70% of PE investors have walked away from a deal over AI exposure, and 40% apply a valuation haircut when digital maturity lags. The audit trail you build in month one is part of the exit story in year four.
Governance isn’t a page in the program. It’s the floor under every band.
The Operating Model
Oversight, accountability, and execution live at different altitudes — and the whole thing fails if the middle tier is missing.
IT, legal, HR, finance, security, plus a business sponsor. Approves policy, budget, and risk exceptions.
Meets monthly · owns policy, budget, exceptions
One named person, near-full-time. The role most programs skip — and why rollouts stall.
One named person · weekly cadence · single point of contact
Two to four people across IT, data, and the business — tools, templates, training, first-line triage.
2–4 people · tools, templates, training, triage
Unblocks budget and cross-department resistance — distinct from the program owner.
Who sets risk tiers, who vetoes a use case, who owns the budget line.
Shipped in the first two weeks — shadow use needs guardrails on day one.
The Acceptable-Use Policy, Page One
Five clauses anyone can hold in their head — shipped before the first skill goes live.
Everything not on the list is flagged on sight.
Customer records, financials, NDA material — in plain language.
Company work never touches personal AI accounts. Zero exceptions.
When AI materially affects a decision, the person is told.
One “can I use this?” channel — answered in a day.
Why it ships first
The Shadow-AI Audit
Before governing what’s planned, we find what’s running — three passes, each catching what the others miss.
Departments self-report under amnesty framing. Nobody gets punished for honesty, so the answers are honest.
AI subscriptions already on corporate cards — counted in the program’s cost picture from day one.
Sanctioned and unsanctioned tools visible at the boundary — the pass that catches everything unmentioned.
What we find: dozens of tools, concentrated where the repetitive pain is worst. That concentration is the program’s demand map — start there.
of shadow AI eliminated at a $1.5B distributor — the sanctioned skills were simply better
unsanctioned usage on company data at a manufacturer, once approved tools beat the workarounds
Risk Tiers
A draft email and a credit decision don’t deserve the same checkpoint.
Drafts, summaries, internal search. A human sees the record.
Anything leaving the company. A human owns the send button.
Hiring, credit, legal, medical, safety. No confidence score overrides this tier.
The Enforcement Machinery
A policy document controls nothing — these run on every skill, every day.
High confidence executes and notifies. Medium pauses. Money and production route through a human.
Who, when, inputs, outputs, confidence — on every run. Nothing happens off the record.
Golden test sets per high-stakes skill. Nothing promotes to production without passing.
Bi-weekly vetting of new and live skills. Problems surface in days, not quarters.
EU data on EU infrastructure, US on US. Residency is architecture, not a contract promise.
No silent model swaps. Always answerable: which version decided, and who approved.
Every action attributable · every change gated
The Skill Review Board, Agenda in Hand
Six standing items, in order — a board without a fixed agenda becomes a status meeting.
Failed golden sets stay out of production. No exceptions.
Every escalation, override pattern, and near-miss on the table.
Pilot-to-production candidates, seven artifacts checked each.
Flagged, explained, or investigated — nothing rides unexamined.
New use cases scored against the risk tiers.
Vendor changes queued behind the eval gate.
Function champions, the program owner, the security lead, the executive sponsor.
Every decision logged — and the log itself is a diligence artifact.
After Launch
The failures that reach the board happen in month seven — after a vendor update nobody read.
Vendor
Model update ships
4.1 → 4.2
Eval gate
27/27 golden cases
Versioned rollout
v4.2 pinned · approver logged
Held + alert
program owner paged · v4.1 stays live
The AI Incident Runbook
What separates a contained incident from a board escalation is a runbook written in advance.
Data egress, harmful output, or drift. Three categories — triage takes minutes, not a meeting.
The skill is paused by configuration — one switch, effective everywhere.
The audit log answers who, what, when, which version — in minutes.
The failing case joins the eval set permanently. Same failure never ships twice.
Legal decides notification from a prepared matrix — not from scratch at midnight.
Vendor Concentration
The continuity discipline you apply to your ERP applies to the model behind your collections desk.
Every AI-dependent workflow rated for what breaks if the vendor disappears
A documented, tested fallback for each critical path
Deprecation notice, pricing caps, data portability — negotiated up front
Model-training opt-outs confirmed in writing
AI vendors folded into existing DR and BCP exercises
The Money
ROI is defined before the pilot, or it’s a story — not a number.
Integration, training, governance, legal review — the license line is usually the smallest number on the page.
Hours saved, cost avoided, error reduction — agreed in writing before the pilot, never after.
Scale dollars release only against proven numbers — a demo doesn’t unlock the second tranche.
Weekly dashboard, monthly committee review, quarterly board update — shadow spend counted in.
Spend Governance
Without per-workflow attribution, AI spend silently erodes the margin it was meant to expand.
Every skill invocation carries its cost — expensive workflows visible while they’re still cheap to fix.
Cost per invoice coded, per ticket deflected, per memo drafted — denominators the CFO already thinks in.
Which model handles which tier is a documented decision with caps and alerts.
The monthly report explains what moved and why — a number that never surprises the board.
The dials behind these reports live in the telemetry layer.
The Quarterly Board Pack
All the machinery above compresses into one document — six items, no appendix.
Five pages the CFO presents without us in the room — that’s the deliverable.
Anchor Frameworks
We anchor to the standards a diligence team looks for — so the answer to “are you aligned” is yes, with evidence.
AI Risk Management Framework
Govern, map, measure, manage — the de facto reference for enterprise AI risk in the US.
AI Management System
The certifiable AI management system — alignment now is cheaper than a scramble at exit.
LLM Security Guidance
The security baseline for prompt-injection and output-handling — failure modes unique to language models.
In regulated industries, our validation aligns to bank model-risk guidance (OCC SR 11-7). One insurance client’s framework became an audit asset — the examiners asked for a copy.
The Exit Lens
The question isn’t whether your AI program gets examined at exit — it’s what the examination finds.
“A compelling AI narrative” with nothing underneath.
A deck with no evidence trail doesn’t just miss the premium — it invites the haircut.
The whole point
The Rest of the Spectrum
Ready to move
We’ll map your risk tiers, name the owner, and stand up the audit trail — defensible from the first pilot to the exit data room.
Talk to LightCI