AI governance is the arrangement that makes an AI system accountable to somebody after it is running, and compliance is the evidence that the arrangement meets a standard somebody else set. The standards come from three places, the international standards bodies, the accounting profession's audit tradition and a growing body of law, and they converged on AI between 2023 and 2025 as the first management system standard for AI, the first city ordinance requiring bias audits and the first comprehensive AI law all arrived within two years of each other. The AWS syllabus splits the subject into the standards a system is measured against, the data practices underneath and the routine that keeps both alive, and the Practitioner course cares because governance is measured by how quickly the record answers.
The sections below take the standards, data governance once a model is involved, the governance routine, the scoping matrix the syllabus names, the four kinds of tool whichever platform is in use, and end with a five question test against the clock.
The standards a system is measured against
ISO publishes the management system standards. ISO/IEC 42001, published in December 2023, is the one for artificial intelligence, describing a management system for developing and using AI responsibly, and it sits alongside ISO/IEC 27001 for information security. Certification against it is how an organisation demonstrates a governance system exists, where an assertion would otherwise have to do.
SOC reports, System and Organization Controls, are audits of a service organisation's controls, defined by the American accounting profession. SOC 2 is the familiar one for security, availability, processing integrity, confidentiality and privacy, and it is what a customer's procurement team asks for.
Algorithm accountability laws are the growing category of regulation requiring disclosure, impact assessment or a right to explanation where automated decisions affect people. New York City's law requiring an annual bias audit of automated hiring tools has been enforced since July 2023, the EU AI Act's high risk regime arrives in December 2027, and a Colorado law covering consequential decisions takes effect in 2026. The specifics vary by jurisdiction and the direction is consistent, which is that a system deciding something about a person is expected to be documented, testable and appealable.
Neither vendor exam tests the text of any regulation. Both test whether a candidate knows which kind of instrument a given name is.
Data governance once a model is involved
Underneath the standards sits the data, and four practices govern it.
Data lifecycle covers a record from collection to deletion, and the useful question is who authorised each stage. Data collected for one purpose and used to train a model is a different purpose, and consent frequently does not travel.
Retention is how long the data is kept. Longer retention is a larger breach and a larger obligation, and training data has a habit of outliving the reason it was gathered.
Residency is where it physically lives. For AI this reaches further than it looks, because inference has to happen in the permitted region too, and many models are unavailable in some regions.
Logging, monitoring and observation are what make an incident reconstructable. For a generative system that means recording prompts, retrieved context and responses, which immediately becomes its own privacy problem, since the log now contains whatever the user typed.
The governance routine
Standards and data practices are kept alive by a routine with five parts.
Policies state what the organisation permits, and the useful ones are specific. Which categories of decision may be automated, which require a person, and what data may reach a third party model.
Review cadence is how often a deployed system is looked at again. Bias drifts, data shifts, and providers update models underneath the product, so a system signed off once is ungoverned.
Review strategies cover what a review consists of. Performance by subgroup, sampled human evaluation, incidents since the last review, and whether the documented limitations still match reality.
Transparency standards decide what gets published. Model cards, disclosure that a user is talking to a system, and stated limitations.
Team training appears in the syllabus and is easy to overlook. People deploying and operating these systems need to understand the failure modes, and an untrained team is the most reliable governance gap there is.
The Generative AI Security Scoping Matrix
The AWS exam guide names one framework explicitly. The Generative AI Security Scoping Matrix is an AWS framework from 2023 that places a generative AI use case in one of five scopes, running from consumer applications an organisation merely uses, through enterprise applications bought with AI inside them, building on a pre trained model, fine tuning one, to training a model from scratch. Security obligations differ sharply along that line, and the matrix exists so a team can work out which ones are theirs.
The four kinds of tool, whichever platform is in use
Governance tooling divides into four jobs, and every platform has a product for each. Holding the job transfers between platforms where a product name does not.
Configuration compliance, evaluating the team's own resources against rules. AWS Config, Google Security Command Center posture management, Azure Policy.
Evidence collection, gathering proof continuously against a framework so an audit is no scramble. AWS Audit Manager, Azure Compliance Manager, and Google Assured Workloads for regulated environments.
The provider's own compliance posture, meaning their SOC and ISO reports, which are separate from the team's. AWS Artifact, Google Compliance Reports Manager, Azure Service Trust Portal.
The activity log, recording who called what and when, which is what makes an incident reconstructable. AWS CloudTrail, Google Cloud Audit Logs, Azure Monitor activity logs.
The pattern worth carrying anywhere is that the first two concern the team's configuration, the third concerns the provider's, and the fourth is the record of what actually happened.
Practise this
The exercise needs one model or feature that is serving real traffic, a stopwatch, and permission to interrupt colleagues.
Start the clock and answer five questions, writing the elapsed minutes beside each as the answer arrives. Count the minutes and leave the effort uncounted, and mark where each answer came from honestly, because something a colleague remembered is a different artefact from something that could be read.
Model or feature .................................
Clock started at ......
Q1 Which exact version is serving traffic now? ___ min
Q2 Who approved it going live, by name? ___ min
Q3 What was it evaluated against before that
approval? ___ min
Q4 On what date, and has it been rerun since? ___ min
Q5 Where is each of the answers above written
down? ___ min
Mark every answer
W written down somewhere I can link to
R somebody remembered it
X could not answer at all
Total elapsed ______ minutes
The elapsed time is the finding. Governance that works produces all five within a couple of minutes from one place, and the usual result is a scavenge across a deployment tool, a chat thread, a spreadsheet and somebody's memory, which is the same search an auditor, a customer's security questionnaire or an incident at nine in the evening would demand at a far worse moment. Every answer marked R is one resignation away from becoming an answer marked X.
The exercise teaches that governance is measured by how quickly the record answers, and whether a policy document exists has little to do with it.