Novaeon.Studio

When your AI agent writes to a core system, whose permissions does it use? OWASP: "Excessive Agency" (LLM06:2025). Our rule: the agent acts as the user, and the target system decides what is allowed. genai.owasp.org/llmrisk/llm0… #AIAgents #AISecurity

We published an open dataset: 3,413 synthetic agent change requests, labelled with risk class and release gates. Plus a LoRA adapter on Qwen3-4B. Risk accuracy: 54.5% → 63.3%. A decision aid, not an approver. huggingface.co/datasets/Nova… #AIAgents #OpenSource

Under DORA, every ICT change at an EU financial entity must be recorded, tested, approved and verified. An AI agent is a change too. Our banking patterns: pre-checks, four-eyes approval, audit evidence. A human approves. novaeon.studio/industries/ba… #DORA #AIAgents

Custom Joule agents run free in a productive Joule Studio runtime through the end of 2026, per SAP's pricing page. In production, usage is billed on consumption. Inventory, usage baseline, owner, budget: settle them now. novaeon.studio #SAP #Joule

Wrong or inconsistent AI answers? Match the fix to the failure. Missing knowledge: retrieval (RAG). Inconsistent behaviour: examples first, fine-tuning later. Which is yours? novaeon.studio #GenAI #RAG

AI Act plan still says "high-risk rules from August 2026"? Update it. The AI Omnibus (in force 27 July 2026) moved Annex III high-risk rules to 2 Dec 2027 and regulated-product AI to 2 Aug 2028. Not legal advice. novaeon.studio #EUAIAct #AIGovernance

Best on paper isn't the best agent. Our gate: 43 deterministic tests, no LLM judge. A dense 27B scored 38/43 vs 39/43, but tried to announce a server reboot it couldn't do. Since the gate: 8 builds tested, 6 stopped. novaeon.studio #AIAgents #AIEvaluation

Not every AI use case needs an agent. Many need a workflow and a human at the right step. Our rule: AI suggests, a person approves, AI acts only where errors are cheap and reversible. Where do you draw the line? novaeon.studio #AIAgents #HumanInTheLoop

Cloud ALM has a Joule agent for requirements. For change and release management after go-live we found nothing concrete in SAP's material. SolMan 7.2 mainstream maintenance ends 2027. Is your change process ready? novaeon.studio #SAP #SAPCloudALM

Your next AI project will probably ask for "MCP access". MCP is becoming the common plug between AI agents and tools. Every server you connect is code your agent will trust. 5 things to secure before go-live: novaeon.studio #MCP #AISecurity

An AI agent that reads your inbox also reads every email a stranger sends you. Prompt injection is #1 in the OWASP Top 10 for LLM apps. So don't only ask "can we block it?" Ask "what could a hijacked agent do?" novaeon.studio #AISecurity #AIAgents

Should one big model pick the model for the job? We tested it on 60 tasks. Alone, our main local LLM kept 9 of 20 hard tasks for itself. Add a small 9B router as second judge (local only if both agree): right tier in 49 of 52. novaeon.studio #LLM #AIAgents

Per the OECD's 2025 SME survey, only 29% of SMEs using generative AI use it in their core activities. Start where it pays: work that is repetitive, rule-heavy, high-volume and error-prone. Measure the baseline first. novaeon.studio/#diagnostic #AIAutomation #SME

SAP says customers control how much autonomy Joule gets. That's a governance decision, not a setting. Treat AI agents like transports: inventory, dev/test/prod, regression tests, autonomy in steps, rollback + audit trail. linkedin.com/feed/update/urn… #AIGovernance #SAP

Want an agent that decides better? Don't fine-tune your big model. Train a small decider next to it. Unsloth took Qwen3.5 0.8B from 20.7% to 74.3% average accuracy on 3 decision benchmarks, on 4 GB of VRAM. Why we keep it separate: linkedin.com/feed/update/urn… #AIAgents #LLM

Learning SAP Joule without spending 1,000 hours? Here's my learning path. Including the module almost every list leaves out. Module 1 – Foundations Skills, agents, assistants: what is what. And where Joule runs: public cloud first, private edition catching up, on-premise only partially. Module 2 – Setup BTP, Cloud Identity Services, formations, Work Zone. This is where you find out whether Joule starts in your landscape at all. Module 3 – Use cases What is generally available today and what is still roadmap. Separate the two before you promise anything to the board. Module 4 – Building Joule Studio (classic and new), custom agents, MCP and A2A. Module 5 – Cost Base AI is included. Agents consume AI Units per action. Free runtime for custom agents ends on 31 Dec 2026. Module 6 – Taking agents to production Missing almost everywhere. An agent acts inside finance, HR and procurement. So it needs what every change in SAP needs: – an inventory entry and an owner – a risk rating and a defined autonomy level – tests and evaluation before go-live – transport and approval instead of "just switch it on" – monitoring, a cost budget and a rollback plan – works council involvement where required, and the EU AI Act in view After 22+ years in SAP release and change management, I'm convinced: module 6 decides whether an agent stays a demo or makes it into operations. Links to all modules (official SAP sources) are in the first comment. Want my module 6 checklist? Comment "checklist" and I'll send it to you. linkedin.com/posts/maik-buch… #SAP #Joule #SAPBTP #CloudALM #AIAgents #ChangeManagement

At SAP Connect this week, SAP announced that general availability of Joule Work, Joule Assistants and the SAP Business AI Platform starts this month, with a phased rollout. Our co-founder Maik Buchenau has spent more than two decades in enterprise IT, much of it in SAP release and change management. His read on the part the keynotes skip: Joule's capabilities arrive first in the latest cloud releases. If your landscape is two upgrades behind, your AI roadmap is two upgrades behind too. Four unglamorous things decide who actually gets value: 1. Release cadence. Customers who defer updates defer AI. 2. Clean core. Joule is strongest on standard SAP processes. Analysts report weaker capability in heavily customised landscapes, 3. Master data. An agent that prioritises collections on bad customer data just makes the wrong calls faster. 4. Change control for agents. Agents are starting to act inside finance, HR and procurement. They need the same discipline as transports: test, approve, document, roll back, The reality check: in the DSAG Investitionsreport 2026, among members with AI use cases implemented, 77% run non-SAP AI in production and only 3% run SAP AI. DSAG points to licensing complexity and heterogeneous landscapes, In other words: the foundation, not just the technology. None of this is new work for SAP teams. It just got a much bigger payoff. What's the biggest blocker in your landscape; release backlog, custom code or data? #SAP #Joule #SAPConnect #S4HANA #AIAgents #ChangeManagement linkedin.com/posts/sap-joule…

What if intelligence could take form? Born from light. Shaped by energy. Connected to something beyond the visible. This is not a story about the future. It is a glimpse of what is already emerging. Novaeon.Studio #NovaeonStudio #AI #GenerativeAI #Innovation