Download to read offline
Uploaded byDataScienceConferenc1
[DSC DACH 25] Governance of Agentic AI - Nescho Topalov.pptx
Agentic AI introduces a new generation of intelligent systems capable of autonomous decision-making and execution, unlocking vast opportunities while raising critical questions about risk, compliance, and ethical responsibility. This session explores how organizations can effectively evaluate, monitor, and govern agentic AI across its lifecycle to ensure transparency, accountability, and trust. Attendees will gain insights into practical frameworks and tools for scaling agentic AI responsibly — balancing innovation with control.
More Related Content
Similar to [DSC DACH 25] Governance of Agentic AI - Nescho Topalov.pptx
More from DataScienceConferenc1
![[DSC DACH 25] Ali Mokh - Symbolic Reasoning in telecom oriented LLMs.pdf](https://cdn.slidesharecdn.com/ss_thumbnails/dscdach25alimokh-symbolicreasoningintelecomorientedllms-251027100200-99093bc8-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Brinnae Bent - Hacking the Blackbox.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/brinnaebenthackingtheblackbox1-251024080947-002ebd1f-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Zrinka Puljiz - Importance of data at the time of LLMs.pdf](https://cdn.slidesharecdn.com/ss_thumbnails/importanceofdataatthetimeofllms2-251024080620-db0b03b5-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Jacqueline Berger - Beyond the Hype: The Real Story of Building...](https://cdn.slidesharecdn.com/ss_thumbnails/dsc-dach-2025jacquelineberger-251024080522-7625155a-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Yudan Lin_The Human - AI Paradox and Effective Fix.pdf](https://cdn.slidesharecdn.com/ss_thumbnails/dscdach2025yudanlinthehuman-aiparadoxandeffectivefix-251024080438-f3a7a9e1-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] From Data to Decisions: Building Responsible AI for the Future ...](https://cdn.slidesharecdn.com/ss_thumbnails/dscdach2025shpresentation-251024080322-abe65846-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Emmanuel Goffi - Moving from Compliance to Post-Compliance Ethi...](https://cdn.slidesharecdn.com/ss_thumbnails/goffi2025-movingfromcompliancetopost-complianceethicsakeychangeforcompanies-251024080222-644893cc-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] AI in Action_ From Industrial Data to Sustainable Impact_Ashraf...](https://cdn.slidesharecdn.com/ss_thumbnails/aiinactionfromindustrialdatatosustainableimpactashrafabushady20251016-251024080148-05e79a02-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Uros Miletic - Common pitfalls in AI adoption.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/commonpitfallsinaiadoption-251024075719-4a9d71ea-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 24] Sofija Pervulov - Building up the Bosch Semantic Data Lake.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/dscsofijapervulovboschupdated-251023082520-942d23c1-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Dr. Jochen Kokemueller.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/251015dsckokemueller-251022092212-44cf2599-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Mario Meir-Huber Demystifying Data Products: From Buzzword to P...](https://cdn.slidesharecdn.com/ss_thumbnails/demdapro-251022092046-55c8d73b-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Stefan Stricker-Agentic_BI.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/stefanstricker-agenticbi-251022091858-f3944612-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Nuno Antonio_BigDataDrivenStrategies.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/nunoantoniobigdatadrivenstrategies-251022091406-23769848-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Ivan Vukasinovic $20M Datapipeline Built by AI in 48 hours.pdf](https://cdn.slidesharecdn.com/ss_thumbnails/ivanvukasinovic20mdatapipelinebuiltbyaiin48hours-251022090438-b9651e46-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC DACH 25] Davor Gasparac - Unscrewing Agile.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/gasparac-unscrewingagile-251022090222-a69d34b8-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 24] Anastasia Shapedko - How Alice, our intelligent personal assi...](https://cdn.slidesharecdn.com/ss_thumbnails/anastasiashapedkoyandexaliceforpeoplewithspeechimpairments-250218225516-7720cf27-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 24] Joy Chatterjee - Balancing Personalization and Experimentatio...](https://cdn.slidesharecdn.com/ss_thumbnails/joy-chatterjee-ab-testing-friday-final-250218225454-d92d0731-thumbnail.jpg?width=640&height=640&fit=bounds)
![[DSC Europe 24] Pratul Chakravarty - Personalized Insights and Engagements us...](https://cdn.slidesharecdn.com/ss_thumbnails/pratulchakravarty-personalizedengagements-250218225450-d3180df3-thumbnail.jpg?width=640&height=640&fit=bounds)
[DSC DACH 25] Governance of Agentic AI - Nescho Topalov.pptx
- 1.
- 2. The DSC Context -DSC DACH = cross-industry data & AI minds - Developers Decision-makers → - 2025: Agentic AI & Responsible Scaling
- 3. The Difference AIAgents - Softwareentitiesthatperformtasksautonomouslybasedonprogrammedlogicorgoals. -Focus:Execution—theyfollowrulesorworkflowsdefinedbyhumans. AgenticAI - AnewgenerationofAIagentswithreasoning,memory,andself-direction. - Focus:Autonomyanddecision-making—theycanplan,adapt,andpursuegoalswith minimalhumaninput.
- 4. The Opportunity and theParadox “By2028,one-thirdofGenAIinteractionswillinvolveautonomousagents.”—Gartner,2024 - Automation & augmentation potential - But: ethical, operational, and reputational risks grow
- 5. Why Governance Matters - Hallucinations -Data leakage - Model drift & bias - Cost explosions “AnAIagentwithoutgovernanceisaliability.”—IBMAIEthicsBoard,2025
- 6. Governance Pillars - Lifecyclegovernance - Risk management - Regulatory compliance
- 7. From Chaos to Control Agents Orchestrator Data&Tools Everyagentshouldliveinsideagoverned ecosystem—orchestrated,monitored,and auditable.That’sthefoundationforresponsible autonomy. Centralized AI lifecycle governance Manage, monitor and govern any AI: model, app, agent or tool; across IBM and 3rd party like OpenAI, AWS, Azure, GCP, Meta, etc. Proactive AI risk and security management Proactively detect and mitigate AI risks, evaluate AI assets, and secure AI deployments with Guardium AI security Trustworthy and dynamic compliance Manage AI for safety and transparency with our regulatory library, automation and industry standards Platform agnostic: Govern any AI Agent, deployed anywhere Ariba
- 8. Agentic AI Risks and Challenges Risks •Misaligned actions • Discriminatory actions • Over- or under-reliance • Unauthorized use • Exploit trust mismatch • Unexplainable or untraceable actions • Lack of transparency Risks • Unsupervised autonomy • Data bias • Redundant actions • Attack on AI agent’s external resource • Tool choice hallucination • Sharing IP/PI/confidential information Challenges • Reproducibility • Traceability • Attack surface expansion • Harmful and irreversible consequences Challenges • Evaluation • Accountability • Compliance • Mitigation and maintenance • Infinite feedback loops • Shared model pitfalls New Emerging areas intrinsic to agentic AI Amplified Known areas intensified by agentic AI
- 9. Key lifecycle governanceactivities For agentic systems Experimentation tracking Track agentic app variants and compare results to inform which to push to production Agentic system metrics, monitoring and alerts Oversee elements such as hallucination, answer relevance, and system drift in production and development Traceability Help developers debug agentic app by tracing each step of the user interaction and agent processing Cataloging of agentic AI applications Single consolidated view of all in development and use
- 10.
- 11.
- 12.
- 13.
- 14. Example: The BIAgent - Conversational business insights, grounded in governed data - Transparent queries, explainable answers - Built on watsonx
- 16. Example: Orchestrating AI Agents -watsonx Orchestrate = create, connect, monitor agents - 400+ ready connectors & tools - Built-in AgentOps for oversight
- 18. The Cost ofIgnoring Governance ⚠️Sandbox success production failure → ⚠️Shadow AI, untracked spend ⚠️Compliance exposure
- 19. Our Approach 1. Co-designwith governance from day one 2. Pilot safely scale confidently → 3. Govern any model, any cloud 4. Combine people, process, and platform
- 20. You Are Welcome:11.11.2025, Vienna
- 21. Visit our boothto continue the conversation! Connect with me: Nescho Topalov CEO & Co-founder Erdbergstraße 52-60/3/20-21, 1190 Vienna / Austria [email protected] www.topideas.digital
- 23. watsonxOrchestrate watches Asanafor new/updatedtasks;readstitle, description,customfields,attachments, andduedates. Pullstherelevant Bynder policiesand brandassets(filteredbymarket,product, channel,language). Runs hybridchecks:hardrulesfor must/forbiddenitems+ watsonx.ai RAG onretrievedpolicypassages. Drafts targetedquestions tothe submitter;postsinAsana auto-rechecks onreply.
- 24. UpdatesAsanafields: ComplianceStatus, RiskScore,PolicyVersion,LastAICheck; attachesa ComplianceReport. Onapproval,syncsByndermetadata and storesafull audittrail. Security&governance:SSO/OAuth,least- privilegescopes,dataminimization, watsonx.governance lineage/monitoring. Outcome:fasterapprovals,fewer reworks,consistentglobalcompliance, andanauditablerecord.
- 25.
- 26.
- 27.
- 28.
- 29. About Us Partner forPractical AI Innovation Based in Vienna, active across the DACH region and beyond Experience across various industries – Health, Telco, Energy, IT
- 30.
- 31. “By 2028, one-thirdof interactions with generative AI (GenAI) services will use action models and autonomous agents for task completion.” Source: Gartner® Press Release, “Gartner Predicts One-Third of Interactions with GenAI Services Will Use Action Models & Autonomous Agents for Task Completion by 2028” watsonx Orchestrate / © 2024 IBM Corporation 31
- 32. Models Problem-solving Logical thinking Pattern matching Assistants Informationretrieval Prescriptive tasks Single-step processes Agents Multi-step processes Autonomous action-taking Self-correcting A fundamental shift is underway for AI Accelerate AI agent deployment Pre-built Agents Get started quickly with pre-built AI agents powered with business logic and seamless integration to the tools that power your business. Build custom designed agents Custom-built Agents Design, deploy, and manage AI agents with ease using pro-code and low-code options. Manage all agents in one place Multi-agent Orchestration Easily deploy and manage any agent for any task within a simple and unified user experience optimized to scale. IBM Data Platform | © 2025 IBM Corporation
- 33. 33 AI Agents An AIagent is an autonomous system that can use tools and collaborate with other agents to plan and act on tasks. After it acts, the agent reflects on the results of its actions, learning iteratively and refining its approach to better align with its defined objectives.
- 34. Pre-built agents End User Collaborator Agents Orchestrator Agent Agent &Tool Catalog v wxO chat 3rd party UIs IBM Data Platform | © 2025 IBM Corporation v Biz apps & Processes 34 Custom-built agents Multi-agent Orchestration
- 35. AI Agents canaugment AI Assistants to tackle increasingly complex tasks and unlock new value for the Enterprise The tipping point – Generative AI Value to the Enterprise Assistants with AI Agents Fixed flow AI assistants The actual tipping point – AI Agents Fixed flow AI assistants + Gen AI Rigid Fixed Context More Flexible More Contextual Highly Flexible Fully Contextual watsonx Orchestrate / © 2024 IBM Corporation 35
- 36. Studio Discover, create andmanage gen AI and digital automations that combine decisions, tasks, skills, and workflows. AI assistants AI-assisted experiences that are human trained and designed Orchestrator Agent AI Agent supervises and manages how work is executed across assistants, agents, and skills AI Agents in IBM watsonx Orchestrate Supervising Routing Reasoning Planning Custom skills – build new or discovery existing Prebuilt skills and app catalog Customers Employees Subject matter experts Empower customers and employees through simple, intuitive and guided conversations Boost productivity with AI and automation Accelerate time to value with pre-built capabilities or build your own IBM Data & AI / © 2024 IBM Corporation Skills Intelligent task, decision, document and workflow automations augmented by generative AI AI Agents Autonomous AI-driven execution of expert tasks Gen AI catalog powered by watsonx.ai™ models Agent catalog Agents for Sales Agents for HR Agents for Procurement Custom built Agents
- 37. The evolution ofGenerative AI for intelligent business automation Fixed Flow Act as programmed Autonomous Flow Plan and self-correct 37 AI-assisted automation Traditional task automation Autonomous AI orchestration Reasoning Planning Routing Self-correction AI Assistants RAG Gen AI skills IDP Workflow Design Decision logic Process Mining Process Modeling Accelerates and optimizes the design and building of automations AI provides an enhanced user experience and drives higher task completion Allows AI to perform the work reducing the need for human intervention
- 38. AI in anenterprise is like Ice Cream …
- 39. … everybody wants toenjoy the ice cream, but it takes a proper cone to do so Agentic Systems Data Products GenAI AI Governance ML-Ops Data Integration Hybrid Cloud Etc.
- 40. AI building blocksof the future 40 Challenges Compliance Manage AI to meet upcoming safety and transparency regulations and policies worldwide-a “nutrition label” for AI Risk Proactively detect and mitigate risk, monitoring for fairness, bias, drift, and custom metrics Lifecycle Management Manage, monitor and govern AI models from IBM, open-source communities and other model providers (e.g. Meta, Mistral AI) Assistant and Agent Orchestration & Rollout Integration with the existing infrastructure, Self-Service, Automation-integration, maturity level of AI depending on the use case,…
Editor's Notes
- #8 Agentic AI introduces novel and amplified risks and challenges, as well as impacts to society. Enterprises cannot responsibly adopt or scale agentic AI without first considering these unique risks and challenges. It is critical to understand the risks and challenges that are intrinsic to agentic AI and that are intensified by agentic AI in order to effectively govern agentic AI.
- #9 The requirements for agentic governance are similar to AI governance generally, with some key differences in both evaluation and lifecycle governance. Capturing and properly documenting metadata to understand the full scope of agent interactions and decisions is crucial. In experiment stage, comparing agents to drive the highest performing, most impactful ones to production. System metrics help build confidence in the underlying models and agents you are deploying. Traceability can help debug and identify issues throughout an agents process And finally developing a catalog of the various agents your organization has built can help understand what is available for deployment or could easily be modified for another use case.
- #10 Here’s an example of starting the process to onboard an agent –To improve customer service, you want to develop a banking chatbot that can answer customer questions. –From the watsonx governance console, you can create an AI use case to describe your business goal. –An AI use case allows teams to start the governance process early to define requirements and assess risk before the first line of code is written. –As part of the AI use case, complete a risk questionnaire to identify risk dimensions, compliance needs, and applicable AI regulations. –A list of associated risks and a corresponding risk rating appear in the AI use case. Applicable AI regulations are also added –When the AI use case is finished, a notification is sent to the agentic AI developer. This allows the developer to review the business goal and associated risks.
- #11 Here is an example of the developer using the tool library –The risk questionnaire identified prompt injection as a primary concern. –The developer can address this risk during agent development by using a set of watsonx governance components. –A library of components and evaluations is available to the developer to incorporate into the chatbot. –The developer searches for ”prompt injection” to see the available tools. –There are two tools available. One of them uses Granite Guardian and the other uses a two-level prompt injection attack detector. –They can compare the two tools side-by-side and review important metrics including quality, latency, and cost. –And can view the details of each tool and identify where it is used. –The two-level prompt injection attack detector is used in 3 use cases. He can view each of the AI use cases to gain trust that the tool is performing well in production. –With the tool at hand, the developer is ready to build the agent.
- #12 Here we are looking at creating the agent –The developer builds the agentic RAG application in Python notebook. –The application will connect to a specific vector database based on the question received. Insurance questions are directed to the insurance vector database. Credit card questions are directed to a different vector database. –The developer adds the prompt injection attack detector to the agent using a code snippet. –They also adds evaluators for hallucination, context relevance, answer relevance, answer correctness, and faithfulness. –You can see the metric results by testing the RAG application on a single input. -The developer programs the agent to handle errors –Low context relevance implies that the documents returned from the vector database were not useful. –For that situation, they program the agent to reply that the question cannot be answered using the data available. –To handle hallucinations, they programs the agent to reply that the question cannot be answered successfully. –Next, the developer tests the agent on a larger dataset of 100 questions and ground-truth answers from both vector databases. –You can see the average score for latency, context relevance, answer relevance, faithfulness, and more. –The developer notices that the answer relevance score is lower than expected. To investigate, he generates a new visualization to determine root cause. –The visualization indicates that hallucinations are the biggest factor in the low answer relevance score. –They adjust the system prompt and generates a new chart. Answer quality improved, but there are still some hallucinations. –and updates the agent to use a larger LLM and then generates a new chart. –The results look much better. –After running evaluations in the notebook, he compares the results of the three experiments using experiment tracking. –Although the larger LLM had the highest answer relevance, latency, and cost were higher. –They decide to go forward with the second agentic RAG application. –After additional validation, the agent is deployed for use. –While in use, the agent is continuously monitored to ensure that performance and quality are maintained.
- #13 (Full script) Demo Video voiceover: watsonx.governance allows you to create an AI use case describing the business goals for the AI agent. Here, we've created the AI use case, Automated investment assistant. From the AI use case, you can associate the related AI agents to the AI use case. We then associate that with an existing AI agent, Portfolio rebalancer, and add an entry for a new AI agent, Fund withdrawal agent. Development of the new agent must follow the organization's governed workflow which includes an initial risk assessment to identify potential risks early in the process. Once deployed, you can monitor performance and behavior of AI agents and the individual models supporting them using the watsonx.governance runtime monitoring features (30 sec script) Demo Video voiceover: watsonx.governance allows you to create an AI use case describing the business goals for the AI agent. Here, we've created an AI use case which you can associate with new or existing agents. These new agents must follow the organization’s governed workflow which includes initial risk assessment to identify risks early on in the process. Once deployed you can monitor performance and behavior of AI agents within watsonx.governance.
- #25 Wo kommt es zu misverständnissen? Welche Prozesse sind sehr komplex?
- #32 At IBM we are focusing on three pillars to scale our AI agent solutions with our customers. The first one is delivering horizontal multi-agent orchestration through our AI agent orchestrator and a unified user interface that provides access to all your agents, assistants, skills and tools from a single touch point. From there we are focusing on accelerating how quickly enterprises can deploy agents by offering pre-built solutions, like vertical domain agents attached to high ROI enterprise domains in HR, sales, procurement and customer care. This is the best way for enterprises to seize on the benefits of this technology quickly with a viable path to expand. Finally, we are going to bring in the ability to build agents natively with agent studio capabilities. So really exciting, we’re covering both the vertical and the horizontal as well as building tied together in a single enterprise-ready solution.
- #35 Assistants with Fixed Flows Fixed state-machine based conversation controller Fixed dialog flows for every task Rules or classic ML for routing between tasks Every new task requires a new flow Anything that falls out of fixed flow is an “off topic” exception Assistants with Fixed Flows + GenAI Fixed state-machine based conversation controller LLMs used to understand conversation context LLMs used as fallback for answering off topic questions Knowledge sources used to ground LLM responses LLMs can invoke robust digital automations to get things done Assistants with AI Agents LLM based meta-agent facilitates user interactions Flexibility to solve more complex tasks using agents, assistants, or skills/tools together Puts more power in the hands of the end-user to unlock new tasks Easily extensible to add new tasks (every new task is an agent tool!)
- #36 To look at this from a conceptual architecture point of view customers, employees and knowledge workers have a simple, intuitive conversational interface to guide them from problem or task to resolution. This feature which we call AI agent chat is a LLM-powered supervisory and routing layer that uses “agentic” capabilities to identify the domain, use case, task or information that is required and surface the best path for the requested activity or outcome. These AI solutions can be created easily with no code with in the AI assistant builder and are packed with AI and generative AI capabilities. The real boost to productivity comes from the skill catalog and studio that allows you to create and manage skills, configure them into multi-step flows and publish them to the skills catalog where you can choose from literally thousands of automations and tasks to power your assistants.
- #37 If we take a look at what's happening in the AI agent and assistant landscape we have entered a sort of continuum where multiple phases of these technologies exist side by side. This may include traditional fixed flow or rules-based chat experiences paired with traditional task automation, not very flexible, a pain to develop and deploy but still with a footprint today for many enterprises that are probably feeling pressure to modernize. Piggy backing on that are hybrid flows that augmented through the addition of generative AI, like RAG, and AI assisted automations that provide enhanced user experience and drive higher task completion. This is also a core component of current enterprise AI strategies, but we are rapidly entering a new era of autonomous flows led by AI agents and “agentic” AI orchestration that can plan, self-correct and complete tasks with little to no human intervention. This multi-faceted approach is a reality and enterprises really need to lean into the future to uncover the greatest benefits.
- #40 As you move from experimentation to deployment you are going to need the right building blocks. These are: Use cases — what do you want AI to help with? A strong data foundation and the right models, including smaller LLMs that are performant and more cost effective. End-to-end governance so you can deploy AI across the enterprise with confidence. And the assistants and agents that will take you to new heights of productivity. Agents, with their higher levels of autonomy, will learn and act on behalf of your teams.