PRISM
How we run AI inside PE-backed companies. One program goes in. A spectrum of working systems comes out — finance, IT, legal, operations — each with telemetry the CFO can defend to the sponsor.
Every function is a band of the same light
Who We Work With
PRISM runs inside their portfolios — and inside the companies they’re about to buy.
Why a methodology
Every enterprise AI rollout fails the same three ways.
We've watched it happen from inside dozens of PE-backed companies. The tooling is never the problem. The rollout is.
Shadow AI persists
Company data is already flowing into public chat tools. Blocking access doesn't stop it — it hides it.
Open-canvas paralysis
A license and a blank prompt box is not a program. Usage spikes in week one and collapses by week six.
No function-specific value
The close doesn't get faster because everyone sat through an "AI 101" session. Value lives inside specific workflows.
The core idea
The Arc
Ninety days to a working program. Six months to an operating system.
Enablement is the wedge. The Company OS is where the value compounds. Track one is the prerequisite for track two — companies that skip it stall every time.
Foundation
SSO, audit logging, and data-loss controls from day one. Role-based model access. Shadow-AI exposure closes in week one.
Function workshops
One function, four weeks, six steps — two at a time, staggered. Each ships 5–8 named skills with an eval set and an owner behind every one.
Telemetry & governance
Adoption dashboards, unit economics, confidence-gated actions, audit trails. By week six the CFO has numbers to defend to the sponsor.
Company OS
The spectrum recombines: a trained workforce, a scored process inventory, and a governed skill library become one layer over every system.
Pillars & Process
Every workstream. Six phases. One chart.
Governance and security start day one. Compliance, people, and budget run parallel with the pilot — so nothing blocks it, and nothing gets skipped.
Each pillar hands off to a named owner as the program matures — by the operate phase, your program owner and champions run the cadence and we step back.
The Unit of Work
One function. Four weeks. Six steps.
The same repeatable engine behind every band of the spectrum. It has run inside retailers, manufacturers, insurers, healthcare networks, and software companies — the steps don't change.
Pre-engagement
Week −1A 45-minute scoping call with the function leader. Out: interview roster, access list, success criteria, a named champion.
Discovery week
Week 18–12 role interviews plus shadowing. Out: a process inventory of 30–60 scored entries and a pain-point heatmap.
Prioritization
End of week 1Every process scored on value × AI-fit × frequency. Commit to the top 5–8; log the rest as wave two.
Build
Weeks 2–3Each process becomes a named, callable skill — piloted with power users until they say "ship it."
Enablement
Week 4Small groups, max eight. We train the skills, not the tool — when to use each, what to escalate.
Office hours
Weeks 5–8A weekly hour to iterate on what shipped and surface wave two. Problems get fixed the week they appear.
Two consultants run two function workshops in parallel, staggered by two weeks — that's how a full program lands in 90 days instead of two quarters.
The Spectrum
Four functions. One foundation.
Each band gets a deep playbook — workflows, skills, worked examples, numbers. Governance, telemetry, and enablement run underneath all of them.
The Foundation — under every band
Quality Bar
Anatomy of a shipped skill.
A skill isn't a prompt someone wrote once. Every skill that reaches production carries seven artifacts — versioned, owned by a function champion, and re-tested on every change.
Prompt template
The workflow itself, versioned like code.
Integration adapters
Wired to the systems where the work lives.
Eval set
10–30 golden cases, run on every change.
Confidence thresholds
Auto-route, ask, or escalate — by score.
One-page user doc
When to use it. When not to.
Training video
Five minutes, hosted where people learn.
Named owner
A champion plus a feedback channel.
Audit trail
Who ran it, on what, with what result.
Proof
The pattern holds across industries.
Anonymized outcomes from recent PRISM programs. The industries change. The shape of the result doesn't.
Mid-market insurance group
Underwriting cycle time down 62%. The governance framework became an audit asset in their next regulator review.
Regional healthcare network
Claim-denial rate down 38% from two functions and eleven skills, shipped in 60 days.
Diversified industrials holding
Four operating companies replicated to 52 skills with 88% cross-OpCo adoption — then a HoldCo-wide Company OS.
Commercials
Priced like an operator, not an hourly vendor.
Three shapes, matched to the engagement. All of them put our fee at risk against the outcome — none of them are time and materials.
Fixed-price phases
One number per phase. A 90-day program is typically $260–420K depending on function count and system depth — scoped in writing before work starts.
Standard for adoption programs
Outcome-priced
You pay against a measured baseline of hours saved or tickets deflected, with the methodology agreed in writing in week one. If the agents don't move the numbers, you don't pay for them.
Standard for agent builds
Production-paid
Invoicing begins when the build is live in production with a real user. Walk away earlier and we eat the cost. Phase two only gets priced after phase one proves out.
Standard for product builds
The exit is part of the design. Post-launch, most teams keep a quarterly sprint or office-hours retainer — but our goal is to make ourselves unnecessary for the day-to-day.
Ready to move
Start with one band of the spectrum.
Pick the function that hurts the most. We'll come back with a process inventory, the first wave of skills, and the number we'd be accountable to.
Talk to LightCI