UiPath AgentHack Submission - DealDesk Expert!

[Daniela_Rosenstein]

UiPath Forum Submission — Deal Desk Agentic Flow

Track

Track 2: UiPath Maestro BPMN — with the coding-agent bonus (built entirely with Cursor + Claude via UiPath for Coding Agents).

The problem — a story every revenue team knows

It’s the end of the quarter. A renewals analyst opens the pipeline and starts going down the list, one opportunity at a time. Most are fine. But buried in there is a $28K renewal with a 45% discount a rep promised to save the logo, and a $350K deal at 32% that quietly breaks the discount ceiling. Each one has to be eyeballed, judged against policy, and — if it’s a special case — emailed up the chain to a manager, a director, sometimes Finance. Then the waiting starts: replies trickle in over days, decisions get buried in email threads, and the rep keeps pinging “any update?”

That is the Deal Desk: the human who manually reviews opportunities, spots the ones that are problematic or at risk, and shepherds the exceptions through approval. It works — but it’s slow, it’s inconsistent from analyst to analyst, and it leaves no audit trail. It’s the most judgment-heavy, least scalable step between a quote and a signature.

What we built — a real deal desk, end to end

Deal Desk Agent is that analyst, as an AI agent — running end to end on UiPath Automation Cloud.

The moment a Salesforce opportunity changes, a Maestro BPMN process wakes up. A coded agent pulls the live deal, looks at the discount and the value the way an analyst would, and scores the risk — then writes a plain-English rationale for why this deal is fine, risky, or a hard no. Clean deals inside policy clear on their own, with zero human touch. The ones that genuinely need a human — the special cases — are routed to exactly the right approvers, sized from the live org chart, and each approver is reached on three channels at once: an interactive Outlook Adaptive Card, a Slack DM, and a native Action Center task. They decide wherever they already are; the same decision resolves the step. A rejection anywhere stops the chain. The agent reads the collected decisions, the rep gets a clean summary, and the entire trail — agent recommendation vs. each human call — is written to UiPath Data Fabric.

It is the manual deal desk, automated: collect, reason, route, sign-off, audit — without the inbox.

See the agent reason (real runs)

These aren’t mock-ups — they’re live runs from Jun 22, where the agent reviewed each opportunity and made the call before any human looked:

• StartupXYZ — $28K at 45% off → Reject. “Discount is margin-destructive for a deal this size; the savings don’t justify the erosion.” The agent caught the at-risk deal a tired analyst might wave through.

• MegaCorp — $350K at 32% off → Approve with conditions. “Exceeds the 25% ceiling — escalate for a multi-year commitment.” It reasoned about the policy threshold, not just the number.

• GlobalTech — $120K at 18% off → Approve. Inside policy, low risk, cleared cleanly.

• Pharma Global — $780K at 22% off → Approve. Large enterprise expansion, already Finance-pre-approved — recognized and fast-tracked with an audit note.

The design principle throughout: the right actor for every step. The agent reasons, the robot communicates, the AWS HITL bridge and Action Center handle the human wait, the BPMN governs, and Data Fabric remembers.

What makes it innovative

• Coded Agents (Python / LangGraph) — three agents carry the intelligence: plan (risk score + LLM rationale + approver chain), render (approval payload), and process_response (interprets each human decision). The agent stays in the loop end to end, not just at kickoff.

• AWS Lambda HITL bridge for the approval loop — a CloudFront + Lambda API (/api/v1/approvals) mints a response token per approver and accepts the decision through a secure token callback, making the wait instant and channel-agnostic.

• Action Center “wait for task” — a native UiPath external task; the WaitDecision robot suspends via UiPath.Persistence.Activities and resumes on decision, so an approver can act right inside the UiPath portal.

• Microsoft Adaptive Cards in Outlook — a real application/adaptivecard+json card embedded in email (originator 61fed71d) with Approve / Reject / Request Info Action.Http buttons and an HTML fallback for clients that don’t render cards.

• Slack Block Kit — the same deal data, recommendation, and rationale mirrored as a DM to both approver and requester, linking back to Outlook.

• UiPath Data Fabric audit — every terminal path writes an ApprovalAudit record (ordered trail, agent recommendation vs. each human decision, comments, timestamps, outcome). A queryable system of record, not a log line.

How it works

1. Trigger + plan — a Salesforce opportunity change starts the BPMN. The plan agent enriches data via the Salesforce DealDesk MCP (AgentHub) and resolves the chain: AE → Manager → Director → VP Sales → CFO → CRO.

2. Decide — deterministic rules (risk_score = min(1.0, 0.5·discount/25% + 0.5·value/$500k)) plus an LLM rationale. auto_approve clears low-risk deals; otherwise the approver loop begins.

3. Approve — per approver, render builds the payload, the WaitDecision robot fires all three channels, and the AWS Lambda HITL bridge mints the token. Any channel resolves the wait; a rejection short-circuits the rest of the chain.

4. Close — process_response interprets the decisions, the requester gets a Slack DM + Outlook summary, and Data Fabric records the audit trail.

Deployed on Automation Cloud in Shared/DealDeskApprovalGlobal: agent package adaptive-approval-agent-core v0.1.17 (3 entry points) · BPMN DealDeskApproval v1.1.9 · Robot WaitDecision v1.3.16 · Solution .uipx v1.2.3 · Salesforce DealDesk MCP AgentHub server live · AWS HITL bridge serving the approval API.

Real Runs with Mock Data

Four cases ran simultaneously; Adaptive Cards, Slack DMs, and Action Center tasks all verified.

Customer	Amount	Discount	Agent Recommendation
GlobalTech Solutions	$120,000	18%	Approve — within policy
MegaCorp Industries	$350,000	32%	Approve with conditions — escalation triggered
StartupXYZ Inc	$28,000	45%	Reject — margin impact too high
Pharma Global Ltd	$780,000	22%	Approve — enterprise expansion, Finance pre-approved

Built with

UiPath Maestro BPMN 2.0 · Coded Agents (Python / LangGraph) · UiPath RPA + UiPath.Persistence · Action Center · Data Fabric · Integration Service (Salesforce + Outlook) · AgentHub MCP · Microsoft Adaptive Cards · Slack Block Kit · AWS Lambda + CloudFront · Cursor + Claude via UiPath for Coding Agents.

Links & supporting files

Item	Link
Demo video	Watch Video
Presentation deck	View Deck
Code	View Code