AI Governance Playbook: 5 Layers From Policy to Assurance

In brief

In my experience, most organisations have written AI principles and few can show one that changed a decision. This playbook turns principles into rules that do, in five layers. Policy turns each principle into a testable rule, Process puts checks into the workflow teams already use, and Tooling automates those checks. People names who decides, by system and by decision. Assurance tests whether the other four work and feeds what it finds back into policy. They come after Minimum Viable Governance (MVG), the 90-day Assess stage, whose documents become their first entries. Between them the layers map all 72 NIST AI RMF subcategories. Governance effort follows MVG's three risk tiers, which sit alongside legal categories such as those in the EU AI Act. Three documented failures show what is missing when principles have no process behind them, and the page ends with five common ways governance fails and the fix for each.

Key Takeaways

→A principle governs only when it changes a decision. That takes a testable rule, a check in the workflow and a named person who acts on the result.
→The Five-Layer Governance Stack runs Policy, Process, Tooling, People and Assurance. Each layer builds on the ones before it, and owners are named first: every system has one from day one.
→Governance effort should follow risk. The three tiers from Minimum Viable Governance set who approves a system and how often it is reviewed: monthly for high, quarterly for medium, yearly for low. They sit alongside the EU AI Act categories and do not replace them.
→The layers are the Align and Assure stages of one model. Minimum Viable Governance is the Assess stage, days 0 to 90, and its documents become the first entries in the layers.
→Between them, the five layers map all 72 subcategories of the NIST AI Risk Management Framework: the 12 that MVG starts and 60 more. A mapped subcategory is a place to do the work, and a programme that follows the layers can show which ones it has not reached yet.

One model, three stages: Assess, days 0 to 90 → Align, months 3 to 6 → Assure, ongoing. You are here: Align and Assure.

An AI governance playbook turns principles into rules that change what a team ships. This one has five layers, running from Policy, which says what you will enforce, through Process, Tooling and People to Assurance, which tests whether the other four work. It covers the Align and Assure stages, after the 90 days of Minimum Viable Governance, and is written for an organisation whose governance has to hold as the number of systems grows.

Why Don’t AI Principles Change Decisions?

I ask AI and risk leaders one question: show me the last time your AI principles changed a decision. Most of the time the room goes quiet. The principles are usually well written. They sit in a document while decisions get made elsewhere, and three things keep them apart.

Committees write principles; governance needs an owner. A committee can agree that AI should be fair. Someone still has to pick a fairness measure for each use, set its threshold and act when a system misses it.
Principles are broad; governance needs specifics. "We will not discriminate" is a principle. "No model that screens job applicants goes live unless its error rates for each protected group sit within a gap we have set, tested on current data" is a rule. Only the second can block a release.
Principles are written once, and the rules around them keep moving. Regulation (EU) 2026/1744 moved the start of EU AI Act obligations for the high-risk systems listed in Annex III from 2 August 2026 to 2 December 2027. A programme that never reviews itself is soon planning against the wrong dates.

What Happens When Principles Have No Process?

Three well-documented cases each show a missing layer. None of them is known to have come from bad intent.

A hiring model that learned the past. Reuters reported in 2018 that Amazon had trained a model to rate job applicants on ten years of résumés, most of them from men. It learned to penalise résumés that included the word "women’s" and downgraded graduates of two all-women’s colleges. Amazon dropped the project. It said its recruiters never used the tool to evaluate candidates, and did not dispute that they had looked at its recommendations. Only a test that looks for this bias will find it, so Policy has to define "fair" before a model is built.
A health model that predicted the wrong thing. A 2019 study in Science examined a commercial algorithm of a kind that, by industry estimates, is applied to about 200 million people in the US each year. It used health costs as a stand-in for health needs, and less had been spent on Black patients with the same needs, so at the same risk score they were considerably sicker. Fixing the gap would raise the share of Black patients picked for extra help from 17.7% to 46.5%. A policy has to cover what a model may predict, as well as how accurately.
A pricing model the market outran. Zillow bought and resold homes through Zillow Offers, its iBuying business. In November 2021 it said it would wind the business down and cut about a quarter of its staff. Its chief executive, Rich Barton, said "the unpredictability in forecasting home prices far exceeds what we anticipated". A pause threshold watched in production is the Assurance layer’s answer to a market that outruns its model.

What Are the Five Layers?

Policy. A policy says what you will enforce, with a measure, a threshold and a scope. If two reviewers could disagree on whether a system complies, it's still a principle. In my experience, most organisations stall here, because the translation needs people who know both the domain and the model.
Process. Put the checks inside the workflow teams already use: design review, pre-production and release. A parallel governance track gets skipped when a deadline arrives. A check that travels with the work does not.
Tooling. Automate the checks once they exist: fairness tests in the build pipeline with open-source kits such as Fairlearn or AI Fairness 360, model documentation generated from training records, drift monitoring with alerts, and a log of every override and exception. Tools automate a policy but cannot write one, so buy them after the first two layers.
People. One named person is accountable for each system and each type of decision. Everyone in the chain is trained for their part, from the data scientist running a bias test to the executive reviewing an escalation.
Assurance. Check that the other four work. Every quarter, review the gates: what passed, what was escalated, what got a waiver and how the waived systems behaved afterwards. Bring in an outside auditor for the most critical high-tier systems, and feed audit findings, incidents and rule changes back into Policy.

The order is a dependency, not a calendar: Tooling automates what Policy defines and Process enforces, and Assurance checks all four. Owners still come first, because every system needs one from day one. The People layer later adds decision rights and training.

How Do Align and Assure Build the Layers?

The five layers come after the Assess stage, Minimum Viable Governance (MVG), whose 90 days give every AI system a place in an inventory, a risk tier and a named owner who can pause it.

Align, months 3 to 6. Build each MVG document out into the layer it opens, as the table below shows. Then work through the layers in order, adding the NIST subcategories in its last column.
Assure, ongoing. Each system comes back for review on its tier’s cadence, starting from the review date it got at the MVG deployment gate, so the first reviews run during Align.

Which NIST Subcategories Does Each Layer Cover?

MVG’s documents answer 12 of the 72 subcategories in the NIST AI Risk Management Framework. Align and Assure map the other 60. The table lists each one once, under the layer that does most of the work, though many touch more than one.

Layer	Starts with, from MVG	Added in Align and Assure
Policy	AI system inventory (GOVERN 1.6); governance charter (GOVERN 1.4)	Legal requirements, trustworthy-AI policy, written rules for how much risk management each tier needs, and third-party policy (GOVERN 1.1, 1.2, 1.3, 6.1); mission, risk tolerance, system requirements and scope (MAP 1.3, 1.5, 1.6, 3.3)
Process	Context and impact record (MAP 1.1); risk register (MAP 5.1, MANAGE 1.2); deployment gate (MANAGE 1.1); escalation and incident procedure (MANAGE 4.1, 4.3)	Decommissioning, risk records kept during delivery and third-party contingency plans (GOVERN 1.7, 4.2, 6.2); business value, task definition, benefits, costs and component risk (MAP 1.4, 2.1, 3.1, 3.2, 4.1, 4.2); human-subject evaluations and user appeals (MEASURE 2.2, 3.3); risk responses, residual risk, resources and non-AI alternatives, unknown risks and third-party monitoring (MANAGE 1.3, 1.4, 2.1, 2.3, 3.1)
Tooling	Baselines and thresholds (MEASURE 1.1); production monitoring (MEASURE 2.4)	Testing and incident practices (GOVERN 4.3); system limits and test design (MAP 2.2, 2.3); test records, performance in deployment-like conditions, the validity, safety, security, explainability, privacy, fairness and environmental tests, and risk tracking (MEASURE 2.1, 2.3, 2.5 to 2.7, 2.9 to 2.12, 3.1); keeping deployed systems useful and monitoring pre-trained models (MANAGE 2.2, 3.2)
People	Accountability matrix (GOVERN 2.1); authority to pause (MANAGE 2.4)	Training, executive responsibility, diverse teams, human-AI oversight roles and a safety-first culture (GOVERN 2.2, 2.3, 3.1, 3.2, 4.1); interdisciplinary context teams, operator proficiency and human oversight processes (MAP 1.2, 3.4, 3.5)
Assurance	The next review date, set at the deployment gate (no MVG subcategory)	Periodic review and outside feedback (GOVERN 1.5, 5.1, 5.2); engagement on impacts (MAP 5.2); reassessing metrics and controls, independent assessors, transparency and accountability risks, test effectiveness, hard-to-measure risks and field validation (MEASURE 1.2, 1.3, 2.8, 2.13, 3.2, 4.1 to 4.3); continual improvement (MANAGE 4.2)

How Does a Principle Become a Policy?

Follow the most common principle down the stack. By the last layer nothing is left to interpret: the system passed its fairness audit or it didn't. Parity tests miss a model that predicts the wrong thing, so Policy checks the target too. The thresholds below are examples.

The principle: "We are committed to fairness in AI."

Layer	What the principle becomes
Policy	Every classification model that makes decisions about people shows similar true and false positive rates across legally protected groups, within a gap the policy sets (for example, five percentage points).
Process	A fairness review at the pre-production gate. A waiver needs written approval from the tier’s approver, with a reason and an expiry date.
Tooling	The fairness test runs in the build pipeline. A model that fails it cannot be deployed.
People	The model owner, by name, is accountable for the result. The governance lead reviews the whole portfolio each quarter.
Assurance	Fairness metrics are monitored for drift every month, and high-tier systems get an independent fairness audit each year.

How Much Governance Does Each System Need?

A content suggestion tool and a credit model carry different risks. Use the tier each system got in Minimum Viable Governance. It sets who approves the system and how often it comes back for review.

High: the system affects a person’s rights, safety, money or job, as credit, hiring, pricing and eligibility decisions do. A VP or named executive approves it, and it is reviewed monthly. The most critical, such as clinical, justice and safety systems, also get an external audit.
Medium: the system shapes business decisions or customer experience, as forecasting, process automation and customer chatbots do. The AI lead approves it, and it is reviewed quarterly.
Low: an internal tool with no consequential output. The team self-certifies, and it is reviewed yearly.

These internal tiers are not the EU AI Act’s legal categories, and the two can differ. A medium-tier chatbot that answers product questions must still be built so people know they are dealing with an AI, unless that is obvious (Article 50). Hiring and credit-scoring systems listed in Annex III are high-risk, with obligations from 2 December 2027 (Article 113, as amended). Practices the Act prohibits, such as social scoring (Article 5), are not a tier at all: no approval makes them lawful in the EU.

Who Owns Each Decision?

Give every decision type one accountable person. A simple split:

The model owner is accountable for one system’s day-to-day compliance.
The AI governance lead is accountable for the portfolio: tiers, gates and the quarterly review.
Legal and compliance are consulted on what the law requires.
The executive sponsor is informed of material risks and every escalation.

Article 26 of the EU AI Act points the same way: deployers of high-risk systems must assign human oversight to people with the competence, training and authority to do it, and the support they need. Two people accountable for one decision usually means nobody is.

How Do You Know It Is Working?

Governance that cannot show its value loses its budget. Track four measures:

Prevention: problem deployments caught before production, and at which gate.
Speed: time from a finished model to a governed deployment, and the share of reviews done on time.
Coverage: the share of AI systems with complete records, and how many run outside governance.
Trust: complaints and appeals from people your systems decide about, and what staff, customers and regulators say.

Speed keeps governance from becoming the bottleneck people route around. If the programme has not stopped or changed a single deployment in a year, either it isn't working or nothing needed stopping. In my experience it is usually the first.

What Are the Common Ways Governance Fails?

For the next leadership meeting: list every AI system in production, give each a tier, and name the high-tier systems that have no owner and no gate. Start with those.

The Governance Playbook worksheet has the policy translation table, the gate map, the tier inventory, the accountability matrix and the failure-pattern check on printable pages.

Subscriber Resource

Download: Governance Playbook Worksheet

Get the complete Five-Layer Governance Stack worksheet: principle-to-policy translation templates, governance gate map, risk-tiering inventory, RACI accountability matrix, KPI dashboard, failure-pattern check and the 12-week planner, ready to print or save as PDF.

Enter your email to get instant access. You'll also get The Brief, every other Tuesday.

Free. No spam. Unsubscribe anytime.

Sources and Changes

NIST AI 100-1, AI Risk Management Framework (AI RMF 1.0), January 2023, Tables 1 to 4.
EU AI Act Article 5, Article 26, Article 50, Article 113 and Annex III, consolidated text at 27 July 2026 on the Commission’s AI Act Service Desk; Regulation (EU) 2026/1744, OJ L, 24 July 2026.
Jeffrey Dastin, "Amazon scraps secret AI recruiting tool that showed bias against women", Reuters, 10 October 2018.
Ziad Obermeyer and colleagues, "Dissecting racial bias in an algorithm used to manage the health of populations", Science 366, 447 to 453, 25 October 2019.
Zillow Group, third quarter 2021 results, press release filed with the SEC on Form 8-K, 2 November 2021.

Changes on 7 October 2026. Wording was tightened throughout. The layers, tiers, NIST table and cases are unchanged.

Changes on 4 October 2026. Cut to a short reference. Survey statistics and the Apple Card, COMPAS and Zillow loss figures were removed. The risk tiers no longer claim to match EU AI Act categories, and social scoring is no longer a tier, because the Act prohibits it. It became the Align and Assure stages of one model with Minimum Viable Governance and gained the NIST table.