Responsible AI Policy

We started Arkintel because we think AI should make work meaningfully better: not just faster, but more considered, better-informed, and more human. That belief comes with obligations.

This policy sets out the principles we design and operate by. It is not a compliance document dressed up in mission language. These are the standards we hold ourselves to, and the ones our customers should be able to hold us to as well.

Some things are simply off the table, regardless of commercial opportunity.

We will not build systems designed to manipulate, deceive, or coerce, whether that means generating disinformation, impersonating people without consent, or nudging users toward outcomes they would not choose if they understood what was happening.

We will not build tools intended for mass surveillance, predictive policing, or social scoring. We will not build AI that automates decisions about people's access to credit, employment, housing, or healthcare without meaningful human review and the right to contest an outcome.

We will not build autonomous weapons or systems designed to select targets or cause physical harm to people. This is not a domain we will work in, on any terms.

We will not accept customers whose primary purpose is to cause harm to people or communities. Use cases that conflict with the principles in this policy are declined, regardless of commercial upside.

This list will grow as the landscape does. We revisit it.

AI assists people; it does not replace their judgement. Every system we ship is built on the assumption that a human should be able to understand, question, and override what the AI does, especially when the stakes are high.

In practice this means: decisions with meaningful consequences for real people are never fully delegated to a model. Our interfaces are designed to surface AI reasoning, not bury it. Audit logs capture what happened and why, so decisions can be reviewed after the fact. And we build in the ability to disengage AI assistance without losing access to the underlying tools.

We think the value of AI is highest precisely when it sharpens human judgement rather than substituting for it.

We do not train our models on your data. Full stop. Every provider we work with is held to the same standard: zero retention and no training on your content.

What you write, store, or process inside Arkintel stays yours, including the outputs the system generates for you. We claim no rights over your content, we do not use it to improve our models, and we do not retain it beyond what is operationally necessary to deliver the service.

How your data is processed depends on how Arkintel is deployed. In our standard deployment, some AI processing is carried out by a small set of vetted third-party providers. All processing takes place within the EU. Every provider operates under a zero-retention agreement: your data is not stored once a request has been served and is never used to train their models.

In our self-hosted deployment, Arkintel runs entirely inside your own environment, air-gapped from the outside world. Your data never leaves your infrastructure, and no third party (including us) ever sees it.

In both cases, processing is governed by European law. GDPR applies, and we have designed our infrastructure to make that compliance structural rather than procedural. Personal and sensitive data is collected only where it serves a clear purpose and held for the minimum necessary time. For how we protect data, data residency, retention schedules, and the providers we rely on, see our Privacy Policy and Security pages.

We are honest about what our AI does, what it does not do, and where it can go wrong.

AI systems make mistakes. They can be confidently wrong, inconsistent across sessions, and poorly calibrated on edge cases outside their training distribution. We document these limitations for every model we ship and surface them to customers so they can deploy responsibly within their own context. We do not paper over failure modes with marketing language.

Where the EU AI Act requires disclosure (for high-risk applications, for systems that interact with the public, for automated decision-making), we provide it clearly and without burying it in footnotes.

We are also transparent about when something has changed: model updates, policy revisions, and significant changes to how data is handled will be communicated directly to affected customers, not just posted quietly to a changelog.

Bias in AI systems is not a theoretical concern; it is an engineering problem with real consequences. We take it seriously as one.

We vet the models and providers we rely on, follow recognised best practices in how we build and deploy AI, and document known limitations clearly. When we identify a concern, we address it or surface it so customers can deploy with full context.

We do not claim our systems are neutral; no model is. Known limitations belong in our product documentation, not hidden in fine print.

Responsible AI starts with secure AI. A system that leaks data, can be manipulated by adversarial inputs, or lacks an audit trail is not a responsible one regardless of its stated values.

Arkintel can be deployed fully air-gapped, and our self-hosted option keeps everything inside your environment. Even in our standard deployment, egress is locked down to a small set of vetted providers and nothing else. Access control is role-based and logged. Every action taken by a user, and by the AI on a user's behalf, is recorded in a tamper-evident audit log. For specifics on how we protect data, see our Security page.

If you discover a security vulnerability in Arkintel, tell us at security@arkintel.com. We acknowledge reports promptly, keep you updated as we investigate, and we will not pursue legal action against researchers acting in good faith. Verified issues are fixed as a priority and disclosed responsibly to affected customers.

These are not features we sell. They are the foundation everything else runs on.

AI inference has a real energy cost, and we factor that into how we build.

Arkintel routes each request to the smallest model that can handle it, automatically stepping up when a task needs more capability. We avoid running inference we do not need and design features to be efficient by default.

We take responsibility for what we ship, and we stay with it.

Deploying an AI system is not a one-time act. Models drift, regulation evolves, use cases change, and failure modes that were not visible at launch can become visible later. We monitor our systems in production, evaluate them on an ongoing basis, and update them when they fall short.

We have a named point of contact responsible for AI policy inside Arkintel, and customers can reach that person directly. When something goes wrong (and at some point something will), we tell customers what happened, what we are doing about it, and what, if anything, they need to do.

We would rather be honest about a failure and fix it than protect our reputation at the expense of the people using our products.

Arkintel is designed for European enterprise and public-sector customers, and we treat the EU AI Act as a design constraint, not an afterthought.

The Act's prohibited practices are prohibited for us too, regardless of whether a customer asks for them. That includes social scoring, manipulative or deceptive techniques that distort behaviour, exploitation of people's vulnerabilities, untargeted scraping of facial images to build recognition databases, emotion recognition in workplaces and schools, biometric categorisation that infers sensitive traits like ethnicity or beliefs, and predictive policing based solely on profiling. Several of these already appear in "What we will not build" above, because we drew our own lines before the law drew them for us.

We classify our AI features against the Act's risk categories. Where we operate in or near high-risk classifications, which includes certain applications in HR, public administration, and access to essential services, we apply the corresponding conformity and transparency requirements.

The Act's obligations come into force on a staggered, and currently shifting, timeline: the prohibitions and the general-purpose AI rules already apply, while the high-risk obligations are being phased in and are, at the time of writing, the subject of amendments at EU level that affect their exact dates. We track these developments, apply each requirement as it takes effect, and tell affected customers in advance when our obligations (or theirs) change.

If you have questions about this policy, how we apply it, or a concern about a specific Arkintel system, email us at contact@arkintel.com. We read it, and a person will reply.

We review this policy at least once a year, and sooner whenever something meaningful changes. It was last reviewed in June 2026; the revision date above reflects that review.