Agent Flows: Fast Multi-Stage Approvals

Key insights

• Agent Flows
  Introduce AI into approval processes by using the Multistage Approval action. They combine AI prompts and vision with human review to make approvals faster and more flexible.
• Multistage Approval action
  Let you create multiple approval stages with different stakeholders and rules. Stages can move forward, skip, or stop based on the request and conditions you set.
• AI approval stage
  Processes inputs such as documents, images, and text to return clear Approve or Reject decisions with rationale. You can route AI results into later human review stages to keep control.
• Human-in-the-loop
  Keeps people involved for complex or edge cases while letting AI handle routine checks. This balances speed with oversight and compliance.
• Conditional routing
  Use conditions to automate decisions, skip stages, or change paths based on criteria. Many skills and patterns transfer from Power Automate, so existing experience helps with design and migration.
• SharePoint attachment and demo steps
  The demo routes an Expense Report stored in SharePoint through AI and human stages. Key steps: add inputs, test the AI stage, add a manual stage, retrieve the SharePoint attachment, ensure inputs are strings, create a test record, and watch for common input bugs.

Summary of the Video

Shane Young [MVP] published a practical walkthrough that demonstrates how Copilot Studio and Agent Flows enable advanced Multistage Approval workflows. In the video, he shows an automated approval pipeline that routes an Expense Report stored in SharePoint through AI and human review stages based on configurable criteria. The presentation includes a live demo, step-by-step configuration, and comparisons with existing automation tools to help practitioners understand how their skills transfer. Overall, the video positions these features as a public preview capability that blends automation speed with human oversight.

Demo Walkthrough and Key Steps

Shane begins with a clear demo workflow, noting the timestamps at key steps such as setting up the Multistage Approval action, adding inputs, and testing the AI stage. Then he shows how to add a manual stage so that human approvers can review decisions flagged by the AI, and demonstrates retrieving a SharePoint attachment during the process. He also highlights practical issues like ensuring inputs are string-typed and creating a test record in SharePoint to validate the flow. Finally, he reproduces an input-related bug to illustrate common pitfalls and how to spot them during testing.

How AI and Human Approval Work Together

The video emphasizes a hybrid approach in which AI stages handle routine, rule-based checks while human approvers address complex or ambiguous cases. For example, the AI stage can analyze document content and vision inputs to return a clear Approve or Reject decision with a rationale, and subsequent human stages can review those outputs before finalizing. This human-in-the-loop model aims to speed up processing for straightforward claims while preserving human judgment where context or nuance matters. Consequently, organizations can reduce cycle times without ceding control over important or high-risk decisions.

Conditional Routing and Dynamic Logic

Shane also covers conditional routing, which lets workflows follow different paths based on approval outcomes or data-driven criteria. In practice, conditions can move a request forward, skip stages, or terminate a process, which adds flexibility compared to rigid sequential approvals. However, he notes that more dynamic paths increase configuration complexity and require careful testing to avoid surprises in production. Thus, teams must balance the desire for flexibility with the operational overhead of maintaining multiple branches and conditions.

Tradeoffs and Practical Challenges

The video candidly discusses tradeoffs between automation and control, including accuracy, speed, and maintenance costs. While AI stages can accelerate routine approvals, they require good prompts, quality inputs, and periodic retraining or tuning to remain reliable; otherwise, false positives or negatives may increase review workload. Moreover, handling attachments and data type mismatches—like ensuring inputs are strings—can introduce subtle bugs that disrupt flows, so testers must validate each input path carefully. Therefore, teams should budget time for testing and monitoring, and accept that initial setup may be heavier than classic single-stage flows.

Implications for Power Platform Practitioners

Shane points out that many skills from Power Automate transfer to Copilot Studio and Agent Flows, which reduces the learning curve for existing practitioners. Nevertheless, the new AI stages introduce design patterns around prompt engineering, vision inputs, and human-review strategies that teams must adopt. As a result, organizations that already use the Power Platform can extend existing automation while building new competencies in AI governance and testing. Consequently, teams should plan training, governance, and staging environments to scale safely.

Final Observations and Recommendations

Overall, the video provides a balanced, hands-on look at how multistage approvals can modernize organizational workflows while exposing practical implementation concerns. For organizations that process many routine requests, adopting AI stages can yield measurable time savings, but they must pair that with human review and monitoring to manage risk. Therefore, readers should view Multistage Approval as a tool that requires deliberate design, incremental rollout, and ongoing maintenance to be effective. In short, the demo offers useful guidance for teams planning to pilot or adopt these capabilities within their business processes.

Keywords

multistage approvals, agent flows, approval workflow automation, multistage approval process, agent-based approvals, automated approval flows, approval flow best practices, enterprise approval automation