DealDesk Agent

Inspiration

Every sales organization has a Deal Desk — the person or team that reviews pricing exceptions, emails managers up the chain, waits for replies, consolidates decisions, and updates the CRM. It is 100% manual, 100% inbox-driven, and almost always slower than the sales team needs. I wanted to find out whether an AI agent could actually do that job — not just summarize a deal, but reason about it, decide who needs to sign off, route the request, and close the loop — with real governance and a real audit trail.

What it does

Deal Desk Agent is an AI Deal Desk analyst that runs end to end on UiPath Automation Cloud.

Collects live Salesforce deal data and resolves the approver org chart (AE → Manager → Director → VP Sales → CFO → CRO) from HiBob.
Scores risk with deterministic business rules, then uses an LLM to draft a contextual recommendation and rationale — why, not just approve/flag.
Auto-approves low-risk deals with zero human touch.
For everything else, runs a sequential multi-instance approver loop, notifying each approver on three channels at once: an interactive Outlook Adaptive Card, a Slack Block Kit DM, and a native Action Center task. A decision on any channel resolves the step.
A rejection short-circuits the rest of the chain; the agent interprets the collected decisions to determine the final outcome.
Writes a full Data Fabric audit record (agent recommendation vs. each human decision, comments, timestamps) and sends the requester a Slack DM + Outlook summary.

How we built it

The brain is three coded agents (Python / LangGraph): plan (data + scoring + rationale), render (approval payload), and process_response (final outcome). Maestro BPMN 2.0 orchestrates everything — a disposition gateway plus the sequential multi-instance approval loop. A UiPath RPA robot (WaitDecision) fans out to all three channels and suspends/resumes via UiPath Persistence, so there is no polling loop in the BPMN. Live data comes through an AgentHub MCP server wrapping a read-only Salesforce Integration Service connection (getSalesforceOpportunity, getSalesforceAccount, searchSalesforceSoql). The whole thing — agents, BPMN, robot, MCP, bindings — was authored with Cursor + Claude via UiPath for Coding Agents and the uip CLI.

Challenges we ran into

Multi-channel HITL that stays in sync. Getting Outlook Adaptive Card, Slack, and Action Center to represent the same pending decision — and letting any one of them resolve the wait — took careful coordination between the response token, the HITL service, and the robot.
Suspend/resume instead of polling. Replacing a polling loop with UiPath Persistence so the BPMN truly idles while waiting on a human was a key design shift.
Real Adaptive Cards in email. Action.Http buttons, originator registration, and a rich HTML fallback for clients that don't render cards.
Keeping connection IDs and bindings real. Phantom/placeholder IDs broke Studio Web validation; everything had to reference the actual tenant connection IDs across the BPMN, bindings, and debug overwrites.

Accomplishments that we're proud of

A genuinely agentic flow: the agent reasons, decides autonomously on low-risk deals, sizes the approver chain from live org data, and interprets the human decisions at the end — it stays in the loop throughout.
Four simultaneous live test runs verified on all three channels (Outlook Adaptive Card, Slack DM, Action Center) on Jun 22, 2026.
A clean separation of concerns — the right actor for every step: agents reason, the robot communicates, the BPMN governs, Data Fabric remembers.
The entire solution was built with coding agents on the UiPath platform.

What we learned

The strongest agent designs combine deterministic rules with LLM reasoning — rules for policy thresholds, the LLM for context and explanation.
BPMN multi-instance loops are a natural fit for variable-length approval chains produced as data, with no hardcoded branches.
Meeting approvers where they already are (email, Slack, the UiPath portal) matters more than any single "perfect" UI.
Coding agents (Cursor + Claude) plus the uip CLI can take a UiPath solution from idea to a deployed, governed automation surprisingly fast.

What's next for DealDesk Agent

Deeper CRM write-back so approved terms flow straight back into Salesforce.
Configurable, tenant-specific policy thresholds and approval matrices without code changes.
Learning from historical decisions to refine the agent's recommendations over time.
Analytics on the Data Fabric audit trail (cycle time, override rates, bottleneck approvers) and expansion beyond renewals to new-business pricing exceptions.