← Back

AI Workflow Pre-Mortem · UiPath Maestro BPMN

FDE Agent

A pre-flight check for enterprise AI workflows — diagnoses where an AI workflow will break before it ships, and stops for a human where it matters.

3rd Prize — Maestro BPMN track

UiPath AgentHack 2026 · 333+ submissions from 104 countries · live finalist presentation to the judging panel, July 2026.

MIT · 2026

What it is

FDE Agent is a pre-mortem for agentic AI workflows. Hand it an AI-augmented business process (BPMN 2.0, Mermaid, or an image) and it scores every node on three risk axes — general failure, security, and human handoff — using a proprietary mapping ontology grounded in 7,959 AI-incident vectors (AIID), OWASP LLM Top 10, MITRE ATLAS, and the MIT AI Risk taxonomy. The output is a RED / AMBER / GREEN heatmap, multi-option mitigations per critical node, and an EU AI Act / K-PIPA compliance trail.

FDE Agent · pre-deployment diagnosis

Three nodes would not survive a pre-deployment audit — and all three sit on the decision path, where a wrong call becomes a legally significant one.

Critical

3

Caution

4

OK

4

Avg composite

2.4 /5

Node risk heatmap

FailSecHand

Intake

Enrichment

Decision

Output

N7 · Auto-Approve / Decline

EU AI ActNIST AI RMFK-PIPA

Why it scores this way

A legally significant decision issued with no human in the loop. Under EU AI Act Annex III §5(b) this is a high-risk system that requires human oversight (Art. 14).

Real-world precedent

Apple Card credit-limit disparity · 2019 · AIID-92

Opaque credit scoring issued lower limits to women with comparable finances; NYDFS opened a probe.

Recommended mitigation

Insert a human approval gate for declines and edge cases — routed through UiPath Action Center before the decision is final.

Human sign-off required before this node completes
Interactive sample of the diagnosis output — select any node. Scores, framework mappings, and incident citations are the demo fixtures from the reference implementation.

Why it matters

A CIO who has to defend an AI deployment to a board or a regulator needs to prove the workflow won't fail in a way they'll have to answer for. Today that means a weeks-long, five-figure pre-deployment audit. FDE Agent turns it into one automated run — with the judgment calls still routed to a human through UiPath Action Center before anything auto-approves.

How it works

The diagnosis itself runs as a UiPath Maestro BPMN process: the thing it diagnoses is a BPMN workflow with AI inside, and the thing that performs the diagnosis is a BPMN workflow too. A Python coded agent parses the input process, maps each node against the risk ontology, retrieves matching real-world incidents via RAG, and computes handoff metrics (including a confidence-decay signal). When a handoff carries real risk, the Maestro run pauses on an Action Center task and waits for human sign-off. The reference vertical is Korean credit underwriting — a working loan-diagnosis sample that maps directly onto the EU AI Act Annex III credit-scoring high-risk case.

Recognition

3rd Prize, Maestro BPMN track at UiPath AgentHack 2026 — selected from 333+ submissions across 104 countries, after a live finalist presentation to the judging panel (July 2026). The Maestro BPMN track judged agentic workflows built on UiPath’s BPMN orchestration layer; FDE Agent placed by running its own diagnosis as a Maestro process.

Links