Login

Responsible AI

Responsible AI

Last updated: June 3, 2026

Nexma builds Jax, an autonomous AI agent for spatial intelligence — spatial design, detection, optimization, and investigation. Because Jax operates on real-world infrastructure where its outputs can carry physical, financial, and safety consequences, we hold ourselves to a high standard of responsibility. This page describes the principles, safeguards, and governance that shape how we develop and operate AI at Nexma, and how those commitments apply concretely to Jax.

1. Our Commitment to Responsible AI

Nexma's mission is to give organizations a spatial intelligence platform they can trust. Jax does not merely generate plausible-looking outputs — it reasons against an explicit world model, validates results against constraints, and surfaces its work for human review. Responsible AI is not an afterthought layered onto the product; it is built into how the platform is designed.

We believe AI should augment expert judgment, not replace it. Jax is engineered to keep a qualified human in the loop at every consequential decision point. The agent proposes, explains, and optimizes; people review, decide, and approve before anything is acted upon.

This page reflects our current practices and will evolve as our technology, the regulatory landscape, and our understanding of best practices mature. We commit to revisiting these principles regularly and to being transparent about how we apply them.

2. Core Principles

The following principles govern every part of how we design, train, deploy, and operate AI at Nexma. They apply to Jax and to any model or system we build or integrate.

Human Oversight and Human-in-the-Loop

Jax is a decision-support agent, not an autonomous decision-maker. For any output with real-world consequences — a network design, an optimization plan, a detection result that triggers action — a qualified human reviews and approves before it is executed. The platform is built so that approval is an explicit, deliberate step rather than a default.

We design interfaces that make it easy for reviewers to understand what Jax produced, why it produced it, and what assumptions it relied on. Human judgment remains the final authority on every consequential decision.

Transparency and Explainability

Jax shows its work. Rather than returning an opaque answer, it exposes the inputs it considered, the constraints it checked against, and the reasoning that led to a recommendation. Optimization results are verifiable — they can be inspected, reproduced, and challenged.

We maintain audit trails of agent actions so that users can trace how a given output was produced. We are also transparent about what our systems can and cannot do, so that expectations are grounded in reality.

Safety and Reliability

Because Jax operates on physical infrastructure, unsafe or unreliable outputs can have material consequences. We apply conservative defaults, multi-layer validation, and uncertainty quantification so that the system fails safe and flags low-confidence results rather than presenting them as certainties.

We test extensively against known scenarios and edge cases, and we build mechanisms to roll back or halt automated processes when behavior falls outside expected bounds.

Fairness and Bias Mitigation

AI systems can reflect and amplify biases present in their data or design. We assess our models for unfair or skewed behavior, scrutinize training data for representativeness, and monitor outputs for systematic distortions that could disadvantage particular regions, populations, or use cases.

Where we identify bias, we treat it as a defect to be corrected. Fairness is an ongoing engineering responsibility, not a one-time checkbox, and we revisit it as our models and their applications evolve.

Privacy and Data Protection

We treat the data entrusted to us with care. Customer data is processed to deliver the service the customer asked for — not repurposed for unrelated ends. We do not use customer data to train our models without explicit consent.

We apply appropriate technical and organizational safeguards to protect data throughout its lifecycle, and we limit access to what is necessary. Our data practices are described in detail in our Privacy Policy.

3. AI Capabilities and Limitations

Responsible use of Jax begins with an accurate understanding of what it does well and where its limits lie. We describe both honestly so that users can apply appropriate judgment.

What Jax Can Do

Jax is designed to assist qualified professionals with spatial intelligence tasks, including:

  • Spatial analysis — interpreting geographic, terrain, and infrastructure data to surface relevant patterns and relationships.
  • Network and route optimization — proposing efficient layouts and plans against defined objectives and constraints.
  • Cost estimation — producing indicative cost projections to support planning and comparison of options.
  • Design generation — drafting candidate spatial designs that a human can refine, validate, and approve.
  • Constraint validation — checking proposed outputs against the rules and limits defined in the active world model.
  • Domain reasoning — applying domain-relevant knowledge to contextualize problems and explain recommendations.

What Jax Cannot Do

Jax is a tool to support expert decision-making, not a replacement for it. It cannot and should not be relied upon to:

  • Replace licensed professionals — Jax does not substitute for the judgment of qualified engineers, surveyors, planners, or other domain experts.
  • Certify or sign off on designs — outputs are not certified deliverables and must be reviewed and approved by responsible professionals.
  • Guarantee real-world performance — projections and optimizations are estimates, not warranties of how a system will behave once built.
  • Override safety requirements — Jax must never be used to bypass applicable codes, standards, or safety obligations.

Known Limitations

Like any AI system, Jax has limitations that users should keep in mind:

  • Edge cases — unusual or rare situations outside the patterns Jax has encountered may produce less reliable results.
  • Novel terrain and contexts — environments materially different from those Jax is accustomed to may reduce accuracy.
  • Incomplete or inaccurate data — Jax's outputs are only as good as the data it receives; gaps or errors in inputs propagate to outputs.
  • Rapidly changing regulations — Jax may not reflect the most recent regulatory changes, which must be independently verified.

4. Data and Training

How we handle data for AI is central to operating responsibly. This section describes our approach to customer data, the basis of our models, and how we maintain data quality.

Customer Data

We do not use customer data to train our models without explicit consent. Data you provide to operate the platform is processed to deliver the service to you, not to improve models for other customers, unless you have opted in.

Where a customer chooses to contribute data for model improvement, we honor the scope of that consent and provide a means to withdraw it. Customer data is segregated and access-controlled in line with our security and privacy commitments.

Model Basis

Jax combines foundation models with an explicit world model and verifiable optimization. This design means Jax is not a free-floating text generator — its outputs are grounded in structured data and checked against defined constraints rather than produced from learned correlations alone.

We select and configure the underlying models with reliability and suitability for spatial reasoning in mind, and we layer our own validation and reasoning components on top of them.

Data Quality and Auditing

The quality of Jax's outputs depends on the quality of its inputs. We apply checks to detect incomplete, inconsistent, or anomalous data, and we surface data-quality concerns to users rather than silently producing results from flawed inputs.

We periodically audit datasets and model behavior to identify drift, bias, or degradation, and we document the outcomes of these audits as part of our governance practices.

5. Safety in Infrastructure AI

Spatial and infrastructure decisions can have physical and financial consequences. We engineer Jax to fail safe and to keep humans in control, using the following safeguards:

  • Conservative marginsJax applies conservative defaults and safety margins so that recommendations err on the side of caution rather than optimism.
  • Human review gatesConsequential outputs pass through explicit review and approval steps before they can be acted upon.
  • Automatic flaggingThe system flags low-confidence, anomalous, or constraint-violating results for human attention rather than presenting them as final.
  • Rollback capabilityAutomated processes are designed so that actions can be reversed or halted when behavior falls outside expected bounds.
  • Multi-layer validationOutputs are checked at multiple stages — against the world model, against defined constraints, and against sanity bounds — before being surfaced.
  • Uncertainty quantificationJax communicates the confidence and limitations of its results so reviewers can weigh them appropriately.

These safeguards reduce risk but do not eliminate it. They are designed to support, not replace, the diligence of the qualified professionals who review and approve Jax's outputs.

6. Monitoring and Improvement

Responsible AI is an ongoing practice. We continuously monitor our systems and improve them through the following mechanisms:

  • Continuous monitoringWe observe system behavior and output quality in production to detect issues, drift, or degradation early.
  • Third-party auditsWe engage independent reviewers to assess our models, safeguards, and practices against external benchmarks.
  • Incident reportingWe maintain processes to capture, triage, and remediate AI-related incidents, and to learn from them systematically.
  • Feedback loopsUser feedback and reviewer corrections feed back into how we improve Jax over time.
  • Benchmark testingWe test Jax against representative scenarios and edge cases to measure reliability before and after changes.
  • Performance metricsWe track quantitative measures of accuracy, calibration, and safety so that improvements and regressions are visible.

When monitoring reveals a problem, we treat it as a defect to be addressed — investigating root causes and updating models, safeguards, or guidance accordingly.

7. Accountability

Clear accountability ensures that responsibility for AI does not fall through the cracks. We have established the following structures and practices:

  • AI Safety OfficerA designated owner is responsible for AI safety practices and for escalating significant risks within the organization.
  • Review boardA cross-functional board reviews high-impact AI decisions, changes, and incidents to ensure they meet our standards.
  • Public reportingWe aim to report publicly on our Responsible AI practices and progress so that we can be held to them.
  • Documented decisionsSignificant decisions about model design, deployment, and safeguards are documented for traceability and review.
  • Escalation proceduresDefined procedures govern how concerns and incidents are escalated, reviewed, and resolved.

Accountability ultimately rests with people, not systems. Nexma takes responsibility for the AI we build, and we expect our customers to apply appropriate professional oversight to the outputs they rely on.

8. Regulatory Compliance

We design our AI practices to align with applicable laws, standards, and emerging regulatory frameworks, including:

  • EU AI ActWe track the requirements of the EU AI Act and work to align our practices with its risk-based obligations where applicable.
  • Israeli AI policyWe follow guidance from Israel's national AI policy and innovation authorities relevant to responsible AI development.
  • NIST AI Risk Management FrameworkWe use the NIST AI Risk Management Framework as a reference for identifying, assessing, and mitigating AI risks.
  • Industry standardsWe align with relevant industry and engineering standards applicable to the domains in which Jax operates.
  • Privacy and data protection lawOur data handling for AI is designed to comply with applicable privacy laws, including Israeli privacy law and the GDPR where relevant.
  • Ongoing complianceWe monitor the evolving regulatory landscape and update our practices as new requirements take effect.

Regulatory compliance is a baseline, not a ceiling. We aim to meet not only the letter of applicable requirements but also the underlying intent of responsible, trustworthy AI.

9. Third-Party AI Providers

Jax relies in part on AI models and services provided by third parties. We select these providers with attention to their reliability, security posture, and approach to responsible AI.

We review the terms and data-handling practices of our AI providers to ensure they are consistent with our commitments — in particular, that customer data shared with them is appropriately protected and is not used to train their models in ways inconsistent with our obligations.

Where third-party models are used, our own validation, world model, and human-review safeguards remain in place. We do not treat third-party outputs as authoritative on their own; they are checked and contextualized within our platform.

We remain responsible for how third-party AI is used within Jax, and we update our provider relationships as our requirements and the available technology evolve.

10. Reporting Concerns

We welcome questions, feedback, and reports of concerns about how Jax or our AI practices operate. If you believe an output is unsafe, biased, or otherwise problematic, we want to hear about it.

You can reach our team using the contact details below:

By email: legal@nexma.ai

By mail: Nexma, Inc., Attn: Legal Department.

We review reports of AI-related concerns seriously and route them to the appropriate team for investigation and response. Where a concern reveals a defect, we treat it as a matter to be corrected.

11. Updates to This Page

We will update this Responsible AI page as our technology, practices, and the regulatory environment change. Material changes will be reflected here, and the date above indicates when the page was last revised.

We encourage you to review this page periodically. Your continued use of Nexma's platform after an update constitutes acknowledgment of the current version.