GOVERNANCE
The question is not whether the model works.It is what may reach production, and how you prove it.
Inside a supervised institution the binding constraint is never capability. It is evidence. A system that cannot show what it saw, what it decided and who could stop it does not reach a customer, however well it performs in a demo.
This section exists to settle one objection and return you to where you were. It is a destination, not a step. What follows is the perimeter an institution can actually defend, and the conditions that widen it.
The model behind this section draws on recurring requirements across the NIST AI Risk Management Framework and its Generative AI Profile, ISO/IEC 42001:2023 and the European Union AI Act. It is an executive diagnostic model — not a certification, a legal opinion, a regulatory determination, or a substitute for a use-case-specific risk assessment.
Governance is not the brake. It is the function that decides what is allowed to run — and most institutions cannot answer where they stand, because they measure activity and present it as capability.
THE FLOOR
A high average authorizes nothing.
Every maturity scorecard in the market averages its domains. Averaging is exactly the wrong operation. An institution with excellent strategy, excellent data and no human-intervention policy does not have a governance problem in one domain — it has no perimeter at all, because the weakest control is the one an incident will find.
CRITICAL FLOOR
critical floor = minimum(risk, data, lifecycle, human intervention, monitoring)
A strong average cannot compensate for a severe weakness in a control domain required for consequential AI.
OPERATING PERIMETER
permitted operating mode = f(overall maturity, critical floor, monitoring readiness)
Maturity is not authorization. Score and control-floor thresholds are applied to progressively more consequential operating modes.
WHAT DECIDES THE MODE
Control intensity should follow consequence.
Autonomy
Can the system only suggest, or can it decide and execute?
More autonomy requires stronger permissions, monitoring and intervention.
Materiality
What is the financial, customer, legal or operational consequence of an error?
Higher materiality requires stronger evidence, approval and accountability.
Five operating modes. Each one has a gate.
| Operating mode | What the system does | What the person does | Gate |
|---|---|---|---|
| AssistGrounding · data access · quality review | Search, summarize, classify, draft and prepare information. | Remains the decision maker and the executor. | maturity ≥ 25 · floor ≥ 25 |
| RecommendEvidence · explainability · threshold monitoring | Propose next actions or decision options. | Accepts, rejects or changes the recommendation. | maturity ≥ 45 · floor ≥ 37.5 |
| Bounded decisionDecision rights · constraints · exceptions | Select actions within explicit policy, materiality and exception limits. | Designs the policy, reviews exceptions and remains accountable. | maturity ≥ 65 · floor ≥ 62.5 |
| Bounded executionIdentity · permissions · limits · rollback | Invoke systems or complete authorized actions under monitoring. | Supervises material exceptions, and can interrupt or reverse execution. | maturity ≥ 75 · floor ≥ 75 · monitoring ≥ 75 |
| High-materiality autonomyMonitoring · audit trail · kill switch · incidents | Requires a specific risk and regulatory assessment for the use case. | Decides case by case, with risk, legal and the business at the table. | never authorized by enterprise score alone |
These weights and thresholds belong to this executive model. They are not regulatory thresholds or industry standards, and a self-declared score without documentation is a hypothesis, not proof that a control is effective.
Maturity is a constraint system, not a leaderboard.
Two institutions can score the same and hold different perimeters. The aggregate is a conversation. The floor is the decision.
RUN IT AGAINST YOUR ORGANIZATION
Both formulas, computed on your institution.
Instrument · 28 questionsTwenty-eight questions answered against current evidence, not future plans. It returns maturity by domain, the weakest control floor, and the operating perimeter that floor supports. A high score is not the objective — the objective is to know what level of consequence the organization can carry without pretending that missing controls do not matter.
Open the assessment →THE CONTRACT
Six conditions before an agent touches a customer.
A control that only reviews the model before launch is not enough once the system can call tools, hold state and execute actions. These six are the ones a CIO should be able to read out loud in a risk committee — and if any one of them has no answer, that is the answer.
- 01
Named owner
One accountable business owner for the workflow, its economics and the outcomes it is permitted to produce.
- 02
Explicit permissions
Tool access, data scopes, transaction limits and action rights are defined — and revocable.
- 03
Human intervention policy
Clear triggers for approval, escalation, supervision, override and forced stop.
- 04
Continuous evidence
Traceability must show what the agent saw, decided, called, changed and escalated.
- 05
Quality and failure thresholds
Completion, escalation, rework, error, policy breach and adverse-outcome indicators, with a level at which the agent stops.
- 06
Release and rollback discipline
Versioning, simulation, controlled rollout, canaries, restrictions and rollback are part of normal operation, not an incident response.
These are not aspirations. They are the contract the agents in the next section are built against, and the reason their coverage matrix is published rather than described.
See what runs →MEASUREMENT
Almost everyone measures level 1 and reports level 4.
Companies love reporting movement. Movement is not capability. An institution that cannot place itself on this ladder cannot defend a perimeter either, because the evidence it collects proves that something was deployed — not that anything was absorbed.
Activity
- Users
- Prompts
- Bots
- APIs
- Pilots
Local efficiency
- Time per task
- Cost per process
- Error reduction
Operating capacity
- Customers per employee
- Products per employee
- Throughput per team
Business impact
- Growth without proportional cost
- Margin improvement
- Operating leverage
The ladder is a measurement hierarchy, not a maturity score. A level-4 number is only credible when levels 2 and 3 are instrumented underneath it.
False signals of automation
Every one of these is a level-1 number wearing a level-4 headline. None of them proves automation, scale or leverage.
Most companies stay at level 1 and present it as if they had reached level 4.
SO WHAT DOES LEVEL 3 LOOK LIKE
The same substitution, industry by industry.
| Industry | Surface indicator | Serious indicator | Economic derivative |
|---|---|---|---|
| Banking | Bots, apps, copilots | Customers per employee, products per employee | Real cost per customer, real cost per product |
| Insurance | Quoting tools, CRM, RPA | Policies per employee, customers per employee | Real cost per policy, real cost per customer |
| Pensions | Campaigns, digital channels | Members per employee, cases per adviser | Real cost per member, real cost per case |
| Payments | APIs, new rails, integrations | Transactions per employee, active merchants per employee | Real cost per transaction, real cost per active merchant |
| Health | Scheduling, chatbots | Patients or episodes per employee | Real cost per patient, real cost per episode |
Every serious automation indicator ends up back at capacity per employee and cost per unit served.
THE EXITS
Six pieces. Each one decides something different.
Listed by what they settle, not by what they discuss. Four are arguments; two are instruments you run against your own organization, or against yourself.
Governance for AI in regulated environments
The operating model: which role answers for each lifecycle decision, and what control accompanies the system once it is running rather than before it launches.
What it decides
Who answers.
Open →AI governance maturity assessment
Twenty-eight questions answered against current evidence, not future plans. Returns maturity by domain, the weakest control floor and the operating perimeter it supports.
Instrument · 28 questionsWhat it decides
Where you stand.
Open →Measuring generative AI productivity
The measurement ladder in full, the false signals that replace it, and the arithmetic that separates a level-2 saving from level-4 leverage.
What it decides
What counts as a result.
Open →AI agents for financial services in Latin America
Seven layers — channels, identity and delegation, orchestration, policy controls, knowledge and tools, human intervention, observability — with the risks each one carries.
What it decides
How it is built.
Open →Six executive profiles in front of generative AI
Four questions (understanding, use, trust and verification) and the six profiles they produce. A table that locates your profile, its main risk and the next step.
Instrument · 4 questionsWhat it decides
Who is deciding with it.
Open →The deterministic engine that questions the probabilistic one
Why a regulated institution cannot demand consistency from a large language model (LLM), what the examiner actually asks for, and the control layer that turns an impossible requirement into one that can be met.
What it decides
What validates the model.
Open →WHAT THIS IS ANCHORED ON
An original executive model, built on recurring requirements.
It reproduces no single standard. Its domains reflect requirements that recur across the frameworks most relevant to AI risk and management.
- NIST AI Risk Management Framework + Generative AI Profile
Lifecycle risk management, governance, measurement and context-specific profiles for generative AI.
- ISO/IEC 42001:2023 — AI management systems
AI governance as a management system: leadership, risk, data, lifecycle controls, monitoring and continual improvement.
- European Union AI Act
For high-risk AI: risk management, data quality, documentation, logging, human oversight, robustness and post-market monitoring.
Model boundaries
- The weights and operating-perimeter thresholds belong to this executive model. They are not regulatory thresholds or industry standards.
- The assessment measures institutional capability at enterprise level. A specific system may require stricter controls by product, customer, jurisdiction, materiality or legal requirement.
- A self-declared score without documentation is a hypothesis, not proof that a control is effective.