Contents 7 sections
Definitions of AI governance tend to describe the goal. The US National Institute of Standards and Technology (NIST) says it developed its AI Risk Management Framework to better manage "risks to individuals, organizations, and society" associated with AI. A second definition, often quoted under the name of the IT audit body ISACA, describes AI governance as the framework of policies, regulations, ethical principles and guidelines that govern AI systems from development through to use. That wording comes from a guest article on the ISACA Now blog, published on 5 December 2024 and written by an engineer at Dell Technologies. It is one author's wording hosted by ISACA, and it is best attributed that way.
A definition does not tell a CIO or a CISO what to set up. This article divides a governance programme into seven working parts, and for each one it names an owner and the evidence an auditor or regulator would ask to see.
What NIST and ISACA give you to start from
The NIST framework was released on 26 January 2023 and is intended for voluntary use. Its core has four functions named Govern, Map, Measure and Manage, and the framework text describes Govern as a cross-cutting function that enables the others. NIST's page also says version 1.0 is being revised as part of the White House AI Action Plan. The page lists no revised version, so 1.0 and the Generative AI Profile of 26 July 2024 remain the documents to work from.
ISACA's own material is more practical than the borrowed definition. Its toolkit Governing AI Across its Lifecycle, dated 15 April 2026, follows an AI system through five stages, from strategy and governance design to change and retirement. It names unclear ownership after deployment and a lack of structured monitoring as common pitfalls. Both sources treat governance as work that continues for as long as a system runs.
The seven parts, with owner and evidence
The owners in this table are our suggestions for a mid-size or large company, and they will differ by sector. The evidence column lists records that each part should produce in normal operation.
| Part | Suggested owner | Evidence it leaves |
|---|---|---|
| Roles and responsibilities | One executive named by the chief executive | Written role assignments and committee minutes that record decisions |
| Operating model | A small governance office under that executive | Intake records with a risk tier, and the approval at each stage |
| Risk and compliance | Risk and legal functions | Risk assessments, signed risk acceptances, a register of laws by system, review dates |
| Policies and standards | Governance office, approved by the committee | Dated policy versions and a log of exceptions |
| Model governance | The business owner of each system, with IT or data science | Inventory entry, test results, change history, retirement record |
| Tools | IT and security | Audit logs of AI activity and exported policy settings |
| Monitoring | The system owner, with security operations | Monitoring reports, an incident log, a timed shutdown test, follow-up actions |
The seven parts in detail
Roles and responsibilities
This part answers who decides, and who is accountable when an AI system causes harm. The NIST framework asks for documented roles and lines of communication (Govern 2.1) and says executive leadership takes responsibility for decisions about AI risk (Govern 2.3), both in its table of Govern outcomes. The US Office of Management and Budget's memorandum M-25-21 of 3 April 2025 told each agency to designate a Chief AI Officer, who may be an existing technology or data chief with significant AI expertise. The largest agencies also had to convene a governance board with members from IT, cybersecurity, legal counsel and privacy, among others.
Companies place the job in different functions. In the IAPP's AI Governance Profession Report, published on 16 April 2025 from a survey fielded in spring 2024, privacy teams held primary responsibility in 22% of organisations and legal and compliance teams in another 22%. IT followed at 17%. Any of these can work, provided one named executive holds the job and has the authority to stop a deployment.
Operating model
The operating model is the route a proposed use of AI takes from request to retirement, with the points at which someone must approve it. The five stages in ISACA's lifecycle toolkit are a usable template. We suggest sorting each request into a risk tier at intake, so that a meeting summariser does not queue behind a system that screens job applicants. The federal memorandum permits agencies to rely on existing governance bodies, and a company can extend its change board or data governance council in the same way.
Risk and compliance
This part identifies what could go wrong with a given system and which laws apply to it. In the NIST framework, the Map 1.1 outcome asks that the intended purposes of a system and the laws specific to its context are understood and documented. Manage 1.3 lists the responses to a high-priority risk as mitigating, transferring, avoiding or accepting. For high-impact uses, M-25-21 requires an impact assessment whose last item is risk acceptance supported by the signature of the person accepting the risk.
Policies and standards
A policy states what staff may and may not do with AI, and standards say how to do the permitted things. In ISACA's AI Pulse Poll, released on 5 May 2026 with more than 3,400 respondents, 38% said their organisation had a formal, comprehensive AI policy, up from 28% in 2025. Schellman, an audit firm, reported in a summary of its 2026 governance survey, published on 29 July 2026, that 57% of more than 500 US enterprise leaders had a formal AI governance policy. The two surveys asked different groups different questions, so the figures should not be compared directly.
For structure, ISO/IEC 42001, published on 18 December 2023, specifies requirements for an AI management system and is written for any organisation that provides or uses AI systems. The evidence for this part is a dated, approved policy and a log of the exceptions granted, with training records kept alongside.
Model governance
Model governance keeps track of each model and AI system from approval to retirement, and it starts with an inventory. NIST's Govern 1.6 outcome asks for mechanisms to inventory AI systems, and Govern 1.7 asks for processes to decommission them safely. An inventory entry should name the owner, the purpose, the data used and the model version, with test results and the approval attached. The guide page on agent registries covers the same record for agents.
The Federal Reserve's SR 26-2 of 17 April 2026 supersedes SR 11-7, the 2011 guidance on model risk management. The attached guidance says of generative and agentic AI models that "they are not within the scope of this guidance". It adds that a bank's own risk management and governance practices should guide the controls for anything the document does not cover. A bank therefore cannot rely on the model risk letter alone as its control framework for a chatbot or an agent.
Tools
Tools record what happens and enforce the rules that the other parts have set. In a Microsoft 365 company, Microsoft's Purview documentation, dated 27 May 2026, says prompts and responses from supported AI apps are captured in the unified audit log. Coverage is the limit. The page points to separate pages for the capabilities each app supports, so check every system in your inventory against what the product records for it.
Monitoring
Monitoring checks that a system in production still behaves as it did when approved, and that someone can stop it. NIST's Measure 2.4 outcome asks that system behaviour is monitored in production. Manage 4.1 asks for post-deployment plans that cover user input, override, incident response and decommissioning. Stopping a system is where the survey data looks weakest. In the ISACA poll of May 2026, 56% of respondents did not know how long it would take to halt an AI system after a security incident. In the same poll, 39% did not know whether a documented process for shutting down or overriding AI systems existed.
Schellman's 2026 survey found that 44% of respondents had documented incident response procedures for AI-specific issues. The evidence for this part is a set of scheduled monitoring reports and an incident log. We would add a test of the shutdown procedure, with the time it took.
The law dates that changed in 2026
Several dates that appeared in 2024 and 2025 guidance have since moved. This section is general information and not legal advice. In the European Union, Regulation (EU) 2026/1744, published on 24 July 2026, amended the AI Act. Duties for high-risk systems listed in Annex III now apply from 2 December 2027. Those for AI built into Annex I products such as lifts or toys apply from 2 August 2028, as the Commission's AI Act page, updated on 19 October 2026, also shows. The transparency rules in Article 50 kept their date of 2 August 2026. Providers of generative systems placed on the market before that date have until 2 December 2026 to meet the marking duty in Article 50(2). The amended Article 4 asks providers and deployers to take measures that support staff AI literacy, and says they need not guarantee a specific level for any individual.
In the United States, the Texas Responsible Artificial Intelligence Governance Act took effect on 1 January 2026. One of its defences covers a company that finds a violation through its own internal review while substantially complying with the NIST Generative AI Profile or another recognised AI risk framework. Colorado's 2024 AI law was repealed and reenacted by SB 26-189, signed on 14 May 2026, which covers automated decision-making technology used in consequential decisions. The law firm Holland and Knight reports that the revised law takes effect on 1 January 2027. In California, the privacy agency's rules on automated decision-making technology were approved on 22 September 2025. The law firm Hunton gives 1 January 2027 as the compliance date for businesses that use the technology for significant decisions.
Where to start
Pick one AI system that is already in use and walk it through the seven parts, asking each time for the record named in the table. A missing record is the first piece of work. For AI agents, the controls in our earlier post on agent approvals and audit trails sit inside parts five to seven.
Where QueryNow fits
QueryNow builds the foundation enterprise AI runs on and the agents that run on it, on Microsoft and AWS platforms. In governance work that usually means setting up the inventory and the audit logging described above, so that each part leaves its record. You can talk to our team about the system you would start with.
QueryNow leads AI transformation for enterprises in the US and Europe. Founded in Germany in 2014, we build the foundation enterprise AI runs on, and the agents that run on it, for pharma, healthcare, manufacturing and financial services organizations on Azure and AWS.
FAQ
Frequently asked questions
What is AI risk governance?
AI risk governance is the set of roles and records a company uses to decide which AI systems it runs and to manage the harm they could cause. NIST describes its AI Risk Management Framework as a way to better manage risks to individuals, organizations and society. In practice the work divides into seven parts, from roles and responsibilities through to monitoring.
Does ISACA have an official definition of AI governance?
The definition usually quoted under ISACA's name comes from a guest article on the ISACA Now blog, published in December 2024 and written by an engineer at Dell Technologies. It is best attributed to that author and not to ISACA as an institution. ISACA's own 2026 toolkit on governing AI across its lifecycle is the more practical source, because it follows a system through five stages.
Who should own AI governance in a company?
One named executive should be accountable, with the authority to stop a deployment. Survey data shows the job most often sits with privacy teams or with legal and compliance teams, and IT holds it in some companies. Any of these can work. Each AI system also needs its own business owner, who answers for its inventory entry and its monitoring.
When do the EU AI Act high-risk duties apply?
After Regulation (EU) 2026/1744, duties for high-risk systems listed in Annex III apply from 2 December 2027, and those for AI built into Annex I products apply from 2 August 2028. The transparency rules in Article 50 kept their date of 2 August 2026. This is general information and not legal advice.
Is SR 11-7 still the model risk standard for banks using AI?
The Federal Reserve's SR 26-2 of 17 April 2026 superseded SR 11-7. The attached guidance says generative and agentic AI models are outside its scope, and that a bank's own risk management and governance practices should guide the controls for anything it does not cover. Its principles still apply to traditional models and to AI models that are neither generative nor agentic.
The Monday digest
The three posts worth your time each week, one paragraph each. From the people building production AI.
