The systems we are willing to put our name on.

AI is moving faster than the rules around it. We are building systems whose behavior is hard to fully predict, on data whose origins are sometimes unclear, for decisions that affect real people.

We do not have all the answers. We do have a working set of commitments. What we will do, what we will not, and how we decide when something is not obvious.

This page is a public commitment. If our practice ever drifts from it, we owe you (and ourselves) an explanation.

Human oversight, by default

AI systems we build for clients are designed to keep a human in the loop on consequential decisions. We treat “fully autonomous” as the exception that requires a deliberate decision and documented reasoning, not the default.

Honest about capability

We tell clients (and prospects) what AI can and cannot reliably do. If a use case is beyond current capability, or only marginally better than a non-AI solution, we say so, even when it costs us the engagement.

Evaluation before deployment

Every production AI system we ship has a written evaluation: what good output looks like, what failure modes we tested for, and the measured rate of each. No “ship it and see.”

Privacy as a constraint, not a tax

We design systems so they hold the minimum personal data they need, for the shortest time they need it. Where on-premise or air-gapped deployment is the right answer for sensitive data, we recommend it, even when it is harder to build.

Bias is a defect

We treat unintended bias the way we treat any production bug: we test for it, we measure it, we fix it, and we report what we found. We do not pretend bias does not exist because it is uncomfortable.

Open by default, secret when warranted

We share methods, evaluation patterns, and lessons publicly through Aproksha Labs. Client-specific data, models trained on it, and confidential business context stay private.

We do not take engagements that primarily exist to:

• Surveil people without their meaningful consent. This includes facial recognition for general public-space monitoring.
• Generate fully autonomous decisions about hiring, firing, lending, sentencing, or healthcare access without human review.
• Produce content that impersonates a real person without their permission, or generate non-consensual intimate imagery.
• Disinformation campaigns, manipulation of democratic processes, or content designed to deceive at scale.
• Weapons systems or use cases targeted at military lethality.
• Any application clearly designed to evade safety regulation or accountability.

This list is not exhaustive. We reserve the right to decline any engagement we believe is materially harmful, even if it is technically legal.

Most of our work is uncontroversial: chatbots, document automation, internal copilots, voice agents, vision QC. Sometimes a request lands in a grey area. When it does, we ask:

• Who could be harmed if this works as intended? If the answer is “people who did not consent to be part of the system,” we slow down.
• Who could be harmed if this fails? The higher the stakes of failure, the higher our bar for evaluation, fallback, and human review.
• Could this be used against the people it claims to serve? Some systems flip from useful to harmful with a small change of context.
• Would we be comfortable explaining this in public? If we would not, we do not build it.

Our practices are informed by, but not limited to, the following frameworks:

• EU AI Act (high-risk system requirements, prohibited practices)
• NIST AI Risk Management Framework
• OECD AI Principles
• India's DPDP Act, 2023 (personal data protections)
• ISO/IEC 42001 (AI Management Systems)

Where regulation is stricter than our practice, regulation wins. Where our practice is stricter than current regulation, our practice stands.

If you believe an AI system we have built is causing harm, behaving unexpectedly, or violating any commitment on this page, tell us:

• Email: support@aproksha.com
• Subject line: “Responsible AI” helps us route it

We will acknowledge within 72 hours, investigate, and report back. If we got it wrong, we will say so and fix it.

AI changes monthly. This page will change with it. When we revise it, we keep a short public changelog so you can see what shifted and why.