01/17
P notes  ·  → advance
Lumi AI
Lumi AI · April 2026

AI in Practice.
Lessons from 50+ agents in production.

Case studies, failures, and the practices that hold up.
Amar Naga + Greg Asher Lumi AI
The reframe
Don't ask
How do we use AI?
Ask
Which problem
costs us the most?
Every working deployment started with a business leader who knew exactly what was broken.
Same situation. Different lens.

Three re-writes. Same work, new frame.

01 Starting point
Pilots built around what AI can do.
Start with the problem that costs the most.
02 Ownership
IT builds it. The business inherits it.
The team that owns the problem owns the solution.
03 Success metric
Automation rates look great on the dashboard.
Monday morning has to feel different.
Production lessons

Four patterns we kept seeing.
We now design around them from day one.

01

Ops team wasn't in the room

The people doing the work define the handoff. Not the project team.
CIO IT SOLUTION OPS ISOLATED
02

Pilot worked. Rollout didn't.

Two-week prototype on real data. Always.
PILOT REAL DATA MISSING FIELDS · FORMAT BREAKS · SURPRISES
03

No kill switch. No audit trail.

Both go in before anything else.
AGENT RUNNING ACTION ACTION ACTION no off button SHUT DOWN PERMANENTLY NO WAY TO INVESTIGATE · NO TRUST RECOVERY
04

Measuring the wrong thing

The business team sets the metric. They know what "correct" looks like.
TARGET AUTOMATION RATE WRONG EXCEPTIONS
"The organizations that get AI right treat it as a business change, owned by the team doing the work, measured by the outcomes they care about."
Business problems we've solved

The highest-value use cases aren't glamorous.

Real-Time Incident Response
45 min manual reporting
3 min
Voice-first. Bus accidents triaged by agent, notifications cascaded to ops in real time.
AI 3 MIN
Agent Fleet Orchestration
$620K inference spend, 4 platforms
40% less
OpenAI, Claude, Copilot, custom models. One governance layer across every AI investment you've already made.
CONTROL PLANE GPT Claude Copilot Custom -40% COST
AP Financial Intelligence
$218K exposure seen
$2.16M
Same invoices, same month. Agent sees patterns across invoices, emails, contracts, and vendor data.
AI EXPOSED $2.16M
Workforce Compliance
Monthly checks
Every 15 min
About 3,200 workers and 10,000+ certifications across 12 sites. Gaps flagged before workers reach the gate.
! 15 MIN
"None required new systems. Once the pattern works, it scales. The same agent built for one operation can be packaged across a group or taken to market."
Change Management

Most AI projects fail in the org, not in the model.

Three tracks. All from day one. Running in parallel.

01
Govern

Who decides.

Org alignment & decision rights.

  • Named business owner per agent
  • Build / halt / KPI rights defined upfront
  • Change champions embedded in each BU
02
Prepare

What we need.

Ecosystem & infra readiness.

  • The right data, not perfect data
  • Sidecar. Teams keep their tools
  • Security defined before build
03
Deploy

How it runs.

Use-case guardrails & trust.

  • Operating envelope signed off pre-build
  • Kill switch + audit from day one
  • Shadow → Assisted → Autonomous
The sidecar is a change management choice, not a technical one. It's the difference between adoption and shelfware.
What we'd build differently

Five principles. Learned the hard way.

01
Orchestration is the real work, not the model.
02
Guardrails are architecture, not afterthought.
03
No code inside your systems.
04
Build on top of existing AI investments.
05
Human-in-the-loop is a design principle.
"Every one of these came from a deployment that taught us the hard way. They're exactly why Argenbright looks less like one project and more like a portfolio pattern."
Where we see the opportunity

Argenbright is the kind of operating environment where the pattern we've built compounds fastest.

Scaled frontline teams

Multi-site reach

Compliance rigor, built in

Diverse operating workflows

Leaders focused on outcomes

Not one use case. A pattern worth building together.
Where we'd explore first

Start where pain is sharpest, outcomes are measurable, and the pattern can scale.

01
Pick one operating wedge
One workflow. Visible pain. Clear owner.
02
Measure what matters
Time, exceptions, quality, compliance. Owned by the team doing the work.
03
Build for reuse
Design so the pattern scales across adjacent businesses.
"The goal isn't more pilots. It's to find a pattern worth proving, then make it scale."
Over to you

Three questions we'd love your perspective on.

01
Which portfolio pattern feels most urgent in your business today?
Workforce coordination · exception handling · proof & compliance · operating visibility.
02
For Unifi, Velociti, or OutScale, which wedge feels closest to live pain?
Scheduling, turnaround, PRM. Or vendor control, proof, mobilization.
03
Where do you need orchestration on top of the AI you already have?
Not a replacement stack. The layer where approvals, exceptions, and controls still live.
Not a pitch. A conversation. Where should we focus together?
Lumi × Argenbright

Thank you.

One problem. Eight weeks. Measurable results.
Let’s find the right place to start.

Amar Naga
amar@lumicorp.ai
Greg Asher
greg@lumicorp.ai
Website
lumicorp.ai
Speaker notes