AI-specific certifications (for example ISO/IEC 42001). We state the frameworks we align to and the practices we follow; we do not claim certifications we have not earned. See the Trust and Security overview for the full posture.
1. Principles
- Human oversight. AI assists a person; it does not replace one. Anything an AI
produces passes a human review surface before it takes effect. No AI feature we build spends money, sends external messages, deploys to production, or changes a legal or financial state on its own.
- Inspectable and deterministic where it matters. We favour designs where the model
produces structured, reviewable output and the consequential step (the database write, the search, the action) is deterministic, testable, and auditable. The client can see why the system did what it did.
- Data minimisation. Only the data needed for a task is sent to a model.
- Client ownership. The client owns the code and the data. We do not subcontract
delivery.
2. Authorising AI use cases
For each engagement the permitted AI use cases are agreed with the client and recorded in the statement of work. A use case is in scope only when the client has authorised it.
- We do not build or operate uses that the EU AI Act prohibits (for example social
scoring, or manipulative or exploitative techniques).
- High-risk AI Act use cases (for example automated decisions on hiring, credit, or
access to essential services) are not built without a prior conformity assessment agreed with the client. By default our deliverables assist a human decision rather than make an automated eligibility decision.
3. Restrictions and guardrails
Systems we build follow these guardrails:
- No autonomous consequential actions. Spending, external communication, deployment,
and legal or financial state changes require a human approval step in the delivered system.
- Untrusted content is data, not instructions. Documents, web pages, and inbound
messages that a model reads are sanitised and never treated as commands to the system.
- No secret exposure. Credentials and secrets are not placed in prompts or exposed
to models.
- No training on client data. We do not use client or end-user data to train models.
4. Risk classification
We classify each AI feature by how reversible it is, whether it has an external effect, and how sensitive the data is. Controls scale with the level.
| Risk | Definition | Examples | Control |
|---|---|---|---|
| Low | Reversible, no external effect, no money, no personal data leaving the client's control | Draft or summarise text, propose an internal note, classify a record | Human can review; runs within the agreed scope |
| Medium | External but bounded and reversible, or small, capped spend | Send a drafted message a person approved, publish a reviewed item, spend under a cap | A person approves the action or the batch |
| High | Irreversible, material spend, legal or financial state change, or contact with people at scale | Production deployment, bulk outreach, anything touching a regulated decision | Explicit human sign-off; reassessed against EU AI Act high-risk criteria before launch |
Any AI feature that interacts with a person (chat or voice) carries a transparency obligation: the person is told they are interacting with AI (EU AI Act Article 50), regardless of risk level.
5. Escalation procedure
- Detect. When a system reaches an action above its agreed authorisation, or an
edge case it is not confident about, it pauses that action rather than proceeding.
- Escalate. The item is raised to a named human: our delivery lead and, where the
action concerns the client's operations, the client's nominated owner.
- Decide. A person approves, rejects, or amends the action. Irreversible actions
always require explicit human sign-off.
- Record. The decision is logged so any action can be traced to the human decision
that authorised it.
We notify the client without undue delay of any security incident affecting their data.
6. Data handling for AI
- No training on client data (restated from Section 3).
- Third-party models are used over their APIs under their terms; the providers we
may engage are listed in the subprocessor register.
- Sanitisation. Untrusted input is normalised and fenced before it reaches a model.
- Audit trail. Delivered systems record AI actions and the human decisions that
approved consequential steps.
7. Transparency
Where a system we deliver interacts directly with a person through AI (a chat or voice interface), that person is informed they are dealing with AI, in line with EU AI Act Article 50. Where interactions are recorded, that is disclosed and handled under the Data Processing Agreement.
8. NIST AI RMF mapping
| RMF function | How we address it |
|---|---|
| Govern | This document; the human-oversight principle; per-engagement authorisation in the statement of work; the operator as accountable owner; a 6-month review cadence. |
| Map | Risk classification of each AI feature (Section 4); assessment against EU AI Act tiers; a subprocessor and data-flow register. |
| Measure | Human review surfaces; audit logging of AI actions and approvals; testing of the deterministic parts of each system. |
| Manage | The escalation procedure; approval steps before consequential actions; the ability to disable an AI feature; incident notification. |
9. Roles and responsibilities
Urbano DX operates as a small, senior team without subcontracting. The operator is the accountable owner for AI governance and the point of contact for reviewer and data-subject requests (urbano@urbanodx.com). Each engagement has a delivery lead responsible for applying this framework; the client remains the data controller for their data.
10. Known limitations
We state limitations plainly:
- No AI-specific certification yet. ISO/IEC 42001 alignment is our roadmap target;
this framework is the first step toward it.
- Governance is applied per engagement, not yet operated as an externally certified
management system.