At the end of February 2026, SAP hosted its Agent Days event at the SAP office in Berlin – a two-day gathering that brought together SAP product managers, transformation & strategy consultants, and customers from across the Middle and Eastern Europe region. The stated purpose was clear: give attendees a grounded understanding of where SAP stands on agentic AI, what the technology actually consists of, and how companies can begin or accelerate their agent transformation journeys.

This blog post is a condensed summary drawn directly from the session transcripts across both event days. It covers the strategic input, the technical architecture, the development approaches, the customer perspectives, and the practical lessons that emerged so far. For full disclosure: Claude 4.6 Opus Reasoning helped to summarize – however afterwards we had humans in the loop to verify the content and language. 

The event opened with hosts Philipp Sutterlin and Sandra Jörns welcoming the international audience. A Mentimeter poll showed that attendees were mostly interested in business cases, agents, strategy, inspiration, exchange, and networking — demonstrating that they had moved beyond basic awareness to active execution.

One recurring theme from the first session: agents are not a new category of software tools. They represent a shift in how software executes work — from human-driven, step-by-step interaction to goal-directed, autonomous action within guardrails. Whether that shift is genuinely transformative or another wave of overpromised enterprise IT will depend almost entirely on implementation discipline, governance, and change management — not on the technology alone.

Anja Schneider, Head of Customer Services & Delivery in Middle and Eastern Europe kicked the event off with her insights. Two of her quotes gave the direction for the event:

“SAP is deep into running end-to-end processes in a stable, reliable, secure environment since many years.”

“Let’s take a look at how we can do agentic AI also in a safe, stable and reliable way where we as humans are still in control.” Use the momentum! And let me and us from SAP understand what is top of mind for you.”

—  Anja Schneider, Head of Customer Services & Delivery in Middle and Eastern Europe, SAP

This was not only a welcoming check-in. It is also the invitation to exchange thoughts and experiences.

SAP's Business AI Strategy and the Agent Portfolio

From Copilot to Autonomous Agent: Positioning the Shift

Richard Grandpierre (Head of Product Management SAP Business AI) framed SAP's overall AI positioning. The argument SAP is making is that the return on investment from AI will not come primarily from generative AI assistants or embedded AI features within individual applications, but from agents that can execute multi-step tasks autonomously across systems and processes.

SAP is not alone in making this opinion — this aligns with the broader industry narrative. But what SAP is specifically claiming is that its unique position comes from the combination of deep process knowledge embedded in its applications, business data that spans ERP, HR, Procurement, and Supply Chain, and an integration layer (BTP) that connects all of these. The argument is that an agent built on top of SAP's stack does not need to re-learn what a purchase order is or how a goods receipt process works — that knowledge is already encoded in the system.

Richard delivered his keynote with live demos showcasing SAP's full Business AI portfolio — spanning embedded AI features, the Joule Copilot, and the emerging wave of agentic AI across Finance, Supply Chain, Spend, and HR.

A key message was the strategic shift from reactive, transactional AI interactions toward a “fleet of agents” capable of autonomously executing end-to-end business workflows. Using the Accounts Receivable agent as an example, he illustrated how agents can automate significant portions of a workflow, doing the actual work rather than just guiding the user through it.

Richard made clear that this transformation will be gradual but is already underway, and that SAP's approach remains open and ecosystem-friendly to ensure seamless integration with partner and customer landscapes.

Setting the Scene – The 10-20-70 Rule as a Starting Point

The event used also an interactive quiz to engage and check different assumptions. One opening question asked participants to define the 10-20-70 rule in the context of AI and digital transformation. According to Patrick Kohler who heads AI Strategy & Advisory at SAP: 10% of AI transformation success comes from Technology, 20% from Data, and 70% from People & Processes.

The audience — predominantly SAP customers and IT decision-makers — scored well on this question. That is notable. It suggests that the audiences attending these events are increasingly moving past the ‘what is AI' phase and are dealing with the harder organizational questions. The 10-20-70 rule is not a new concept in enterprise IT. But it bears repeating as organizations encounter agentic AI, because the temptation to treat agents as a purely technical deployment problem remains strong.

Patrick introduced in his keynote a strategic AI Playbook built around six components, including a multi-year, IT-business-aligned strategy, a dedicated AI program with its own funding and governance structure, and a clear KPI framework to measure impact.

For the SAP context specifically, he recommended a three-step approach: identify your own starting point, set priorities, then scale. He stressed that there is no one-size-fits-all solution — every organization starts from a different position. The overarching message: the technology is ready and available; the real work lies in organizational enablement.

The Three Layers of SAP's Agentic AI Stack

Across multiple sessions, SAP's positioning can be distilled into three layers:

• Joule as the Conversational Front-End: Joule is SAP's AI assistant, accessible across SAP applications. It acts as the primary interface through which users interact with agents — asking questions, triggering workflows, and receiving outputs in natural language or dynamically generated UI components.
• Pre-Built SAP Agents: SAP is delivering a growing library of agents that are embedded within its applications — for finance, HR, procurement, supply chain, and other functional areas. These are production-ready agents designed for SAP-standard processes.
• BTP as the Agent Runtime and Development Platform: The SAP Business Technology Platform is where all agents — both SAP-built and custom-built — run. It provides the infrastructure for tool connectivity, memory, orchestration, and governance. The AI Agent Hub in Joule (not to be confused with the Generative AI Hub on BTP, which is different) allows organizations to manage and deploy agents across their landscape. Connected there is SAP Build Code and Joule Studio for AI Agent creation for customers and partners who want to design, test & create their own agents.

How SAP Agents Actually Work — Technical Architecture

Inside the Agentic Loop

Mathis Boerner, Central Product Manager for AI Agents at SAP, walked through the internal mechanics of an agent in a way that was accessible without being oversimplified.

The core architecture of an SAP agent follows a pattern that is consistent across most agent frameworks — including Claude Code and other leading coding agents:

• Query Decomposition and Planning: The agent receives an intent and breaks it down into actionable tasks. It forms an initial plan for how to proceed.
• The Agentic Loop: This is the recursive core. At each step, the agent evaluates: Do I need to continue? Am I done? Can I solve this? If the answer is “continue”, it selects an action.
• Action Execution via Tools: Tools are how the agent perceives its environment and takes actions. They connect to backend systems, search the web, retrieve data, trigger processes, or call APIs. Different types of tools exist for different purposes.
• Humans in the Loop: Mathis was explicit that this is not optional. One of the available actions for any agent is to ask the user — to request guidance, approval, or clarification. ‘We want collaborative agents and humans need to be in the driver's seat.'
• Memory and State: Agents maintain context across steps. SAP's framework includes mechanisms for both short-term (within-session) and longer-term agent memory, which Mathis described as important for more complex, multi-session workflows.

“One very important tool is also humans in the loop. We want collaborative agents and humans to be in the driver's seat. So, one action is often always to say: ask the user, ask for guidance, ask for approval.”
— Mathis Boerner, Central Product Manager AI Agents, SAP

A question from the audience during Mathis Boerner's session on AI Agents at SAP raised a practical point: What about organizations that are not wall-to-wall SAP? Mathis answer was direct: All agents run on BTP, and BTP is designed for interoperability. Custom agents can connect to third-party systems via tools, and the platform supports bringing together SAP and non-SAP environments. That said, the depth of out-of-the-box connectivity is naturally deeper for SAP systems.

Pro-Code & Low-Code Agent Development

Jochen Schneider who heads BTP AI at SAP presented SAP's ambition to democratize agent development across all roles in the enterprise — from business users through to experienced developers. He walked through the different development modes available in SAP Build: business users can configure agents using natural language, low-code developers work through a visual interface, while pro-code developers can deploy to BTP from their preferred IDE using any framework of their choice.

Central to his talk was the concept of Intent-Based Development (IBD): Rather than writing code, users describe their business intent, and the system derives the appropriate solution. He also introduced the MCP Server capability within SAP's portfolio, enabling interoperability with external tools and third-party agents.

The core message: Choosing the right development layer is key — not every agent requires a developer to build it.

Joule Studio and the AI Agent Hub

SAP's Joule Studio is the environment in which developers and technically capable business users can define the tools that agents use to interact with backend systems. It sits within BTP and is the bridge between the agent's reasoning layer and the actual SAP or third-party systems it needs to act upon.

The AI Agent Hub is the management and deployment interface for agents across the SAP landscape. It is separate from the Generative AI Hub on BTP, which focuses on LLM access and model management. This distinction caused some confusion in the audience Q&A, and Mathis acknowledged that the naming overlap is not ideal.

An important architectural note: Custom-built agents are developed on BTP, and all agents — whether SAP-built or custom — run on BTP. This creates a unified runtime even in hybrid landscapes, though the depth of integration varies depending on which systems the agent tools connect to.

Joule: From Assistant to Application Generator

One of the live demonstrations during the event showed Joule in a mode that goes significantly beyond conversational Q&A. In the demo, a user asked Joule to analyze sales order issues. Instead of returning a table or a text response in the chat interface, Joule generated a fully interactive card — a dynamic application within the SAP system — displaying the sales order issue data in a browsable format.

The user then asked Joule to convert the card into a bar chart showing issues per month. The system regenerated the visualization on the fly. A follow-up prompt — create a high-level summary, what should I be concerned about? — returned a structured analytical summary identifying fulfilment issues and operational efficiency patterns.

This is a meaningful shift in what an AI assistant does. Rather than being a search or query interface layered on top of an application, Joule in this mode is generating the application layer itself — dynamically, in response to the user's intent. The implications for how SAP UX evolves are significant, though it is still early in production rollout.

Driving Agent Transformation — Strategy and Change Management

The Three-Step Kickstart Approach for Agent Transformation

Phillip Sütterlin presented a structured approach on Agent Transformation – developed and tested with early customers. He described this as a three-step model for kickstarting agent transformation:

• Step 1 – Inspiration: Give the organization a clear understanding of what agents are, what SAP's agent portfolio looks like, and what is technically possible. This is the awareness and framing layer also called Agent Days.
• Step 2 – Use Case Discovery: Working in groups organized by SAP system domain (e.g., SAP S/4HANA, SuccessFactors, Ariba), teams explore, identify specific agentic use cases relevant to their organization. A typical output is 40–50 use cases.
• Step 3 – Use Case Deep Dive / Hackathon: The validated and prioritized high-value use cases go into a hands-on implementation sprint. Pre-work involves ensuring prerequisites are met like identifying the right APIs to enable the agents to execute respective target tasks. The output is a working prototype and a concrete implementation path.

The validation between Steps 2 and 3 is critical. Having 40 use cases is not the same as having 40 good use cases. Validation criteria should include business value, technical feasibility (API availability, data quality), organizational readiness, and the degree to which human oversight is preserved in the workflow. For starting with agents it is also beneficial to start in areas where the business criticality and risk is not too high.

“From intent to a defined business problem — we really help you nail down the problem you want to solve, and track whether you finally improve on the KPI you aimed for.”
— Philipp Sütterlin – SAP Business AI Transformation Consultant

The Build-Run-Govern Framework

Across multiple sessions, SAP's recommended approach to agent deployment was structured around three phases: Build, Run, and Govern.

• Build involves defining the business problem precisely, selecting the right combination of app, agent, and process, and using SAP's development tooling on BTP. The recommendation is to use as much code generation as possible to reduce development effort.
• Run involves operationalizing the agent — but not just technically. This means establishing observability (Can you see what the agent is doing?), monitoring (Are outcomes within expected parameters?), and agent memory management (Does the agent retain useful context across sessions?).
• Govern is the layer that organizations tend to underinvest in. It includes defining company-specific best practices, establishing guardrails, and having a clear escalation path for when agents produce unexpected outputs. One thing is clear: With multiple agents running across multiple systems simultaneously, mistakes will happen (like currently with Humans). The question is whether the organization has the governance infrastructure to catch up and correct them.

“There will be mistakes — not just with one agent, but with all the agents running across all of your systems. So really understanding what the company-specific best practices are really helps.”

Early AI Adopters and the Rollout Strategy

One of the practical sections of the event addressed how to approach the internal rollout of agent initiatives — not just the technology deployment, but the organizational diffusion.

Suggestions are to identify early adopters, grant them sandbox autonomy to experiment, and leverage peer-to-peer influence and learning. When an individual or team builds something that works, make the result visible. Broadcasting lighthouse wins — concrete examples with measurable outcomes (e.g., ‘50% efficiency increase in this specific workflow') — drives adoption in the broader organization faster than top-down mandates.

The concept of formalizing shadow IT was also raised. In many organizations, motivated employees are already experimenting with AI tools outside official channels. Rather than suppressing this, a structured agent transformation program can channel that energy: Give teams a legitimate sandbox, capture what they build, and use successful experiments as the foundation for broader deployment. This approach also de-risks the rollout — by the time agents go into production across the organization, several edge cases have already been discovered and handled.

Customer Panel: Customer Perspectives and Real-World Experiences

Anja Schneider moderated the customer panel closing the first day. The session focused on the “how” of operationalizing AI in real enterprise environments, moving beyond SAP's own portfolio to hear from customers living it in real world practice. 

Sheila Cushman, CoC Lead for SAP Architecture & Technology at BMW Group, shared that BMW's CIO has set a clear strategy: “AI for everyone, everywhere” — giving every employee access to AI through BMW's own platform and Microsoft Copilot. On the agentic side, BMW deliberately focuses on standardized, out-of-the-box SAP agent products rather than custom builds, expecting the combination of agentic AI and natural language access to SAP systems to revolutionize transactional business processes.

A key practical insight from Sheila was the concept of the “circle of control” — the idea that the appropriate level of agent autonomy should scale with the scope of impact. Agents operating within a single user's personal workspace can be given broad autonomy with minimal oversight, but as the circle expands to team workflows, cross-system processes, or financial transactions, governance requirements must increase proportionally. Agents that can initiate purchases, approve exceptions, or alter master data need tighter controls and explicit human checkpoints.

“If I go for a bigger circle, then I probably need IT involved, and I need to have the governance in place.”

This tiering approach is practically valuable: Not every use case requires the same governance overhead, and starting small reduces risk while accelerating time to value. On use case identification more broadly, Sheila emphasized that cross-functional collaboration is essential — bringing together digitalization groups, solution architects, and business process experts, as no single group holds all the knowledge needed.

Alex Gonzalez, Global Director IT & Digital Innovation at Nestlé outlined two strategic pillars for AI:

• “Everyday AI” — augmenting employees to be more effective through AI assistants with humans in the loop, and
• Process Reimagination — rethinking end-to-end processes like Hire-to-Retire and Order-to-Cash.

Alex noted that Nestlé is not yet at the stage of fully reimagining processes around agents. He emphasized that most processes involve many stakeholders, systems, and data touchpoints — which makes the journey complex and thoughtful.

Falko Lameter, CIO at Kaeser Kompressoren, contributed to discussions on the practical realities of agent implementation and use case validation — noting the challenge many companies face: ending up with large lists of ideas (e.g., 500 use cases in a spreadsheet) without a clear internal framework to validate which ones are truly valuable. Kaeser runs a deliberate dual-platform strategy (“Google and SAP Business AI is our strategy”) combining the strengths of both ecosystems rather than betting on a single vendor. His team has already moved from ideas to impact: a Visual Spare Parts Search and an AI-powered nameplate recognition tool are changing how field service engineers work day-to-day, both initiated bottom-up by motivated employees rather than top-down mandates.

Several consistent themes emerged across all three panelists and the broader customer discussions at the event:

• Human in the loop by design — Knowing where to place human oversight is a fundamental design decision, not an afterthought. Governance is not a blocker; it is a design requirement. The fastest-moving organizations had defined minimum viable governance — clear ownership, escalation paths, and monitoring — from day one.
• Use case discovery and prioritization — Generating ideas is easy; identifying which are technically feasible, business-valuable, and organizationally ready is harder. Customers who had made the most progress invested significant time in structured validation before starting development. No single team can do this alone — business, IT, and process experts must collaborate.
• Data and API readiness as hidden dependencies — Bad data leads to bad agent decisions. Equally, agents act through tools that require APIs, and in many SAP landscapes those APIs either don't exist in the required form or are poorly documented. Both must be addressed as pre-work, not discovered mid-implementation.
• Change management and trust — User adoption, not technical failure, was a primary risk factor cited by customers who had deployed agents. Transparency and explainability are non-negotiable for building employee trust. Agents that work correctly but aren't used deliver no business value.
• Workforce evolution — Agents will progressively shift human roles from executing tasks to orchestrating and moderating fleets of agents — making workforce reskilling and organizational redesign an essential part of any agentic AI program.
• Governance: Don't wait for a perfect governance framework before starting. Grant a small group of motivated employees sandbox autonomy, let them build, and use their results — including failures — to learn and to inform your broader rollout strategy.

Development Tracks and Hands-On Guidance from Day 2

Pro-Code and Low-Code Agent Development on BTP

Day 2 sessions split into more technical and more strategic tracks. On the technical side, the focus was on how to actually build agents on SAP BTP — both for developers (pro-code) and for business users or citizen developers (low-code/no-code).

SAP's position is that agent development should be accessible at multiple skill levels. Pro-code development on BTP offers full flexibility — developers can define custom tools, implement complex orchestration logic, and connect to any system with an API. The low-code path, likely through SAP Build / Joule Studio, enables less technical users to configure agents using pre-built components and connectors, without writing code.

The practical guidance from the sessions was to match the development approach to the use case complexity. Simple, well-defined use cases with standard SAP connectivity are good candidates for low-code configuration. Complex, cross-system use cases with custom business logic require the pro-code path. Mixing the two — starting low-code and extending with custom tools where needed — is also supported by the platform architecture.

From Intent to Business Problem: A Practical Design Framework

One of the sessions presented a framework for moving from a vague intent (“We want to use agents for procurement”) to a well-defined business problem (“we want to reduce the time to resolve purchase order exceptions by routing them automatically to the right approver with full context”).

The framework emphasized that a good agent use case has three characteristics:

• a clearly defined trigger (What initiates the agent?),
• a bounded scope of action (What can the agent do, and what is outside its authority?), and
• a measurable outcome (How do we know if it worked?).

Without all three, an agent deployment cannot be evaluated and cannot be improved.

The session also addressed the combination of app, agent, and process. An agent does not replace an application or a process — it operates within and across them. Getting the right combination of static application logic, dynamic agent reasoning, and defined process flow is the architectural challenge. SAP's tooling on BTP is designed to support all three layers together.

Evaluation, Observability and Agent Memory

Several sessions touched on what happens after an agent is deployed — a phase that receives less attention in most AI discussions than it deserves. Key points were:

• Evaluation frameworks (evals) are necessary to systematically assess whether an agent is performing as intended. Unlike traditional software, where correctness is binary, agents operate probabilistically — an eval framework defines what “good enough” looks like and tracks performance over time.
• Observability during runtime means being able to see what an agent is doing, step by step, as it executes. This is both a debugging tool and a trust-building mechanism — for IT teams and for end users.
• Agent memory management is important for use cases that span multiple sessions or require the agent to learn from past interactions. SAP's platform includes mechanisms for this, but the design of what to remember — and what to discard — is a configuration decision that requires also human thought.

Practical Tips for Organizations Starting Their Agent Journey

Tip 1: Start with the business problem, not the technology

Define the trigger, the bounded scope of action, and the measurable outcome before selecting any tool or platform. Agents built around a vague intent tend to deliver vague results. Their use needs to solve business problems.

Tip 2: Invest in use case validation before development

A structured validation step — assessing business value, technical feasibility (API availability), and organizational readiness and demand — avoids wasting hackathon time on use cases that will not make it to production.

Tip 3: Audit your API landscape early

Agents act through tools, and tools require APIs. Often organizations discover that the APIs they need either do not exist or are insufficiently documented. This is the single most common hidden constraint on agent development timelines.

Tip 4: Leverage early adopters – grant sandbox autonomy to motivated employees

Early adopters within the organization will find the edge cases that your governance framework has not anticipated. It is better to find them in a controlled sandbox than in a production deployment.

Tip 5: Define minimum viable governance before go-live

You do not need a complete AI governance framework before deploying your first agent. But you do need clear ownership, an escalation path for unexpected behaviour, and basic monitoring. These are non-negotiable.

Tip 6: Make results visible inside the organization – celebrate quick wins

Concrete, named examples with measurable outcomes (“Team X reduced exception handling time by Y%”) are the most effective driver of internal adoption. Abstract promises about AI potential are not.

Tip 7: Calibrate autonomy to the scope of impact

Use the “circle of control” mental model. Agents operating within a single user's workspace can have broader autonomy. Agents that affect financial transactions or cross-system processes need tighter controls and more human checkpoints.

Tip 8: Do not treat Joule, the AI Agent Hub, and the Generative AI Hub as the same thing

They are distinct products that serve different functions. Get clarity from SAP on the specific product components relevant to your use case and your landscape and map these to your existing BTP entitlements. Next to help.sap.com there is lots of free digital learning on learning.sap.com to make yourself knowledgeable.

Tip 9: Plan for the run phase from the start

Observability, monitoring, agent memory management, and evaluation frameworks are not post-deployment concerns. They need to be designed into the solution architecture from the beginning.

Tip 10: Ask SAP specifically about requirements and compatibility

Before committing to a use case that depends on specific agent capabilities, confirm which SAP release versions are required and what the compatibility matrix looks like for your landscape.

Summary and Outlook

The SAP Agent Days MEE in Berlin delivered a picture of where SAP and customers stand on agentic AI. The agentic loop architecture is consistent with how serious agent frameworks work in practice. Joule Studio and AI Agent Hub are functional products, not slide-deck concepts. Joule's ability to dynamically generate application components — not just conversational responses — represents a genuine shift in what an AI assistant does within an enterprise system.

The harder questions — around governance, API readiness, change management, and production-grade reliability — were addressed more candidly than is typical for such events. Multiple speakers acknowledged that mistakes will happen and that the 70% people-and-process dimension cannot be treated as a post-implementation concern.

For SAP customers considering where to invest attention in 2026, the Agent Days validated a few priorities: Start with structured use case discovery, do the API groundwork before, invest in governance design from day one, and use early adopters and lighthouse results to drive internal adoption. None of these are new principles — they apply to any significant enterprise software initiative. But they are especially important with agentic AI, where the potential for both impact and unintended consequences is higher than with conventional automation. 

The overall mood was overwhelmingly positive and energized. Words like inspiring, informative, insights, and eye opening dominated — a strong signal that the event delivered on its promise to equip SAP partners and customers for the agentic AI era.

How to Start Your Agent Transformation?

The conversations at SAP Agent Days MEE made one thing clear: The organizations making the most progress are not waiting for a perfect strategy — they are starting small, validating fast, and building momentum through real use cases and real results.

If you are ready to take the next step, we are here to support you — whether that means identifying where to begin, building internal capability, or accelerating what you have already started.

Here is how customers can work together with SAP:

• Use Case Discovery Workshops: A structured, cross-functional session to surface, validate, and prioritize your most promising agentic AI opportunities. Modelled on SAP's   We help you move from 500 ideas to 5 that are ready to be used by you.
• Upskilling & AI Literacy Programs: Explore SAP's AI learning resources as a starting point and let us design a program tailored to your organization to develop hands-on AI skills.
• Agentic AI Hackathons: A focused, time-boxed sprint to prototype a live agent use case — with real data, real APIs, and a real go/no-go decision at the end. The fastest path from conversation to conviction.
• AI Transformation Roadmapping: If you are looking for strategic clarity, we can help you map your current state, define your governance model, and sequence your agent portfolio for maximum impact.   
• Visit our AI Experience Workshops in Warschau (22.4.), in Ratingen for MHT / Automotive (23.4), in Linz (May) or the AI Executive Roundtable in the SAP Garden at 1st of April.

Please contact us if you want to engage in a discovery workshop or any strategic session – the colleagues from the AI Strategy and Advisory Team are happy to help.  

Frequently Asked Questions (FAQ) an Agentic AI & SAP

The following questions reflect common themes raised by participants during the event's Q&A sessions and breakouts.

Q1: Do agents only work if I have a full SAP landscape?

A: No. All agents — whether SAP-built or custom-built — run on BTP, which supports interoperability with third-party systems. Custom agents can be built to connect to non-SAP systems via tools and APIs. However, the depth of out-of-the-box integration is greatest for SAP systems.

Q2: What is the difference between the AI Agent Hub and the Generative AI Hub?

A: These are two different products. The AI Agent Hub is accessible within Joule and is the interface for managing and deploying agents across the SAP landscape. The Generative AI Hub is a component of SAP BTP focused on LLM access and model management. They are complementary but serve different functions.

Q3: What is the agentic loop and why does it matter?

A: The agentic loop is the recursive core of how an agent operates: It evaluates its current state, decides whether to continue, selects an action (calling a tool, asking a human, retrieving data), executes it, and repeats until the task is complete or it determines it cannot proceed. Understanding this loop helps in designing appropriate guardrails and human-in-the-loop checkpoints.

Q4: How do we identify good agent use cases?

A: Good use cases have three characteristics: a clearly defined trigger (what starts the agent?), a bounded scope of action (what is the agent authorized to do?), and a measurable outcome (how do we know it worked?). SAP recommends a structured use case discovery process — working by system domain in groups, generating 40-50 candidates, and then validating on business value, technical feasibility, and organizational readiness before selecting which to develop.

Q5: What is the right governance approach for agent deployments?

A: SAP's guidance is to define minimum viable governance before go-live — not to wait for a complete framework. This includes clear ownership of each agent, an escalation path for unexpected behaviour, basic monitoring and observability, and an evaluation framework for tracking performance over time. As agent scope and autonomy increase, governance requirements increase proportionally.

Q6: How do we handle the 70% people-and-process dimension?

A: Start with early adopters who are motivated to experiment. Grant them sandbox autonomy. Capture what they build, including the failures. Use concrete lighthouse results — named examples with measurable outcomes — to drive adoption in the broader organization. Treat change management as a design requirement from the start, not a post-deployment activity.

Q7: What technical prerequisites are needed before starting agent development?

A: The most underestimated prerequisite is API readiness. Agents act through tools, and tools require APIs. Before a hackathon or development sprint, audit which APIs exist in your landscape, which are stable, and which need to be created or extended. Also confirm which SAP release versions are required for the specific agent capabilities you plan to use.

Q8: Can business users (non-developers) build agents?

A: Yes, with limitations. SAP supports both pro-code development on BTP (for developers) and low-code configuration through SAP Build Joule Studio (for business users). Simple use cases with standard SAP connectivity are good candidates for low-code approaches. Complex, cross-system use cases with custom business logic require developer involvement. A mixed approach — starting low-code and extending with custom tools — is architecturally supported. However, with low-code agents you need technical know-how.

Q9: How does human-in-the-loop work in practice?

A: Human-in-the-loop is implemented as one of the available actions in the agent's tool set. At any decision point, the agent can be designed to pause and request user input, approval, or clarification before proceeding. This is configurable and should be a deliberate design decision based on the risk profile of the use case. Higher-stakes actions (e.g., initiating a financial transaction) should always include a human approval step.

Q10: What is the realistic timeline from use case discovery to production deployment?

A: This was not given a specific timeline in the sessions, and for good reason — it depends heavily on API readiness, organizational complexity, and governance maturity. What was described is a three-phase approach: inspiration and discovery (event-based, a few days), validation and deep dive/hackathon (weeks), and production deployment (months). Organizations that skip or rush the discovery and validation phases consistently encounter.

 

Were you at SAP Agent Days Berlin? We'd love to hear your takeaways in the comments. And if you're planning your own agent transformation journey, feel free to reach out — or join the conversation here in the SAP Community.