Skip to content
Responsible AI

Commitments we can be held to, not values on a wall.

Every statement below maps to a control in our platform or a step in our delivery method. If we cannot point at the mechanism, we do not make the claim.

Commitment 01

A person stays accountable for consequential decisions

Where an output affects someone’s money, health, employment or access to a service, a named human holds the decision and has the evidence to overturn it.

How it is enforced
Human-in-the-loop checkpoints defined in ZNYX policy, not in a prompt
Approver identity recorded on every consequential action
Override paths tested before release, not assumed
Commitment 02

Every output can be traced to its sources

If a system cannot show what it drew on, it does not ship. Provenance is a build requirement, not a later enhancement.

How it is enforced
Per-answer citation and retrieval provenance
Immutable audit entry linking input, context, model and outcome
Reconstruction of any decision for the full retention period
Commitment 03

Your data does not train anyone’s model

Client content stays inside the client boundary. We do not use it to improve our products, and we contract for that rather than promising it.

How it is enforced
Deployment inside your tenant, region or air-gapped estate
Zero-retention agreements with model providers
Contractual prohibition on secondary use, in every engagement
Commitment 04

We test for harm before launch and after

Fairness, safety and robustness are evaluated on the population the system will actually serve, then monitored while it runs.

How it is enforced
Golden datasets built with domain experts, not scraped defaults
Adversarial and jailbreak testing on every release
Drift and quality monitoring with defined intervention thresholds
Commitment 05

No lock-in to a single model or vendor

Concentration is a governance risk as much as a commercial one. Systems we build can move between providers without a rewrite.

How it is enforced
Model-agnostic orchestration with routing policy
Documented exit path and dependency register per system
Full IP, source and configuration transferred at handover
Commitment 06

We say what the system cannot do

Limitations are documented for the people using the system, in language they can act on, and repeated at the point of use rather than buried in a manual.

How it is enforced
Model cards written for operators, not only for auditors
In-product confidence and refusal behaviour designed deliberately
Known-failure list maintained per deployment
Where we draw the line

Work we will not take.

This list is enforced at the proposal stage and written into our contracts. It has cost us business, which is rather the point of having it.

Covert manipulation
Systems designed to influence behaviour without the person knowing they are interacting with AI.
Emotion inference at work or school
Affect recognition used to assess employees, candidates or students.
Untargeted biometric scraping
Building or enriching facial recognition databases from open sources.
Social scoring
General-purpose scoring of people that determines access to unrelated services.
Fully autonomous consequential decisions
Denying a claim, a loan or a service with no human able to review it.
Synthetic media without disclosure
Generated likenesses or voices presented as real to an audience.
Governance of ourselves

Who decides, and how often.

Responsible AI fails when it is nobody's job. Ours sits with a named committee that meets monthly, publishes its decisions internally, and can stop a release.

Raise a concern
Responsible AI CommitteeEngineering, assurance, legal and delivery leadership. Meets monthly, minutes retained.
Release authorityThe committee can block any product release or client deployment. That authority has been used.
Escalation pathAny employee or client can escalate directly, anonymously, without going through delivery leadership.
External reviewIndependent assessment of our AI management system annually under ISO/IEC 42001.

Found something we got wrong?

Report a harmful output, a security issue or a policy gap. We acknowledge within one business day and publish a summary of material findings to affected clients.

support@zitrino.com