Artificial Intelligence Blog Posts

Agentic ABAP with MCP: what is really solving problems for you?

Five field notes from a few months of agents on top of the ADT API in my S/4HANA 2023 lab, and five questions I would like to hear your answers to.

AI Product Management: Evals Are Not the Last Step

Most development frameworks put evaluation at the end. For deterministic software, that was late but recoverable. For agentic AI built with coding agents, it is the wrong sequence - and the cost shows up in production, not in the review.

How Companies Are Approaching MCP: Reflections From Amsterdam

This blog explores how companies are putting MCP(Model Context Protocol) into practice, drawing on insights from AGNTCon + MCPCon Europe. It covers protocol evolution, tool discovery, skills and agent execution, alongside our work on Open Resource Di...

AI Product Management: Not Every Decision Needs a Language Model

Three weeks ago a logistic regression beat eight LLMs at reading job postings. I gave it a rematch against a model built for tables. TabPFN-3.5 beat the lion, beat GPT-6 Astra and Claude Fable on the same postings, ran in 12 seconds per thousand, cos...

Why I stopped letting LLMs decide what to fetch

Letting the LLM decide what to fetch works. It's also non-deterministic, sequentially slow, and hard to audit.A short note on declarative pre-fetch as the default for enterprise agents and where tool calling is still the right answer

My productivity is boosted with Joule work desktop, this is the reason why

There is a long-standing debate about the use of IA and productivity in the corporate environment. This article explores the real-life experiences of an AI Consultant and how the mix of experience, collaboration, and reskilling generated outstanding ...

An Engineering Manifesto for AI-Assisted SAP Development — Let's Build It Together

AI helps us learn and write code. We have been shipping code faster than ever. Someone with no IT background can create software now. But software is not just code. It is structure, security, governance and much more. We deal with sensitive data. An ...

When Agents Multiply: Human Oversight in the Autonomous Enterprise

When we started this blog post series, we asked a hard question: how do we keep humans meaningfully involved as AI systems get more autonomous every day? Today, that question is more urgent than ever. SAP’s vision of the Autonomous Enterprise promise...

Governance at Scale: Why a Clean Core Is What Makes Agent Governance Enforceable

SAP has spent 2026 building governance infrastructure for AI agents — a verification workflow through AI Agent Hub, policy enforcement through MCP Gateway. Both promise trust badges and audit trails. Neither can deliver on that promise against an obj...

From AI Assistance to Human-on-the-Loop: A Compliance Detection Case Study

What if your compliance team could stop chasing false alarms and start focusing on what really matters? That’s exactly what happens when AI moves from being a helpful assistant to a tool that automatically handles routine cases.

Side-by-Side Without Relocating the Debt: Governance Patterns for CAPM (Series 3, Part 4)

Part 3 showed how SAP's Custom Code Migration Agent handles the mechanical half of clean core work. This part addresses what the agent explicitly hands back to a human: the objects that get side-by-side extended into CAPM. Done under deadline pressur...

Why Human Oversight Matters More Than Ever and What It Means

With the advent of AI agents, there is increasing interest in a topic that had been around for a long time in the AI Ethics world: Human oversight and determination. But why did it become so important now, and what does it mean?

From Predictions to Real Decisions: How SAP RPT-1.5 and Joule Agents Orchestrate AI Workflows

A deep-dive into building AI workflows in Joule that turn tabular structured data into key supply chain decision engines using SAP-RPT-1.5 Tabular AI model.

August Developer Challenge: A look at n8n with Vibe Coding

n8n workflows, as you saw in the Week 3 challenge, can be defined by JSON text file, and therefore you can easily vibe code to create complex and useful workflows. I will show you what I did do manage the challenge itself.

Email-driven SAP Service Quotation approval with n8n and OData

A personal experiment: an ABAP report sends a service quotation ID to n8n, n8n applies multi-stage email approval rules, and only the last Approve PATCHes the quotation to Released in SAP. Not a replacement for SAP Business Workflow or Flexible Workf...

Smarter Token Usage With Your AI Coding Agent

Most developers trying to cut AI costs focus on model selection. But the bigger lever is what your agent generates in the first place. Two open-source plugins — Caveman and Ponytail — tackle this from different angles: one cuts the verbosity of agent...

AI Product Management: The User Who Left the Room

The operator did not vanish when the agent arrived. She split. What she knew became the agent's context and bounds. What she checked before committing became the supervisory surface. The only casualty was the dashboard, and most agents are being buil...

Agent-to-Agent Protocol (A2A): From First Principles to Production

Most AI systems stop at a single agent talking to a single model. A2A is the open protocol that lets agents from different frameworks and teams discover each other, delegate work, and exchange results. This article shows you how to build it from scra...

August Developer Challenge – New Joule Studio

Everyone is abuzz about the new and improved Joule Studio, announced at Sapphire in May and expected out in the fall. We wanted to take this month’s developer challenge to prepare you for all the excitement that is likely to occur at SAP TechEd in Oc...

The Octopus Brain

How a distributed nervous system inspired the way I organize Claude's knowledge.An octopus has around 500 million neurons in a massive nervous system. Where humans concentrate most neurons in the head, an octopus does the opposite. About two-thirds l...

Humanoidd robots are a catalyst, not the solution

The hype is on on humanoid robots - price drops make them payable, increasing capabilities, better sensors, more computing power and stroner batteries make them affordable and applicable in a variety of use cases - but is the humanoid robot the answe...

The Agentic AI Team - The Architecture That Doesn't Exist Yet Is a Team

The Architecture That Doesn't Exist Yet Is a TeamPost 11, The Agentic AI Team seriesA claims agent denies a man's insulin reimbursement after a refrigeration outage. Nothing about the denial breaks a rule. The exclusion is ambiguous, the agent reads ...

The Agentic AI Team - The Failure No One Is Watching

The Failure No One Is WatchingPost 8, The Agentic AI Team seriesEighteen months after an agentic product ships is the moment it is most likely to be quietly broken and least likely to be caught. Not because anyone stopped watching. Because everyone k...

The Agentic AI Team - The Empty Column

The Empty ColumnPost 7, The Agentic AI Team seriesTake any team that ships an agentic product and ask them to draw their team structure on a whiteboard. They will draw it fast, because they have drawn it a hundred times before. A product manager deci...

The Agentic AI Team - Is the PM Still the Hero?

Is the PM Still the Hero?Post 6, The Agentic AI Team seriesSearch for "product manager" and "agentic AI" together and you will find no shortage of PMs telling other PMs that the agentic era finally makes the PM the center of everything. The engineer ...

The Agentic AI Team - Do We Still Need Scrum?

Do We Still Need Scrum?Post 4, The Agentic AI Team seriesA delivery team plans a two-week sprint the way it always has. Monday, the backlog is groomed. Tuesday, the sprint starts. Over the next two days, an agent drafts most of the committed features...

The Agentic AI Team - What Does a Team of 10 Developers Look Like Now?

What Does a Team of 10 Developers Look Like Now?Post 3, The Agentic AI Team seriesPicture a ten-person engineering team six months into using a coding agent well. Nobody writes much code from scratch anymore. What they do, all day, is read pull reque...

The Agentic AI Team - The Employee You Never Onboarded

The Employee You Never OnboardedPost 10, The Agentic AI Team seriesA company stands up an agent that triages inbound support tickets, drafts responses, and escalates the hard cases to a person. Procurement logged it as a software purchase. IT logged ...

The Agentic AI Team - The Green Checkmark Nobody Owned

The Green Checkmark Nobody OwnedPost 9, The Agentic AI Team seriesAn enterprise agent handles vendor onboarding. It checks a new supplier against policy, flags exceptions, and clears the routine cases without a human in the loop. Before launch it ran...

The Agentic AI Team - The Architect Was Right All Along

The Architect Was Right All AlongPost 5, The Agentic AI Team seriesThere was always one person on the project who was not really on the team. She did not sit in the standup, did not own a backlog, did not ship a feature. She showed up early, at the p...

The Agentic AI Team - Is There Still a UX?

Is There Still a UX?Post 2, The Agentic AI Team seriesA claims adjuster opens her laptop Monday morning. Overnight, an agent processed a hundred and forty spoilage claims: burst pipes, power outages, a warehouse freezer failure. It paid a hundred and...

The Agentic AI Team - The Team Didn't Shrink. It Reshaped.

The Team Didn't Shrink. It Reshaped.Post 1, The Agentic AI Team seriesA team ships an agent that handles procurement. It reads a purchase requisition, checks it against policy, compares suppliers, and places the order, with no human in the middle for...

The Agentic AI Team - The Sixth Archetype the Maps Forgot

The Sixth Archetype the Maps ForgotPost 0, The Agentic AI Team seriesThere was a time when photography was a planned profession. Film gave you thirty-six frames, each one cost money, and none of them existed until a lab developed the roll days later....

From Strategy Doc to Runtime Rule: Embedding SAP AI Standards into Claude Code

What if your AI coding assistant already knew every relevant product standard before you typed a single line?

Multi-Agent LLM Pipeline with SAP RPT and other SAP AI Core LLMs for World Cup 2026 Match Prediction

What happens when you combine SAP's Relational Pretrained Transformer with a multi-agent LLM system and ask it to predict World Cup 2026 match scores? Surprisingly decent football analysis — and one very honest timeout.

Claude Generates So Much — Here's How I Actually Keep Up

Claude Generates So Much - Here's How I Actually Keep UpAI is producing an almost comical amount of output these days. In a single session, Claude can scaffold a feature, write the tests, refactor two files, explain the architecture, and suggest thre...

AI Product Management: The Human Side of Agentic AI Nobody Designs

Post 8 — AI Product Management seriesIn December 1978, United Airlines Flight 173 ran out of fuel while circling Portland, Oregon. The aircraft was mechanically sound. The crew knew the fuel state was critical. The captain, fixated on a landing gear ...

AI Product Management: What Happens to Agents Nobody Plans to Stop

Post 9 — AI Product Management seriesIn 2021, a research team at the University of Michigan published an audit of the most widely deployed sepsis prediction algorithm in American hospitals. Epic Systems had marketed the Epic Sepsis Model with a claim...

AI Product Management: The Two Documents That Replace the PRD

The PRD worked when the hard part was building. With agents, building is the easy part, a prototype takes an afternoon. The hard part is deciding what the agent should do, where its authority stops, and how anyone knows when it's wrong. That work spl...

AI Product Management: Autonomy Is Earned, Not Scheduled

"After 30 days, we'll move the agent to full autonomy." That's a schedule, not a safety criterion. A clean record usually means the hard cases haven't come up yet, not that the agent can handle them. Why autonomy has to be earned against evidence, an...

Is LLM fine-tuning really dead?

I’ve been hearing a lot lately about fine-tuning being “dead,” but the reality is much more interesting. In this SAP Community post, I share what fine-tuning actually is, where it still shines, and why it continues to matter for real-world enterprise...

AI Product Management: What Keeps an Agent From Becoming a Liability

Drift is not one thing. An agent can drift in four independent directions simultaneously: model, data, supervision, and scope. None of the four announces itself. The combination can be more dangerous than any single vector. This post covers the four ...

AI Product Management: What You Need to Watch After You Ship

An agent that is not monitored does not sit and wait. It continues to act, autonomously, at scale, until a supplier calls to ask about orders that never arrived. This post covers the six observation instruments every agentic product needs, why automa...

AI Product Management: How Do You Know the Agent Is Actually Working?

The QA mental model that has served you for twenty years breaks in exactly three places when software starts making decisions instead of executing code. Pass@K replaces binary pass/fail. Compound probability means component accuracy and system reliab...

AI Product Management: The Four Things Every Agent Needs Before It Ships

Most agentic products ship with only the agent designed. The supervisory system, how humans monitor, intervene, audit, and recover from what the agent does, is an afterthought. This post covers the four runtime artifacts every agentic product require...

AI Product Management: The Cost Model Your Business Case Is Missing

Business cases for agentic AI are systematically 40 to 60 percent too low. Not because people are careless. Because the cost structure is genuinely different, token pricing is not what you expect, and the governance stack nobody budgeted is the line ...

AI Product Management: Not Every Problem Deserves an Agent

The most expensive mistake in agentic AI does not happen in production. It happens before a single line of engineering is written, when a team decides to build an agent for a problem that should have been solved with a clearer dashboard or a better r...

Claude Code: Best Practices for Developers

A practical guide to working effectively with Claude Code. Covers workflow principles, planning strategies, model selection, and the full tooling ecosystem.Who This Is ForDevelopers — from first-time Claude Code users to power users. Part 1 covers pr...

AI Product Management: Seven Things You Need to Understand About AI Before You Build With It

Seven mental model shifts every product manager needs before any agentic AI project begins. Not because they are advanced. Because they are prerequisites. From suggestion vs. action to the two channels most teams never build — the concepts that separ...

AI Product Management: A Field That Doesn't Have Its Frameworks Yet

Agentic AI is not just smarter software. It is probabilistic, it acts autonomously, and it does not know when to stop and ask for help. This series maps the frameworks product managers need to design the system around the agent, not just the agent it...

Explore this Area