Concept card

Human-in-the-Loop Checkpoints for Autonomous Workflows

As of 2026-10-06

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is Human-in-the-Loop Checkpoints for Autonomous Workflows?

Human-in-the-loop in SAP agents is concrete: skill tools can require user confirmation (Yes, proceed / No, cancel), content-based agent tools carry human_approval_required, standard agents pause for decisions or produce proposals for review, and deployments pass a human approval gate. Choose the checkpoint by reversibility and volume, and design it to meet EU AI Act Article 14 where it applies.

What it is

A human-in-the-loop (HITL) checkpoint is a point where an agent stops, shows a person what it intends to do or has prepared, and continues only on that person's decision. In SAP's agent stack this is no longer an abstract principle: several mechanisms are documented, and SAP's own agents use them. This card lists them, then gives the design rules for choosing and calibrating checkpoints.

Why it matters

SAP gives you concrete HITL controls (skill-tool confirmation, human_approval_required, standard-agent pauses, deployment and verification gates); designing without them wastes what the platform already provides.
Checkpoint placement decides both risk and value: too few and errors reach the ledger, too many and the agent delivers no autonomy.
For HR and credit uses, Article 14 oversight requirements make the checkpoint design an auditable compliance artefact.

Key points

Agent builder: 'Require user confirmation before running' on skill tools, with a prompt for the approval message; Yes, proceed / No, cancel quick replies.
Content-based agents: human_approval_required and human_approval_prompt per tool.
Standard agents: PPO agent pauses with quick replies or alternatives; accruals agent proposals are reviewed before posting.
Lifecycle gates: GitHub Actions approval for production (new Joule Studio); verification before publication in SAP AI Agent Hub.
Match checkpoint to reversibility and volume; confidence-threshold routing and drift rules are patterns you implement, not Joule settings.
Three distinct control points, not one feature: per-tool approval (inside a turn), agent-initiated pauses for missing input (mid-plan), and lifecycle/version gates (outside any conversation) — verify all three, not just one.
SAP documents that an agent treats notifying the user as the final action of its turn; design approval steps as a turn boundary, then resume, rather than instructing 'inform and continue'.
EU AI Act Article 14 requires the approval prompt itself to let a reviewer understand, interpret, override or stop the system — a content-free 'Proceed? Yes/No' satisfies the mechanism but not the substance of oversight.

Terms used on this page

Human-in-the-loop checkpoint: Point where an agent pauses and continues only after a person approves, rejects or completes its proposal.
human_approval_required: Tool property in Joule content-based agent definitions that forces user approval before the tool runs.
Approval fatigue: Degradation of review quality when approvers see mostly routine items and approve reflexively.
Article 14 (EU AI Act): Human-oversight obligation for high-risk AI systems: understand, interpret, override or stop the system.
Lifecycle gate: A checkpoint on an agent's version rather than on a single run — for example a GitHub Actions production-deployment approval or SAP AI Agent Hub verification before publication.
Confidence-threshold routing: Pattern that sends an agent's low-confidence outputs to a human reviewer while letting high-confidence ones proceed automatically; implemented in custom tool logic or workflow, not a standard Joule setting.
Turn boundary: The end of one agent conversational turn; SAP documents that notifying the user is treated as the agent's final action, so approval steps should be designed to end a turn rather than continue within it.
Automation bias: The tendency of a human reviewer to over-trust an AI-generated proposal, named explicitly in Article 14 as a risk that checkpoint design must counter.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →