Responsible Artificial Intelligence

At SERGENTA AI, we build artificial intelligence systems that are designed to support people. This is not a marketing position — it is a design principle that shapes every decision we make about how our AI is developed, deployed and governed. We believe that the most valuable applications of AI are those that enhance human capabilities, not those that attempt to replace human judgement.

AI Should Augment People

AI excels at processing large volumes of information, identifying patterns, generating drafts and performing repetitive tasks. Humans excel at understanding context, exercising judgement, navigating ambiguity and taking responsibility. The most effective systems combine these strengths — using AI to handle the heavy lifting while humans make the decisions that matter.

This is what we mean by augmentation. AI in our platform is designed to help people work faster, smarter and more effectively — not to make decisions on their behalf without their knowledge or consent. An AI agent might research a topic, draft a document or analyse data, but the person who reviews, approves and acts on that work remains in control.

Human Decision-Making Remains Important

There is a temptation, particularly as AI systems become more capable, to delegate increasingly consequential decisions to them. We resist this temptation. Decisions that affect people, organisations or society — hiring decisions, financial decisions, legal decisions, safety decisions — should remain with human beings who can understand the full context, exercise moral judgement and accept accountability.

Responsible Innovation

Innovation and responsibility are not opposites. The most durable innovations are those built responsibly — considering their impact, anticipating their misuse and designing safeguards from the outset. Responsible innovation does not mean moving slowly; it means moving thoughtfully, with awareness that the systems we build affect real people in real ways.

Transparency and Accountability

We are transparent about where and how AI is used in our platform. We document our AI systems, their capabilities and their limitations. We do not claim that our AI is perfect or that it can do things it cannot. When our AI makes a mistake — and it will, because all AI systems do — we acknowledge it, investigate it and work to prevent it from recurring.

Accountability means that someone is responsible for every AI system we deploy. Not a model, not an algorithm — a person. Someone who can answer for the system's behaviour, address concerns and make changes when needed.

Long-Term Thinking and Engineering Excellence

AI is not a short-term opportunity — it is a long-term transformation. We build with a long-term perspective, investing in foundations that will endure. Swiss engineering principles — precision, reliability, quality — guide our AI development just as they guide everything else we build. We would rather build something that works reliably than something that impresses momentarily but fails under real conditions.

Principles are not aspirations that live on a page. They are operational commitments that guide design decisions, development practices and governance processes. The following principles define what we mean by responsible AI at SERGENTA AI.

How Principles Become Practice

Principles are only meaningful if they shape behaviour. We embed our AI principles into our development process through several mechanisms: AI risk assessments during design, model governance reviews before deployment, human oversight controls in production, audit logging for accountability and continuous monitoring for reliability and safety.

When a proposed AI feature conflicts with one of our principles, we do not quietly override the principle. We either redesign the feature to align with the principle, or we do not build it. This is what it means for principles to be commitments rather than slogans.

Governance is what separates responsible AI from aspirational AI. Without governance, principles are words on a page — well-intentioned but unenforced. With governance, principles become practices — embedded in the way we design, build, deploy and operate AI systems.

Governance Framework

Our AI governance framework defines who is responsible for what, how decisions are made, what controls are required and how compliance is verified. It covers the full AI lifecycle — from initial research and design through development, testing, deployment and ongoing operation. The framework is proportionate: more powerful or consequential AI systems receive more governance attention.

Decision-Making

AI-related decisions — what to build, which models to use, how to deploy, what safeguards to implement — are made through structured processes that involve the right stakeholders. Significant AI decisions are reviewed by people with expertise in the relevant domain, not made unilaterally by developers or product managers.

Internal Reviews

AI systems undergo internal reviews at key points in their lifecycle. Design reviews assess proposed AI features for risk, fairness and alignment with our principles. Pre-deployment reviews verify that required controls are in place. Post-deployment reviews evaluate whether the system is performing as expected and whether any unintended consequences have emerged.

Risk Management and Documentation

AI risk management is integrated into our broader risk management framework. AI-specific risks — bias, hallucination, prompt injection, data leakage — are identified, assessed and mitigated through structured processes. Documentation captures what risks were identified, what mitigations were implemented and what residual risks remain.

Operational Controls

In production, AI systems operate under operational controls that limit their behaviour. Least-privilege access ensures AI agents can only access what they need. Human approval gates ensure consequential actions are reviewed. Audit logging ensures that AI activities are traceable. Rate limiting and output validation prevent abuse.

Policy Development

Our AI policies are living documents. As AI technology evolves, as we learn from experience and as regulatory expectations develop, our policies are updated. Policy development is collaborative — involving engineers, security specialists, compliance professionals and leadership — ensuring that policies are both technically sound and practically implementable.

The concept of human oversight is simple: AI systems should not take consequential actions without human review and approval. In practice, this requires deliberate design — building the right checkpoints, approval flows and escalation paths into every AI system we deploy.

Human-in-the-Loop

Human-in-the-loop means that a human being is actively involved in the decisions an AI system influences. AI can research, analyse, draft and recommend — but the human reviews, validates and approves before action is taken. This is not about slowing things down; it is about ensuring that judgement is applied where judgement matters.

Review Processes

AI-generated outputs are subject to review before they are acted upon. The depth of review depends on the consequences of the action. A draft email might receive a quick scan. A data analysis that informs a business decision receives careful scrutiny. An action that affects a person's rights, data or safety receives formal review and explicit approval.

Escalation

When AI encounters situations it cannot handle — ambiguous inputs, edge cases, potential policy violations — it escalates to a human rather than guessing. Escalation paths are designed into our AI systems, ensuring that uncertainty triggers human involvement rather than autonomous action.

Decision Support, Not Decision Replacement

AI in our platform is designed as decision support — providing information, analysis and recommendations that help humans make better decisions. It is not designed as decision replacement — taking decisions out of human hands entirely. The distinction matters: decision support amplifies human capability; decision replacement removes human agency.

Enterprise Approvals

For enterprise customers, human oversight extends to organisational controls. Enterprises can configure which AI capabilities are available to which users, set approval thresholds for different types of actions and define escalation paths that match their organisational structure. This ensures that AI operates within the governance framework that each enterprise defines for itself.

AI systems can feel like black boxes — inputs go in, outputs come out, and the process in between is opaque. This opacity erodes trust. If people cannot understand how an AI system arrived at a particular output, they cannot evaluate whether to trust it, question it or override it. Transparency and explainability are the antidotes to this opacity.

Explainable AI

Explainable AI is the practice of designing AI systems whose behaviour can be understood by humans. Full explainability is technically challenging — particularly for large language models, whose internal processes involve billions of parameters. But while we may not be able to explain every computational step, we can provide meaningful transparency about what the system does, what data it uses and how its outputs should be interpreted.

Understanding AI Outputs

AI outputs should be presented in ways that help humans understand them. This includes showing the sources or references an AI used, indicating confidence levels where appropriate, flagging uncertain or speculative content and making it clear when content is AI-generated. Users should never be in a position where they cannot tell whether they are interacting with AI or a human.

Confidence and Uncertainty

AI systems do not always know when they are wrong. A large language model can produce a confident-sounding answer that is factually incorrect — a phenomenon known as hallucination. We design our AI systems to be honest about uncertainty where possible, and we educate users about the limitations of AI outputs. Treating AI outputs as authoritative without verification is a risk; treating them as useful inputs that require human review is responsible.

Traceability

Traceability goes beyond logging individual activities. It means being able to trace the full chain of events — from the user's request, through the AI's processing, to the final output and the human's decision. This end-to-end traceability is essential for investigating issues, demonstrating compliance and building trust with customers who need to understand how our AI operates.

Documentation

We document our AI systems — their purpose, their capabilities, their limitations, the data they access, the controls that constrain them and the oversight mechanisms in place. This documentation is maintained throughout the AI lifecycle and is available to relevant stakeholders, including enterprise customers who need to evaluate our AI practices.

Enterprise Reporting

Enterprise customers need visibility into how AI is used in their environment. We provide reporting that shows which AI features are active, what actions AI agents have taken and what human decisions have been made. This transparency enables enterprises to govern their own AI usage and demonstrate compliance with their internal policies and external obligations.

AI systems behave differently from traditional software. They can produce unexpected outputs, be influenced by biases, be manipulated through novel attack vectors and make decisions that affect people in ways that are difficult to predict. These characteristics require a risk management approach that is specifically designed for AI.

AI Risk Identification

We identify AI-specific risks through structured assessments, threat modelling, testing and ongoing monitoring. These risks include both technical risks (model behaviour, security vulnerabilities) and societal risks (bias, misinformation, privacy impact).

Risk Assessment and Mitigation

Identified risks are assessed for likelihood and impact, then prioritised. Mitigation strategies are developed and implemented — which may include technical controls (input validation, output filtering, access restrictions), process controls (human review, approval gates) or policy controls (usage restrictions, documentation requirements).

Continuous Evaluation

AI risk management is not a one-time activity. AI systems change over time — models are updated, data shifts, usage patterns evolve and new threats emerge. We continuously evaluate our AI systems, monitoring their behaviour, assessing new risks and adapting our controls. What was safe yesterday may not be safe tomorrow, and our risk management process reflects this reality.

Bias and Fairness

Bias is one of the most significant risks in AI. AI systems learn from data, and if that data reflects historical biases, the AI will reflect and potentially amplify them. We assess our AI systems for potential bias, we design them to avoid discriminatory outputs and we maintain human oversight for decisions that could affect people differently based on protected characteristics. We do not claim to have solved bias — it is an ongoing challenge — but we take it seriously and work to mitigate it.

Hallucinations and Accuracy

Large language models can produce outputs that are fluent, confident and wrong. This phenomenon — hallucination — is a fundamental characteristic of current AI technology, not a bug that can be fully fixed. We design our systems with awareness of this limitation: outputs are reviewed by humans before consequential action, uncertainty is communicated where possible and users are educated about the need to verify AI-generated content.

AI is a new security surface. Traditional application security controls — authentication, authorisation, input validation, encryption — are necessary but no longer sufficient when AI is involved. AI systems can be manipulated through prompt injection, tricked into revealing sensitive information, coerced into producing harmful outputs or exploited to access data they should not see.

AI Security

AI security encompasses the controls and practices that protect AI systems from manipulation, misuse and exploitation. It includes technical controls (input validation, output filtering, access restrictions), process controls (human oversight, monitoring) and governance controls (documentation, risk assessment, model governance).

Model Protection

AI models are valuable assets. They can be targets for theft, reverse engineering or adversarial manipulation. We protect our models through access controls, monitoring and operational practices that limit exposure to unauthorised access or tampering.

Data Isolation

One of the most critical AI security controls is data isolation. Customer data must not leak from one customer's AI context into another's. We design our AI systems with strict isolation — AI agents access only the data they are explicitly authorised to access, and conversation contexts are separated between users and organisations. Customer data is not used to train shared models.

Identity and Access

AI agents have identities — and like all identities, they must be authenticated, authorised and monitored. AI agents operate with least-privilege access, can only perform actions explicitly granted to them and are subject to the same access management lifecycle as human users. Their activities are logged and auditable.

Monitoring

AI systems are monitored in production for security-relevant events. Unusual access patterns, unexpected outputs, anomalous behaviour and potential manipulation attempts are detected through automated monitoring and investigated by our team. This monitoring is essential because AI systems can behave in ways that were not anticipated during testing.

Responsible Deployment

How AI is deployed matters as much as how it is built. We deploy AI systems gradually, monitor their behaviour in production and have the ability to roll back or disable features if issues are detected. Responsible deployment means being prepared for the unexpected — having the controls, monitoring and processes needed to respond when AI systems do not behave as intended.

• AI agents operate with least-privilege access — no implicit permissions
• Customer data is isolated — never used to train shared models
• Consequential actions require human approval
• AI activities are comprehensively logged and auditable
• AI systems are monitored for anomalous behaviour in production
• Output validation and filtering prevent harmful content
• Rollback and disable capabilities exist for every AI feature

Every AI system we build passes through a structured lifecycle. Each phase has defined activities, deliverables and governance checkpoints. This lifecycle ensures that responsible AI is not an afterthought but a property of the development process itself.

Research: Is AI the Right Tool?

The first question in any AI project is whether AI is the right solution. AI introduces complexity, risk and cost that traditional software does not. If a problem can be solved effectively without AI, it should be. When AI is the right tool, research focuses on understanding the problem space, evaluating available approaches and identifying the risks and trade-offs involved.

Design: Building in Responsibility

During design, responsible AI is built in — not bolted on. Access patterns are designed with least privilege. Human oversight checkpoints are defined. Data flows are mapped and isolation boundaries are established. Threat modelling identifies potential attack vectors. Risk assessments evaluate potential harms. These activities shape the design before a single line of code is written.

Testing and Validation

Testing and validation are where responsible AI is verified. Does the system do what it is supposed to do? Does it avoid doing what it is not supposed to do? Are the controls effective? Are there biases in the outputs? Are there edge cases that produce harmful results? These questions are answered through structured testing before the system is deployed.

Deployment and Monitoring

Deployment is gradual and monitored. We do not deploy AI systems and assume they will work perfectly. We watch them — checking for unexpected behaviour, performance issues and security anomalies. If problems are detected, we can roll back, disable or adjust the system.

The field of AI governance is in its early stages. Best practices are still being established, regulations are still being written and the technology itself continues to advance at a rapid pace. We approach AI governance with humility — recognising that what seems sufficient today may need to evolve tomorrow.

International Standards

International standards for AI governance are being developed by organisations around the world. We monitor these developments closely and intend to align our practices with recognised standards as they mature. We do not claim compliance with standards that are still being defined — but we design our practices with awareness of the direction in which standards are moving.

Emerging AI Regulations

The regulatory landscape for AI is evolving rapidly. The EU AI Act is the most prominent example, but other jurisdictions are developing their own AI regulations — each with different approaches, requirements and timelines. We monitor these developments and adapt our practices accordingly. We do not claim compliance with regulations that are still being implemented, but we design our systems with principles that align with the direction of regulatory evolution.

Enterprise AI Maturity

Enterprise customers are at different stages of their own AI maturity. Some are just beginning to explore AI; others have established AI governance programmes. We meet customers where they are — providing the documentation, controls and transparency they need to evaluate and trust our AI, regardless of their own maturity level. As our customers' AI governance matures, we will evolve our capabilities to meet their growing expectations.

Continuous Governance

Governance is not a project with an end date. It is a continuous practice that evolves alongside the technology it governs. We regularly review our AI governance framework, incorporating lessons from experience, developments in the field and feedback from customers and stakeholders. What we describe on this page is our current state — not our final state.

Future Certifications

As AI-specific certification frameworks emerge and mature, we will evaluate whether they are appropriate for our organisation. We intend to pursue relevant certifications as part of our long-term governance roadmap — but only when the frameworks are well-established and our practices are genuinely ready for assessment. We will not pursue certifications for marketing value; we will pursue them when they represent a meaningful validation of our practices.

Ongoing Improvement

Above all, we are committed to ongoing improvement. AI technology will continue to evolve. New risks will emerge. New mitigations will be developed. New regulations will be enacted. Our commitment is not to a static set of practices but to a process of continuous learning, adaptation and improvement. Responsible AI is not a destination — it is a practice.