We build AI for credit desks, claims queues and grievance counters. That obligates us to a higher bar than a chatbot: sourced answers, human decisions, and the humility to say “I'm not sure.”
Not a manifesto — the constraints our systems are actually built to meet.
This page sets out how Newron approaches the responsible development and deployment of AI. It applies across our products and custom engagements, and it informs the contracts we sign.
Responsible-AI decisions are owned by the same people who build the systems — not delegated to a separate committee that never sees the code. High-impact features are reviewed before deployment for sourcing, oversight and failure modes, and we revisit those decisions as systems change.
We train and fine-tune on data we are permitted to use, with attention to provenance and licensing. For customer deployments, customer data is processed under the customer's instructions and is not used to train shared models without an explicit, contracted agreement. We train on India-hosted data with residency commitments where required.
Every production workflow ships with an evaluation harness. We measure accuracy on the tasks that matter, track regressions over time, and — where the workflow affects people — test for disparate impact across relevant groups. We prefer reproducible, honest benchmarks over cherry-picked demos, which is part of why we open-source some of our evaluation tooling.
Newron is built to assist expert decision-makers, not replace them. Systems are configured so that consequential actions — approving a loan, settling a claim, responding to a citizen — require a human to review and confirm. Interfaces are designed to surface uncertainty and the underlying evidence rather than hide them.
AI systems make mistakes, can reflect biases in their training data, and can be confidently wrong. We document known limitations for each deployment, design for graceful failure and abstention, and we are explicit with customers about what a system should and should not be relied upon to do.
If you believe a Newron system has behaved unfairly or harmfully, we want to know. Contact us via our contact page; concerns are routed to the team responsible for the relevant system and used to improve it.
Our systems are designed to assist, not to decide autonomously. Consequential actions require human review and confirmation.
We test for disparate impact on the populations a workflow affects, document known limitations, and work with customers to monitor outcomes in production.
Not without an explicit, contracted agreement. Customer data is processed under the customer's instructions inside their environment.
We design for abstention. A system that surfaces uncertainty and the relevant policy clause is more useful — and safer — than one that always answers.
If you're deploying AI where the stakes are real, let's talk about how to do it defensibly — and what we'd refuse to build.
