AI Security - Principal Security Engineer at Zendesk
Última atualização em 22 de abril de 2026
How grounding, guardrails, and human-in-the-loop drive trustworthy automation.
The AI trust gap
Many organizations are hitting a wall with AI adoption, and it isn't because of a lack of speed. According to the latest McKinsey Global Survey on AI, inaccuracy is the top AI-related risk that organizations have experienced over the past year - and the one they are most focused on mitigating.
Concerns about AI reliability are a true blocker for adoption - and no wonder. If your customers don’t trust the answers they’re getting, they won't use your AI agent. A fast response is a business risk if it is a hallucination - a confident but entirely invented answer. In this new era of autonomous service, the goal isn't just to automate support - it’s to be reliably helpful at scale.
Grounded accuracy–the reliability baseline
To solve the risk of inaccuracy, you start with grounding. Imagine an AI agent is a brilliant librarian who has read every book ever written. They are incredibly knowledgeable, but they don't know which book contains your current company policies. Without a guide, they might quote a rumor from a tabloid as confidently as a fact from an encyclopedia.
Grounding, through Retrieval Augmented Generation (RAG), is like handing that librarian a single, verified handbook and saying: “Only answer using the information in this volume.” This ensures the AI prioritizes your official customer FAQs and internal handbooks over the noise of the public internet. By anchoring the AI's speed to your specific, verified facts, grounding prevents the system from providing unverified information.
From grounded accuracy to true AI reliability
Grounding however, is just one element that ensures AI reliability. , provides a response that is relevant and moves the customer towards resolution, while minimizing data exposure.
At Zendesk, we believe a reliable response is defined by:
Intentional accuracy: The system doesn't just find a random fact; it accurately maps the user’s intent to the specific policy or account information relevant to them.
Verified integrity: Every generated answer is strictly supported by your approved knowledge sources, without the AI making assumptions or guesses.
Authorized exposure: The system respects privacy by ensuring an answer only reaches a user who is authorized to see that specific data.
Safe fallbacks: The AI always follows the company's guidelines, and when the situation is ambiguous, it falls back and hands it off to a human agent.
Continuous improvement: The system doesn't stay static; it uses what we call a Resolution Learning Loop™ to ensure every interaction is smarter than the last.
Reliability in practice: How we deliver on the promise
Trust isn’t a happy accident - it’s the result of how we treat these five elements in the real world. This is how we ensure reliable AI in practice.
Engineering AI-ready knowledge
Reliability starts with organized, regularly updated knowledge sources. You can't have a reliable AI agent if your information is hidden in silos or trapped in outdated documents. The good news is Zendesk makes it easier to build AI-ready knowledge that is accurate, connected, and structured to be searchable and retrievable. Using Zendesk knowledge builder, service leaders can quickly build and standardize content by turning trending issues into clear, consistent articles with a repeatable structure. Then, knowledge connectors bring content from across your existing systems into a single searchable network, so your AI has the full picture without relying on scattered sources.This ensures your AI is always pulling from a clean, verified source of truth that is optimized for machine retrieval.
Security and Brand Guardrails
An accurate answer is not reliable if it reveals sensitive data to the wrong person. This is why we apply strict access limits: the bot can access only what the specific end-user is permitted to see. To further protect your brand, we use prompt shielding to block malicious inputs and data sanitation to remove personally identifiable information (PII) before it ever reaches the model.
Safe Fallbacks: The importance of saying “I don’t know”
A trustworthy AI system knows its limits. A smart AI strategy involves designing safe fallbacks, such as bounded guidance or immediate human hand-offs, when AI isn't sure of the correct answer. Using these fallbacks demonstrates that you take risks seriously. It reassures customers and employees that the system is designed to admit a limit rather than guess an answer.
The Resolution Learning Loop - improving reliability over time
To drive continuous improvement, Zendesk uses a cycle where AI and human service teams work together to refine outcomes. This isn't a one-time setup; it’s an ongoing practice of refinement where every interaction generates data to improve the next:
Learning from outcomes: The loop analyzes interaction data to identify knowledge gaps or outdated procedures. It provides recommendations to update your help center articles and SOPs so the AI never has to guess based on old information.
Guidance through Copilots: Agent and Admin Copilots provide real-time guidance based on ticket and knowledge context. As humans use these insights to update articles or refine workflows (via Action Builder), the AI’s performance becomes faster and more consistent.
Constant Quality Assurance: Zendesk QA acts as a digital safety net, automatically reviewing interaction data across AI and human agents to surface exactly where quality, tone, or accuracy can be improved.
The Outcome: Reliability you can scale
When these five elements work together, AI reliability stops being a technical checkbox and starts being a business strategy. You no longer have to choose between the speed of automation and the safety of your brand. By creating a system that is grounded in facts, protected by guardrails, and constantly refined by human expertise, you build the foundation necessary to deploy AI across every customer touchpoint with total confidence.
AI supported by this connected ecosystem stops being an experiment and starts being an engine for growth. You will see more issues resolved automatically, fewer repeat contacts, and a support team that is empowered to focus on the complex, human side of service - giving customers the right help the first time, every time.
Béatrice Moissinac
AI Security - Principal Security Engineer at Zendesk
Béatrice is the Principal Security Engineer for AI Security at Zendesk, where she focuses on applying AI research to cybersecurity, trust, and safety—exploring how to build AI applications and products securely and responsibly. She previously held positions at IBM, Credit Suisse, and Okta and has two Master's degrees and a PhD in Computer Science. In her free time, Béatrice enjoys long-distance running, camping and backpacking, and building Legos. Her favorite algorithm is Dynamic Time Warping.
