Skip to content
All work

Enterprise AI Copilot · Agentic Workflows

Finn — the employee copilot that actually does the work

WEXDirector of Product for Agentic AI2023–Present

Most assistants answer questions; Finn finishes the job. It's the smartest coworker on the team — available 24×7, it understands you, your workflows, and your team's knowledge graph, and completes real work across core systems with a human in the loop on anything sensitive.

Or a quick task
Finn
Employee copilot · online
24×7
Acme Corp · Renewal sync · meeting starts in 30 min
Finn is gathering context
Google Calendar
Snowflake
Dynamics
Your brief · Acme Corp · Renewal sync
Calendar
  • 2:00 PM today · 30 min
  • 3 attendees · incl. their VP of Operations
Snowflake signals
  • Product usage −18% MoM — churn risk
  • Open opportunity: $420K
  • Renewal in 45 days
Dynamics notes
  • Last touch 12 days ago
  • Asked about volume pricing
  • 2 support tickets open
Recommended angle

Lead with the usage dip and the 45-day renewal; come ready with a volume-pricing option.

Sending to your Slack & email…
Proactive by design — Finn reads your day, gathers the facts from every system, and nudges you before it matters.
Run this agent loop live — tool use, approval gate, no sign-up
Role
Director of Product for Agentic AI
Organization
WEX
Status
In GA · 2+ years
Impact
$2M+ / yr in productivity
Availability
24×7 · human-approved actions
Most recent
Migrated to AWS managed agents

The problem

10,000+ hours lost to work no one should have to do

Employees were spending 10,000+ hours on trivial, repetitive tasks — requesting PTO, filing access tickets, chasing routine approvals. Zoom out and the same pattern repeats across the enterprise: manual workflows, everywhere, quietly taxing every team.

  • Cognitive load and constant context-switching across a dozen disconnected systems.
  • Slower approvals as routine requests sat waiting in queues.
  • Fragmented knowledge islands — the answer existed, but no one knew where to find it.
  • New employees hesitated to interrupt a peer for every small question or workflow.

None of it showed up cleanly on a P&L — which is exactly why it had been allowed to persist.

The idea

Finn doesn't describe the steps — it submits, routes, and tracks the request itself, 24×7.

It understands you

Your role, your context, and your history — so the help is personal, not generic.

It knows your workflows

And runs them end-to-end — submitting, routing, and tracking — instead of just describing the steps.

It taps your team's knowledge graph

The policies, docs, and systems that usually live in one person's head, available to everyone, instantly.

What I did

I led Finn from a blank-page discovery to a production copilot that completes real work across WEX's core systems.

Today Finn has been in general availability for over two years, handling real workflows for employees across the company — Workday approvals and PTO, Jira and access requests, expense, knowledge retrieval, and conference-room booking — with a human approving anything sensitive.

Most recently, I led the migration to a managed agent runtime on AWS to extend the copilot's scalability and give it a more robust foundation for the next wave of agents.

I was the product lead: I owned product strategy, the agent UX, and the evaluation framework, partnering closely with an ML platform team of ~6 engineers and our design team the whole way — from discovery through a rapid, tight iteration cycle.

Under the hood

How one request becomes finished work

An orchestrator plans each task and delegates to specialist agents. Reads run automatically; anything that writes to a system of record pauses for a human. Every step is logged and evaluated.

Employee asks — chat · web · mobile

Finn orchestrator

Plans the taskMemory & contextEvaluationsAudit log
Guardrails & policy applied to every step
HR agent
IT & access agent
Expense agent
Knowledge agent
Reads run automatically · writes pause for human approval
HRISWorkday-class
ITSMJira-class
CRMDynamics-class
Data warehouseSnowflake-class
Docs & policiesSharePoint-class
Peer agentsvia A2A

Systems of record

One request, end to end: plan → delegate → act — with every step logged, evaluated, and gated by a human where it counts. System names are representative, not literal.
See this loop run for real in the live demo

How it connects

One copilot, wired into the systems where work actually happens

Finn's value comes from reach. It completes tasks by acting across the enterprise's real systems of record — not by handing back a link.

Workday

HR approvals, PTO & time off

Atlassian / Jira

Tickets & service management

Jira Forms

Structured service-management intake

Snowflake

Analytics & enterprise data

Dynamics

CRM & customer context

Dovetail

Research & product insights

SharePoint

Policy source of truth

Agents via A2A

Agent-to-agent collaboration

Impact

Real work, finished

$2M+
in annual productivity impact
10K+
hours of trivial work targeted for automation
24×7
availability with human approval on sensitive actions
2+ yrs
in general availability across the company

Stack

AWS managed agentsGoogle Vertex AIGuardrailsPolicy layersMCPAgent-to-Agent (A2A)

Happy to walk through the real specifics, the agent architecture, and my specific role in a conversation.

Get in touch