Agentic AI Delivery Model

An agentic deliverable differs from a BI deliverable in one decisive way: the software acts. A dashboard informs a person who decides; an agent drafts, posts, releases or sends, sometimes after a confirmation and sometimes not. That shift changes scope, acceptance criteria, testing and after-care. This card describes how to structure delivery using what SAP now documents for Joule Studio and SAP AI Agent Hub, and how to reflect it in a statement of work.

The lifecycle SAP documents

SAP AI Agent Hub's step-by-step guide organises agent governance into five stages, which make a good backbone for a delivery plan:

1. Plan and Build: record architecture decisions (model vendor, agent framework, verification criteria, data-handling rules) and map each planned agent to the business capability, applications and existing MCP servers it will use, which tells you its risk profile and owners before you build.
2. Discover and Provision: bring agents into the inventory (SAP, custom, extended and non-SAP), assign owners.
3. Observe and Analyze: monitor usage and quality (telemetry from SAP Cloud ALM), business value and AI risk.
4. Secure and Govern: verify agents against criteria, publish verified agents to the AI Agent Portal.
5. Optimize and Decommission: retire agents that miss value targets or are unused, feed lessons into the next cycle.

Build and test inside Joule Studio

For custom agents, the build work follows the tool (C113):

Why it matters

• SAP now documents an agent lifecycle end to end (Agent Hub stages, Joule Studio steps, GitHub production gate); a delivery plan that ignores it will fail governance review.
• Agent UAT tests actions, tool choice and escalations on real data; test effort and data remediation must be priced upfront.
• Post-go-live monitoring and decommissioning are part of the deliverable, because agent errors compound after hypercare ends.

Key points

• Agent Hub lifecycle: Plan and Build → Discover and Provision → Observe and Analyze → Secure and Govern → Optimize and Decommission.
• New Joule Studio: generated tests (tool invocations, formatting, error handling), AI evaluation (intent alignment, requirements coverage, spec adherence), Try mode with traces.
• Production: developers build in non-production only; admin-triggered GitHub Actions workflow is the approval gate; Monitoring tab tracks failed sessions and workflow errors.
• UAT structure: trigger, context, reasoning and tool choice, action, checkpoint; each with happy, exception and adversarial cases on real data.
• SOW must include data remediation, AI Unit consumption, API-policy and AI Act checks, monitoring handover and data-use sign-off.
• Decommissioning is a lifecycle stage, not an afterthought: SAP's Agent Hub explicitly retires agents that miss value targets or sit unused, because idle agents still consume governance effort and, in some designs, AI Units.
• Agent UAT must run on the customer's real master data, not a demo dataset — a wrong vendor name is cosmetic in a BI report and a payment error once an agent can post.
• The continuous-improvement clause in Joule Studio's terms lets SAP process customer instructions, logs and feedback under contract; get data-protection sign-off during scoping, not after go-live.

Terms used on this page

AI-powered evaluation

Joule Studio step that scores a generated solution on intent alignment, requirements coverage and specification adherence.

Try mode

Interactive testing on the Joule Studio Solution page with a chat, a traces panel of tool calls and reasoning, and logs.

Approval gate

In the new Joule Studio, the administrator's manual trigger of the GitHub Actions workflow that deploys a solution to production.

Decommission

Agent Hub lifecycle stage in which inactive, redundant or rejected agents are retired from the inventory.

AI Agent Hub lifecycle

SAP's five-stage agent governance model: Plan and Build, Discover and Provision, Observe and Analyze, Secure and Govern, Optimize and Decommission.

Agent UAT

User acceptance testing structured around trigger, context, reasoning/tool choice, action and checkpoint, run with happy-path, exception and adversarial cases on real data — distinct from a BI report's UAT, which only checks a number.

Continuous improvement clause

Contractual term under which SAP may process customer instructions, configurations, execution logs and feedback from Joule Studio to improve the product.

AI Agent Portal

The publication point in SAP AI Agent Hub where agents that pass verification are made visible and available after the Secure and Govern stage.

Sources

1. SAP Help — AI Agent Hub step-by-step: Plan and Build
2. SAP Help — Observe and Analyze (telemetry, business value, AI risk)
3. SAP Help — Secure and Govern (verification, AI Agent Portal)
4. SAP Help — AI Agent Hub step-by-step: Optimize and Decommission
5. SAP Help — What is Joule Studio? (new, intent-based)
6. SAP Help — Joule Studio: Testing (automated tests, AI-powered evaluation)
7. SAP Help — Joule Studio: Deployment (managed runtime, GitHub approval gate)
8. SAP Help — Role Management for Joule Studio
9. SAP Help — Joule Studio: Continuous Improvement (customer data use)
10. SAP Help — Project Lifecycle Management and Versioning
11. SAP Help — Test Your Joule Agent
12. SAP Help — Joule Studio classic edition: Known Limitations
13. SAP Help — Accounting Accruals Agent (prerequisites, scope item J58)
14. SAP Help — Production Planning and Operations Agent
15. SAP Help — Joule Agents Overview in SAP SuccessFactors (assistants, permissions, commercial model)
16. ERP Today — What Customers Need to Know About SAP's API Policy (Sep 2026) — cited for: Agentic AI Delivery Model
17. SAP News Center — Autonomous enterprise, business transformation management and SAP AI agents at scale: AI Governance Assistant (2026-09-22)