Who runs your AI agents after they go live?

AI agents have a failure surface that traditional SAP AMS and DevOps do not. Token spend drifts as conversation context grows. Model behaviour shifts on vendor-side version updates you did not schedule. Hallucinations and tool-call errors land in real SAP transactions. Compliance evidence needs to capture not just what the system did, but what the reasoning chain was when it decided. AI Agent Operations is the discipline of running those concerns in production: observability over token cost, latency, accuracy, and tool-call success; governance for every proposed action before it commits to SAP; auditing that produces tamper-evident trails the regulator can read; and lifecycle management for prompt versions, model versions, retrieval indices, and grounding data. Treating it as a feature of standard AMS misses what is actually breaking — which is rarely the ABAP layer.

Both, under one operating model. The five-layer agent architecture documented on the DEBCOR AI Platform page is the same whether agents were authored on Joule Studio 2.0 (RISE customers' native track) or assembled externally on LangGraph with Anthropic Claude over MCP (the track that runs everything Joule does not cover natively yet). Operations land on the platform layers consistently: Layers 4 and 5 — Governance and Auditing — become ongoing services rather than one-time deliverables. The Company Intelligence Layer compounds knowledge of the customer's specific landscape across both tracks, so a behaviour learnt while operating a custom procurement agent informs how a Joule finance agent is grounded. This is what makes a single operating contract across both build tracks honest.

Same vocabulary as DEBCOR's Core SAP and BTP Runtime managed services. Silver is watch-the-system — monitoring, incident response, token-spend reporting, weekly health reviews. Gold is tune-the-system — Silver plus prompt and grounding refinement, exception-pattern remediation, accuracy and cost optimisation cycles, compliance attestation support. Platinum is evolve-the-system — Gold plus new agent development inside the operating budget, Joule entitlement activation, governance policy evolution as the SAP estate changes, and AI-platform roadmap fit with the customer's S/4HANA plans. Tiers are mix-and-match across landscapes — Platinum on AI Agent Operations with Gold on Core SAP and Silver on BTP is a common pattern. The Coverage Matrix on this page details what is and is not in each tier line by line.

Model and platform versioning is part of the operating contract, not an excuse. For Anthropic Claude and Google Gemini (via SAP's Generative AI Hub) the version pin is explicit per agent — version changes are tested in a staging environment against a regression suite of the customer's actual transactions before promotion, with rollback always available. For SAP Joule, where SAP controls the underlying model and quarterly release cycle, DEBCOR participates in the Early Adopter Care programmes for Joule Studio 2.0 and the managed runtime, so customer impact is identified and remediated before it reaches the production landscape. Compliance evidence captures which model version handled which transaction, for the period the customer's regulator requires. The standard read: model change is a planned event, not an outage.

Four mechanisms, layered. First, scope: each agent is authorised against exactly the SAP transactions and master data it needs — a finance agent cannot read HR data, an AP-clerk session cannot post a journal it is not allowed to. That alone removes most blast-radius scenarios. Second, the Governance Agent layer (Layer 4) validates every proposed action by an Orchestration, Intelligence, or Worker agent before it executes against SAP — segregation of duties, approval thresholds, anomaly detection on transaction patterns. Third, grounding: the Company Intelligence Layer keeps agents anchored to the customer's actual chart of accounts, vendor master, pricing keys, and history; hallucination is much rarer when the model is reasoning over verified retrieved facts, not generating from prior. Fourth, audit and reversal: every action is logged tamper-evidently with reasoning chain; anomalous actions are flagged for human review before the next dependent action runs. Errors still happen — the architecture limits what they can break and ensures they are visible.

Both, depending on tier. Silver and Gold are business-hours operations with 24×7 incident response on defined-severity events. Platinum is 24×7 operations — appropriate when the agents are running revenue-critical transaction flow (order capture, payment processing, AP automation, customer-service tier-1 deflection) and an outage or accuracy regression has measurable P&L impact within hours. The honest framing: not every AI use case warrants 24×7 spend. A weekly close-acceleration agent doesn't. A real-time order-validation agent processing thousands of orders an hour does. DEBCOR's recommendation is anchored to where the AI sits in the business process, not to which tier sells. The Coverage Matrix on this page details monitoring windows, SLAs, and escalation patterns by tier.