All resources

AI & Automation · Resource 11 of 15

AI Orchestration Ledger

A governance framework for the agentic enterprise — the CoS's instrument for turning AI ambition into a governed operating model.

14 min readAI governanceOriginal framework

Resource 11 · Governance

The AI Orchestration Ledger

A governance framework for the agentic enterprise.

The Chief of Staff's instrument for turning AI ambition into a governed, productive operating model. Built for the reality of 2026: widespread shadow use, rising agent autonomy, and regulation catching up fast.

The AI-Pilled CoS resource covered how a Chief of Staff becomes personally AI-fluent. This resource covers the next problem: how the organization gets there, and why the operator in the room — not the CTO, not Legal, not the CEO alone — is the one who can actually run the instrument.

Before you read further

Two audiences, one instrument, applied differently

This resource was written primarily for operators inside traditional organizations — companies that grew by hiring humans into functions and now need to govern AI activity that is already happening underneath that structure. That is the majority case in 2026, and the framework below is calibrated to it.

If you are an AI-native founder or early operator, the Register still applies — but the surfaces you are registering are different. You are not inventorying what your sales team quietly pastes into ChatGPT. You are registering the agents you are already onboarding, their role definitions, their decision boundaries, the guardrails that keep them from sending the wrong ad live at 11pm on a Friday. The ledger is the same instrument. What fills it changes.

This resource fits you if

You run the operating rhythm for a team where humans are the primary actors and AI is being added underneath. You already suspect AI use in the company is wider than leadership can see, and you need to make it visible without punishing anyone for it.

You'll get value, but adapt it if

You are building AI-native from day one. Agents are on the team before humans are. The Register is still the right instrument — the columns still matter — but your population of entries is weighted toward agent specs and guardians rather than shadow discoveries. A companion resource on the Register for agent-first companies is coming next in this series.

How this resource is organized

Why → How → What → Apply

1

Why

The gap, who owns it, and which posture your org needs now.

2

How

The governance approach before any instrument — the paved road.

3

What

The Register, discovery by company stage, and a worked example.

4

Apply

Stewards, the operating rhythm, the disclosure dividend, a 30-day start.

Phase 1 · Why

The gap, the owner, the posture

Before any instrument: name the problem, claim the role, diagnose what your org actually needs.

The Problem

Four numbers that name the gap

The gap between "leadership wants AI" and "the organization is using it safely and productively" is not a technology gap. It is a coordination gap. That is the CoS's home field.

74%

of companies plan to use agentic AI at least moderately within two years.

Only 21% have a mature governance model for autonomous agents.

Deloitte, State of AI in the Enterprise 2026

98%

of organizations report some form of unsanctioned AI use by employees.

Shadow AI adds an average of $670,000 to breach costs.

CrowdStrike 2026 Global Threat Report · IBM 2025 Cost of a Data Breach

80% / 28%

of leaders say their org is mature at basic automation. Only 28% say the same for AI agents.

The gap between automation maturity and agent maturity is the operating problem.

Deloitte 2025 Tech Value Survey, n=550

Aug 2, 2026

EU AI Act enforcement begins.

Fines up to €35M or 7% of global revenue for non-compliance.

Regulation (EU) 2024/1689

Ownership

Why the Chief of Staff owns this

Three roles are already circling this problem and none of them can solve it alone:

The CTO

Sees: Infrastructure and models.

Misses: Cannot see the workflow.

Legal & Compliance

Sees: Risk and regulation.

Misses: Cannot see the productivity loss of saying no.

The CEO

Sees: Ambition and velocity.

Misses: Cannot see what the finance team is actually pasting into ChatGPT on a Tuesday.

The CoS sees across all three. The CoS is trusted by all three. The CoS already runs the operating rhythm that turns decisions into follow-through. What follows is not a new job for the CoS. It is the job, named.

Diagnosis

Which posture does your org need now?

Before any framework or instrument, the CoS answers a single diagnostic question: what kind of governance problem does this organization actually have right now? Most companies need all three postures at once, applied to different surfaces. The CoS's job is to know which belongs where — and to start with the posture that matches the most pressing reality.

01

The Velocity Posture

The org is under-using AI. Governance risks becoming a brake.

CoS move: Remove friction. Fund sandboxes. Publish sanctioned tool lists. Celebrate use cases publicly. Make the approved path fast enough that no one bothers with a workaround.

When to use it: Early-stage adoption, specific functions lagging, new acquisitions being integrated.

02

The Visibility Posture

Use is widespread but unmanaged. You are in the 98% shadow AI reality.

CoS move: Inventory without punishment. Declare amnesty. Build the ledger. Treat disclosure as learning, not enforcement. Pair every discovered shadow use with a paved-road alternative within 30 days.

When to use it: Most mid-size and scaling organizations in 2026 are here and do not know it.

03

The Veracity Posture

Autonomous agents are acting at scale. The risk is wrong actions, fast.

CoS move: Named Stewards per agent. Incident playbooks. Periodic human audits even of out-of-the-loop systems. Separation of concerns: the agent that creates is never the agent that verifies.

When to use it: Any surface with customer impact, regulatory exposure, or financial authority.

Posture tells you where to start. The next question is how — before you pick up any tool.

Phase 2 · How

The approach, before the instrument

One philosophy sits under every governance decision that follows. Get this wrong and the instrument will not save the program.

Approach

The Paved Road Principle

Every governance decision the CoS makes flows from one philosophy: build the well-lit path, do not build walls. Prohibition drives shadow use deeper. A good sanctioned option pulls it back into the light. This principle sits under every section that follows — the Register, the operating rhythm, the thirty-day starter. If the approach is wrong, the instrument will not save the program.

Borrowed from platform engineering

Build the well-lit path

Instead of building walls around forbidden behavior, build a well-lit, well-maintained path that is genuinely easier to follow than the alternative. Prohibition drives shadow use deeper; a good sanctioned option pulls it back into the light.

  1. 01

    Surface what people are actually doing

    Not what the AI policy says they should be doing. Interviews and declared-use amnesty beat network scans.

  2. 02

    Fund the paved version

    If marketing is pasting customer data into a consumer chatbot, the answer is not a memo; it is an enterprise-sanctioned tool with the same ease of use.

  3. 03

    Retire the cowpath

    Close off the shadow option only after the paved one exists, not before. The sequence matters.

With the approach settled, the instrument becomes straightforward. The Register is where the paved road gets drawn.

Phase 3 · What

The Register — and how to fill it

With approach settled, the instrument becomes straightforward. The Register is where the paved road gets drawn.

The Instrument

The Register

Every AI use case in the organization belongs on one ledger, with five columns the CoS is uniquely positioned to maintain. Reviewed quarterly. Owned by the CoS. Consumed by the CEO, CTO, and Legal in three different languages.

Column
What it captures
The CoS question behind it
Surface
Where the AI touches the business: customer data, board materials, financial modeling, hiring, code, vendor comms, internal comms.
"What part of how we actually operate does this touch?"
Sanction State
Approved, Sandboxed, Shadow, or Prohibited.
"Do we know this is happening? Do we want it to?"
Autonomy Position
Human-in-the-loop, Human-on-the-loop, or Human-out-of-the-loop (Deloitte 2026).
"Who is watching, and how closely?"
Blast Radius
Who is affected when it goes wrong: one person, one team, customers, regulators, or the brand.
"If this fails silently for a month, what breaks?"
Steward
Named human accountable for outcomes, not just uptime.
"Whose name is on this when the board asks?"

Autonomy Position · the three values defined

The autonomy spectrum that matters

Deloitte's 2026 framing is the clearest on the market. The CoS should adopt the language and stop letting teams conflate these three.

Autonomy

Human-in-the-loop

AI drafts, human reviews and approves before anything ships.

Examples: Board briefs · offer letters · customer-facing copy · financial models.

Default for anything with external audience or legal weight.

Autonomy

Human-on-the-loop

AI acts, human monitors and can intervene.

Examples: Auto-filed meeting notes · expense categorization · inbound lead qualification.

Human is not in the critical path but is watching the dashboard.

Autonomy

Human-out-of-the-loop

AI acts and the system audits.

Examples: Anomaly detection · log triage · scheduled report generation.

Continuous monitoring replaces case-by-case review.

Most organizations in 2026 are trying to operate at out-of-the-loop speed with in-the-loop assumptions, or the reverse. The Register forces the conversation.

A blank Register is not the hard part. Filling it honestly is. How the CoS does that changes with the size of the company.

Discovery

How to populate the Register by company stage

Discovery is a judgment problem, not a scanning problem

No network-monitoring tool will surface a finance analyst pasting an earnings draft into a consumer chatbot at 9pm. The data the Register needs comes from conversations, not dashboards. But the method scales differently depending on how many conversations there are to have.

Four stages below, each with the CoS-appropriate discovery method and the tooling that makes it tractable. Claude — or the AI thought partner of your choice — is part of the method itself, not a substitute for the conversations.

Stage 01 · Seed

The hallway method

Typically under 50 people · pre-Series A

Method

Direct conversation with every functional lead. At this stage the CoS can realistically know every AI use case in the company by name.

Goal: A complete Register in one week of 20-minute conversations.

Tooling: A shared doc. No platform spend required. The CoS's real work is building the muscle of disclosure before the company grows past the point where direct visibility is possible.

Stage 02 · Series A–B

The functional audit

Typically 50–250 people · post-product-market-fit

Method

Structured interviews with function leads (Eng, Product, Sales, Marketing, Finance, People, Ops), asking each to name their team's top three AI uses. Cross-reference with a short all-hands survey.

Goal: A Register that captures ~80% of real use in two to three weeks.

Tooling: A spreadsheet or Notion database, plus an AI thought partner to synthesize interview notes into Register rows. This is where Claude starts earning its place in the workflow.

Stage 03 · Series C+

The delegated discovery

Typically 250–1,000 people · scaling globally

Method

The CoS cannot personally interview every function any longer. Instead: appoint a Function AI Liaison per department, give each a simple intake template, and run a quarterly Register rollup. Pair with procurement data (any SaaS tool with 'AI,' 'Copilot,' 'GPT,' or 'Agent' in the name or latest release note) to catch embedded AI that users may not flag.

Goal: Liaison-led intake plus procurement cross-check.

Tooling: Shared Register in a governance platform or purpose-built Notion template; AI thought partner for rollup synthesis; procurement/SaaS management data as a cross-check.

Stage 04 · Public or PE-owned

The governance platform

Typically 1,000+ people · regulated or sponsor-scrutinized

Method

At this scale, the Register becomes a governed asset with auditable history, role-based access, and integration to existing risk and procurement systems. Discovery combines Liaison-led intake, automated AI-tool discovery (CASB, SaaS management platforms), and periodic third-party assessments.

Goal: The CoS's role shifts from discovery to orchestration across named deputies.

Tooling: Dedicated AI governance platform (Credo AI, Enzai, Holistic AI, or equivalent); integration with existing GRC stack; quarterly attestation from each Function AI Liaison.

AI-Assisted Discovery

Two prompts that make Claude a real discovery partner

At Stages 2 through 4, the CoS is synthesizing more interview notes than any human should process manually. A dedicated AI thought partner — set up as a Claude Project, ChatGPT custom GPT, or equivalent — turns raw notes into Register rows in minutes. Below are two prompts the CoS can use immediately. Adapt freely.

Prompt

You are acting as my Chief of Staff thought partner for AI governance. I am going to paste notes from an interview with [FUNCTION LEAD NAME, ROLE]. Extract every distinct AI use case mentioned or implied, and for each one produce a draft Register row with these five columns: Surface, Sanction State, Autonomy Position, Blast Radius, Steward. Where information is missing, flag it as "needs follow-up" rather than guessing. Keep each row under 40 words per cell. Here are the notes: [PASTE].

Prompt

Here is our current Register of AI use cases across the company [paste the table]. Act as a governance analyst and tell me: (1) which rows are most likely to expose us to the EU AI Act as high-risk, (2) which rows have a mismatched Autonomy Position given their Blast Radius, (3) which rows are missing a named Steward and should not stay that way. Rank each finding by how urgent it is to resolve this quarter, and explain your reasoning in one sentence per row.

These prompts assume the CoS has set up a dedicated Claude Project with organizational context loaded — org chart, function names, known AI vendors. Without that context, the quality of the synthesis drops materially. Resource 9 in this series covers the setup.

An instrument and a discovery method are still abstract until they meet a real company. The rest is applied.

Phase 4 · Apply

Working example, accountability, rhythm, starter

An instrument and a method are still abstract until they meet a real company. The rest of this resource is applied.

Worked Example

TrackFlow — the Register on a real company

Worked Example · TrackFlow

What the Register looks like on a real company

Returning to TrackFlow, the fictional B2B SaaS company from the P&L Field Guide. ARR sits at $19M with growth decelerating for eight consecutive quarters. Gross margin has dropped from 71% to 63%. Equity remaining is $2.1M, down from $11.4M. A funding conversation is coming, and it will not be the easy kind.

This is the environment in which AI governance stops being theoretical. Every use case below either helps the next board meeting go better, or makes it worse. The Register is how the CoS makes that distinction legible.

Surface

Board narrative drafting

Sanction

Approved

Autonomy

Human-in-the-loop

Blast Radius

Brand & investors

Steward

CoS

Surface

Sales forecast modeling

Sanction

Shadow

Autonomy

Human-on-the-loop

Blast Radius

Regulators (revenue recognition)

Steward

CRO (proposed)

Surface

Customer support triage

Sanction

Sandboxed

Autonomy

Human-out-of-the-loop

Blast Radius

Customers

Steward

VP Customer Success

Surface

Candidate screening in ATS

Sanction

Shadow

Autonomy

Human-on-the-loop

Blast Radius

Regulators (EU AI Act, high-risk)

Steward

Head of People (proposed)

Surface

Investor update copy

Sanction

Shadow

Autonomy

Human-in-the-loop

Blast Radius

Investors & brand

Steward

CEO (via CoS)

Reading the Register

What five rows tell the CoS in under a minute

Three of five rows are Shadow.

This is not a TrackFlow failing. It is the 2026 baseline. The Register's job is not to shame the company into existence; it is to name reality so the next decision is informed. At 82% equity collapse, the CoS does not have the option of pretending the finance team is not already using AI to stretch the runway.

The ATS row is the most dangerous on the page.

Candidate screening is explicitly classified as high-risk under the EU AI Act. Shadow use on this surface with a funding round approaching means a diligence finding waiting to happen. This row moves to Sandboxed this week, Approved within the quarter, or it disappears.

The sales forecast row is the most consequential for the next board meeting.

If revenue recognition is being influenced by an ungoverned model, the CFO needs to know before the auditors do. The CoS brings this to the CFO before the QOR, not during it.

The investor update row is the most quietly damaging.

Investor communications drafted by AI without a named Steward is how a well-meaning CEO ends up defending a claim they never actually made. The Register forces the name.

Accountability

The Steward model

Every AI use case on the Register has a named human Steward. The Steward is not the engineer who built it or the vendor who sells it. The Steward is the business owner of the outcome.

This is governance as role clarity, which is why it slots directly into the Decision Frameworks resource. In DACI terms, the Steward is the Approver. In RAPID terms, the Steward Decides. The engineer Recommends. The CTO is Consulted on technical risk. The CEO is Informed on strategic exposure. Legal is Consulted on regulatory exposure.

When an agent misfires, the question is not "whose system broke?" It is "whose judgment was this supposed to extend?" The Steward answers.

Operating Rhythm

The Quarterly Orchestration Review

An operating ritual the CoS runs alongside the QBR. Not a separate meeting — a standing section of the existing executive operating rhythm.

One page per Surface

Four questions

  1. 1.What is running today? (Sanctioned, Sandboxed, and known Shadow uses.)
  2. 2.What changed this quarter? (New uses, retired uses, autonomy shifts.)
  3. 3.What broke? (Near-misses count. Silence is a finding, not an absence of one.)
  4. 4.What is next? (Proposed changes to autonomy position, stewards, or sanction state.)

Read three ways

Classic CoS translation work

CEO

Strategic velocity and leverage

"Where is AI actually producing output we couldn't get before?"

CTO

Technical risk and architecture debt

"Where are we accumulating agent sprawl we'll regret?"

Legal / Compliance

Regulatory exposure

"What on this ledger would we struggle to defend to an EU AI Act auditor?"

The Soft Skills Are the Mechanism

The Disclosure Dividend

You cannot orchestrate what you cannot see

Why honesty is the governance instrument

The hardest part of the Register is not building it. It is filling it honestly.

Employees disclose AI use when disclosure is treated as learning. They hide it when disclosure is treated as enforcement. The governance literature is consistent: prohibition-first approaches produce less visibility and more risk, not less risk. A paved-road approach with transparent declaration reveals far more than any network scan.

This is where the soft skills of a Chief of Staff become the actual governance mechanism, not adjacent to it. Trust, discretion, follow-through, the ability to surface uncomfortable information without punishing the messenger — these are not nice-to-haves that make the Register easier to build. They are what make the Register possible at all.

You cannot orchestrate what you cannot see. You cannot see what people are afraid to show you.

Get Started

The thirty-day starter

If you are reading this as a sitting CoS without an AI governance program, here is the minimum viable version. It does not require budget or headcount. It requires a calendar and a willingness to ask.

01

Week 1

Declare amnesty

A short CEO-signed note: "We want to know what AI tools you are using. No tool will be banned in the next 30 days as a result of disclosure. We are building a better version." Open a simple intake form.

02

Week 2

Interview the heaviest users

The marketing ops person, the sales engineer, the FP&A analyst, one engineer, one recruiter. Ask what they use, why, and what a sanctioned version would need to do to be worth switching.

03

Week 3

Build the first Register

Five columns, one row per discovered use case. Assign a proposed Steward for each. Leave the Autonomy Position blank for now if you need to.

04

Week 4

Present to the Triad

CEO, CTO, and you. One decision per row: Approve, Sandbox, or Retire-with-replacement. Assign the replacement work. Schedule the first QOR.

Why no agent replaces this role

The AI-Pilled CoS resource made the case that no AI agent can replace a human Chief of Staff. The Orchestration Ledger is the sharpest proof.

An agent cannot hold an amnesty conversation with a nervous team lead. An agent cannot read the room when the CTO and Legal disagree on autonomy position. An agent cannot decide that this quarter, the org needs velocity more than visibility, or the reverse. An agent cannot be the named Steward — by definition.

The governance of AI is a deeply human problem dressed in technical clothing. That is exactly the kind of problem a Chief of Staff is built to hold.

Key research cited

  • Deloitte 2026 State of AI in the Enterprise — 74% planning agentic AI, 21% with mature governance; in/on/out-of-the-loop framing.
  • Deloitte 2025 Tech Value Survey — 80% automation maturity vs 28% agent maturity across 550 US leaders.
  • CrowdStrike 2026 Global Threat Report — 98% of organizations report unsanctioned AI use.
  • IBM 2025 Cost of a Data Breach Report — shadow AI adds avg $670,000 to breach costs; 63% of breached orgs had no AI governance policy.
  • McKinsey State of AI 2025 — 88% of organizations use AI in at least one business function; redesigning workflows is the top success factor.
  • Gartner / Forrester 2026 — >40% of agentic AI initiatives at risk of abandonment by 2027 due to governance and ROI fundamentals.
  • EU AI Act (Regulation 2024/1689) — enforcement begins 2 August 2026.