Hernan Huwyler
AI Risk & Compliance Leader| Academic Director IE | Decision-Making | AI Threats and Controls | Responsible Technology | Quantitative Risk | Compliance | Certified AI Architect
Copenhagen, Denmark
Actions
Hernan Huwyler is a Global Risk Management Executive and the Academic Director at IE Law and Business School (Executive Education). With a dual background as an MBA and CPA, Hernan specializes in the intersection of Law, Technology, and Quantitative Risk.
He focuses on "translating" complex regulatory frameworks (EU AI Act, GDPR, ISO 42001) into actionable architectures for engineering and business teams. Unlike traditional compliance officers, Hernan advocates for data-driven risk management, moving beyond "heatmaps" to utilize Monte Carlo simulations, convolution models, and Python-based risk quantification.
Hernan has led Risk, Control, and Audit functions for major multinational corporations across Europe and the Americas. He writes and speaks extensively on Adversarial AI, Algorithmic Auditing, and the financial quantification of Cyber Risk. Past keynote venues include Risk Awareness Conferences (>2,000 risk managers), IIA and ISACA Summits, ProcureCom Europe, IKN Madrid, IE University, and closed-door sessions for international organizations and national regulators. Hernan’s talks consistently rank in the top 5% of conference feedback scores because he makes complex security requirements, business needs and regulations feel like a competitive advantage.
Speaking Topics:
AI Governance: Implementing ISO 42001 and navigating the EU AI Act.
Risk Quantification: Why "Impact x Probability" fails and how to use convolution models.
Adversarial AI: Balancing transparency with security to prevent data poisoning.
AI ROI: Risk-adjusted financial modeling for Generative AI deployments.
AI Architecture: Decision-making and planning on how to plan AI developments.
Ready-to-Submit Talk Proposals:
Session 1: The Technical/Quant Talk
Title: Kill the Heatmap: Quantifying AI Risk with Python & Monte Carlo Simulations
Abstract: Most organizations manage AI risk using colorful 5x5 "heatmaps." This is mathematically flawed and dangerous for high-stakes AI deployment. In this talk, we will ditch the gut-check methodology. I will demonstrate how to use convolution and Monte Carlo simulations to model the "tail risk" of AI hallucinations and regulatory fines. You will leave with a method to calculate the actual financial exposure of your models—no expensive GRC software required, just math and logic.
Session 2: The Governance/Security Talk
Title: The Transparency Trap: How to Audit AI Without Creating Vulnerabilities
Abstract: New standards like ISO/IEC 12792 demand AI transparency, but in the age of Adversarial Machine Learning, "oversharing" is a security vector. If you reveal your full architecture and training sources, you invite data poisoning and model inversion attacks. This session outlines a "Tiered Disclosure Architecture." We will discuss how to route specific data to auditors, regulators, and users differently, ensuring compliance with the EU AI Act without handing the keys to bad actors.
Session 3: The Business/Strategy Talk
Title: The Real Cost of Intelligence: A Risk-Adjusted Framework for AI ROI
Abstract: Companies are deploying Generative AI based on productivity hype, but they are ignoring the "OpEx Iceberg"—model drift, inference costs, and liability reserves. This talk presents a financial framework for calculating the true ROI of AI. We will cover how to factor in the "Risk Delta" (new risks vs. automated risks), calculate the Levelized Cost of AI (LCOAI), and determine when to build vs. buy based on process uniqueness. A session for leaders who need to justify the cloud bill to the CFO.
Area of Expertise
Topics
From Raw Data to Reliable AI:Building Data Models, ML and Deep Learning Pipelines That Actually Work
Artificial intelligence does not start with the model. It starts with the data.
This session takes a practical journey from raw enterprise data to production-ready Data Science and AI, connecting Data Science, Data Modelling, Python, R, Machine Learning, Artificial Intelligence, Data Engineering and Deep Learning.
The focus is on the problems that determine whether an AI initiative produces reliable results or becomes an expensive experiment: data that is unsuitable for the intended use case, biased or unrepresentative datasets, compromised data integrity, leakage between training and test data, and the difficult balance between privacy and realistic data.
Participants will see how to build the analytical chain from data → data model → engineering pipeline → statistical analysis → machine learning → AI → deep learning → validation → production monitoring.
The session will start with data fitness. A dataset can be technically clean and still be unsuitable for the model or decision it is supposed to support. We will examine how to define data quality in relation to the use case, including completeness, accuracy, consistency, timeliness, label quality and representativeness. Participants will also learn how to document provenance, assumptions, transformations and known limitations.
We will then move into data modelling and engineering, showing why the structure and lineage of data matter before any Python or R model is developed. The session will cover schema validation, entity relationships, enrichment, feature construction, versioning and reproducible datasets. The objective is to create a reliable analytical foundation rather than simply produce a larger dataset.
The practical modelling section connects Python and R to the complete Data Science workflow. Python will be used to illustrate data preparation, exploratory analysis, feature engineering, machine-learning workflows and model evaluation, while R provides a complementary environment for statistical analysis, modelling and validation. The emphasis is on choosing the right method for the problem rather than treating a particular programming language or algorithm as the solution.
Participants will then move from traditional statistical and machine-learning approaches into Artificial Intelligence and Deep Learning. The session will explain when conventional models can be preferable to neural networks, when deep learning adds value, and how model complexity affects data requirements, interpretability, validation and operational risk.
A central theme is data integrity. Training data can be incomplete, manipulated, poorly labelled or contaminated. The session will cover dataset provenance, version control, restricted write access, reproducible snapshots, anomaly detection and controls for third-party or externally sourced data. The goal is to make it possible to answer a fundamental production question: Which data produced this model, and can we reproduce it?
We will also address bias and representativeness. Participants will learn how to examine datasets by relevant population or operational segments, identify gaps, evaluate label agreement and test whether model performance changes materially between groups. For AI systems subject to the EU AI Act, these practices become especially important because Article 10 addresses the relevance, representativeness, completeness, quality and bias of datasets used for high-risk AI systems.
Another practical problem is data leakage. A model can appear highly accurate because information from the future, the test set or closely related records has unintentionally entered the training process. We will demonstrate how to establish training, validation and test sets correctly, detect duplicates and near-duplicates, use time- or entity-based splits where appropriate, and maintain a genuinely held-out test set.
Finally, the session examines synthetic data. Synthetic data can help overcome limited or sensitive datasets, but it can also reproduce bias, lose important relationships or amplify errors through repeated model-generated training. Participants will learn how to evaluate synthetic data through three practical dimensions: fidelity, utility and privacy.
The session is designed for professionals who want to understand what happens between an AI idea and a reliable production model. It combines technical concepts with practical controls and examples that Data Scientists, AI Architects, Data Engineers and technology leaders can apply directly.
From Agent Sprawl to Agent Control: Governing Copilot Studio Access, Costs and Production Readines
AI agents are moving from experiments to operational systems, but governance is often lagging behind deployment. In Microsoft environments, Chief AI Officers and AI Architects increasingly have to answer practical questions that traditional AI governance frameworks do not solve: Who can access an agent and its connectors? What data can the agent reach? Who owns it when the original maker leaves? How much does experimentation cost? When is an agent actually ready for production?
This session provides a practical governance model for Microsoft Copilot Studio and Agent Builder, focused on the controls that matter after an agent has been built.
We will examine four governance problems that are becoming critical in 2027: fragmented administration across Microsoft 365 and Power Platform, uncontrolled connector and knowledge-source access, agent sprawl and underused or ownerless agents, and poorly controlled Copilot Credit consumption. Microsoft’s current guidance confirms that credit consumption can begin during building, preview and evaluation, making governance necessary before production. It also recommends controls at tenant, environment, agent and operating-process levels.
The session will then turn these problems into practical controls. Participants will learn how to establish least-privilege access to external connectors and tools, distinguish environment-level governance from agent-level controls, assign accountable ownership, administer user and group access, monitor agent usage and consumption, and design a controlled path from development to testing and production.
A major focus will be publishing validation. An approval to publish does not by itself demonstrate that an agent is safe, useful or production-ready. We will examine a practical validation gate covering identity and access, data sources, connector permissions, expected behavior, evaluation results, failure modes, cost, logging, ownership, business accountability and rollback or retirement criteria.
The session will also cover common administration traps in Agent Builder, including the difference between owners and chat users, group-based access, organizational sharing and the relationship between an agent's access and the underlying SharePoint data. Microsoft currently supports multiple individual owners for Agent Builder agents, while groups cannot be assigned as owners.
Finally, we will address cost governance. Participants will see how to combine environment-level capacity controls with agent-level limits, notifications and hard stops, and why monitoring consumption only after deployment is too late. Microsoft currently provides agent-level monthly limits and hard-stop controls through the Power Platform admin center.
The session is delivered by Prof. Hernan Huwyler, AI Governance, Risk & Compliance specialist and Claude Certified Architect – Professional (CCAR-P), issued by Anthropic. The CCAR-P credential is Anthropic's professional-level architect certification, covering enterprise-scale Claude deployments, integration architecture, evaluation, governance and responsible deployment.
The emphasis throughout is practical: what to configure, what to monitor, what to validate, who should own each control, and which governance failures are likely to appear when organizations scale from a handful of agents to hundreds.
Beyond Copilot Hype: Quantifying AI Risk & ROI in Dynamics 365 F&SCM
AI is embedding itself into Dynamics 365 Finance & Supply Chain Management from forecasting and cash-flow predictions to procurement scoring and autonomous agents. Yet most project managers, consultants and partners are still explaining AI on productivity narratives, not on quantified business value or risk-adjusted return. This session challenges the feature-first approach. We will examine how AI changes the financial, operational, and regulatory risk profile of ERP implementations, and why traditional risk heatmaps fail in enterprise AI deployments.
You will learn a practical framework to:
• Calculate the Risk Delta (new AI risk introduced vs. legacy risk reduced)
• Model the levelized cost of AI, including monitoring, drift, inference, and governance overhead
• Align EU AI Act and ISO 42001 requirements to D365 architecture without slowing innovation
• Communicate AI exposure and ROI credibly to CFOs and boards
Using real-world enterprise examples, we will translate governance into architecture, and architecture into financial impact.
If you are a Dynamics partner navigating Copilot, AI agents, or advanced automation, this session will help you price smarter, deploy safer, and turn compliance into competitive advantage.
From AI Hype to AI Accountability: Risk, Assessments and Controls for Responsible Adoption
Organizations are deploying AI into production systems that affect credit decisions, hiring, fraud detection, and customer outcomes. Most of them have a responsible AI policy. Very few have controls that a regulator, an auditor, or a plaintiff's attorney would accept as evidence.
This session closes the gap between what an AI governance policy says and what systems can actually prove. In 30 minutes, I will walk through the three failures I see most often in AI risk programs, superficial risk assessments that cannot be audited, ethical principles that cannot be enforced technically, and ownership structures where nobody is actually accountable and show participants what to build instead.
Participants will see how to replace subjective analysis with quantitative risk models that translate bias exposure, model drift, and regulatory penalty scenarios into financial terms a board can act on. They will learn why system-level controls, middleware that validates every output against policy rules before it reaches the user, are more defensible than model-level alignment that jailbreak research consistently shows can be bypassed.
This is not a session about principles. It is a session about controls, numbers, and evidence. If participants are responsible for AI systems in production and you cannot answer five questions , what is the metric, what is the threshold, who owns it, how often is it tested, and what happens when it is breached, this session will show you how to get there.
Participants will leave withpractical insights they can apply on Monday morning, whether you are a CAIO building a governance program, a security architect designing AI threat models, or an engineering leader trying to figure out what the EU AI Act actually requires from your deployment pipeline.
Hernan Huwyler
AI Risk & Compliance Leader| Academic Director IE | Decision-Making | AI Threats and Controls | Responsible Technology | Quantitative Risk | Compliance | Certified AI Architect
Copenhagen, Denmark
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top