An AI assistant that drafts a proposal and an AI system that rejects a job applicant may share underlying technology, but they should not share the same approval process. Enterprise AI governance begins with that distinction: what the system is allowed to do, whose interests it affects and who can intervene when it fails.
Direct Answer: AI governance is the system of accountability, policies, approval processes and controls an organization uses to oversee AI throughout its lifecycle. Enterprises should connect each AI use case to applicable laws, assign a business owner, evaluate its risks, restrict its data and action permissions, and retain evidence that required safeguards work.
This guide focuses on U.S. enterprise operations and relevant European Union exposure, with regulatory status checked as of September 30, 2026. The operating framework below is a recommended approach; legal requirements are identified separately.
AI Governance Must Distinguish Laws, Standards and Company Rules
A workable program separates obligations imposed by law from frameworks selected by the business and restrictions written into its own policies.
Figure 1. Company rules should be more restrictive than the law wherever potential harm exceeds the organization's tolerance, even when no AI-specific statute addresses the use.
Laws and regulations establish requirements for covered organizations and activities. Their applicability depends on the jurisdiction, business role, affected people and particular use of the technology.
Voluntary frameworks and standards help organize governance. The NIST AI Risk Management Framework is voluntary and organizes risk management around Govern, Map, Measure and Manage. ISO/IEC 42001:2023 specifies requirements for an organizational AI management system.[1][2]
Company rules establish approved tools, permitted data, review requirements and limits on automation. Enterprises should make those rules more restrictive where the potential harm exceeds their tolerance, even if no AI-specific law directly addresses the use.
Contractual commitments add another layer. Review customer and supplier agreements for restrictions on processing, subcontractors, confidentiality and AI use before deciding that an otherwise permissible workflow is authorized.
Keep these layers separate in the compliance register. A standard can support a control program, but adopting it does not establish that every deployment satisfies every applicable law. Internal approval likewise cannot waive a legal obligation.
Determine Which AI Regulations Apply to Each Business Use
Start with a deployment record rather than a country checklist alone. Record where the enterprise operates, where affected people are located, where outputs are used, what decisions the system influences and whether the organization develops, supplies or deploys it.
The following requirements illustrate why that detail matters. They are selected enterprise examples, not an exhaustive global legal inventory.
U.S. Enforcement Includes Existing Business Laws
AI use does not remove existing restrictions on unfair or deceptive conduct. The FTC's September 2024 Operation AI Comply targeted alleged deceptive AI claims and AI-enabled schemes, including unsupported claims about professional capabilities.[3]
Employment requires a separate review. Employers' obligations under the Americans with Disabilities Act extend to hiring technology, including technology purchased from another company. Screening methods must avoid unlawful disability discrimination, and reasonable accommodation requirements still apply.[4]
For enterprise teams, the practical implication is to involve the function responsible for the underlying activity. Marketing should substantiate claims. HR should evaluate selection methods and accommodation procedures. Legal and compliance should identify additional sector requirements for lending, insurance, health care or other regulated decisions.
New York City Applies Specific Conditions to Covered Hiring Tools
New York City's Local Law 144 restricts employers and employment agencies from using covered automated employment decision tools unless a bias audit was completed within one year of use, required audit information is publicly available and required notices are provided.[5]
Do not treat every HR application as covered or every vendor audit as sufficient. Document the tool's actual role in screening or selection, determine whether the legal definition applies and assign responsibility for notices and publication.
Colorado's Requirements Changed in 2026
Colorado's SB26-189 repealed and reenacted the provisions established by its earlier AI law. The replacement addresses automated decision-making technology that materially influences consequential decisions, including covered employment, housing and financial decisions. Developer documentation requirements begin January 1, 2027.[6]
The enacted summary identifies consumer notices, descriptions following adverse outcomes, correction rights and meaningful human review. Developers and deployers must retain compliance records for at least three years.[6]
An enterprise using an older Colorado compliance checklist should reconcile it against the replacement law. Preserve useful risk controls, but distinguish company policy from duties that the current statute actually imposes.
California Separates Effective Dates From Compliance Deadlines
California's finalized privacy regulations define covered automated decisionmaking technology around processing personal information to replace or substantially replace human decisionmaking. Article 11 addresses use for significant decisions and sets January 1, 2027, as its compliance date.[7]
Requirements include pre-use notices and access and opt-out rights, subject to specified exceptions. The definition of a significant decision excludes advertising to a consumer; that exclusion should not be interpreted as an exemption from other privacy requirements.[7]
For a marketing or sales operation, classify the specific processing activity. An advertising audience and an automated eligibility decision require different analysis even when both use customer records.
The EU AI Act Can Reach Enterprises Outside Europe
The EU AI Act covers relevant providers placing AI systems on the EU market and deployers located in the EU. Its scope also includes certain providers and deployers outside the EU where the system's output is used in the Union, subject to the Act's exclusions and qualifications.[8]
A U.S. headquarters address therefore does not settle applicability. Map the actual service, users and output destinations.
Figure 2. The Commission's current implementation timeline incorporates the Digital Omnibus amendments. These are different obligations with different triggers.[9]
| Milestone | Application date |
|---|---|
| General provisions, AI literacy and initial prohibited practices | February 2, 2025 |
| General-purpose AI model rules, subject to applicable transitions | August 2, 2025 |
| Article 50 transparency rules | August 2, 2026 |
| Certain additional prohibitions and a transition for specified preexisting synthetic-content systems | December 2, 2026 |
| Annex III high-risk AI rules | December 2, 2027 |
| High-risk AI embedded in regulated products covered by Annex I | August 2, 2028 |
These are different obligations with different triggers. A later high-risk deadline does not postpone transparency duties already applicable to a covered system.[9]
Classify the Workflow, Not Just the AI Product
Approve a defined use with defined limits. A product name, procurement approval or enterprise subscription cannot describe every way employees might apply the technology.
Figure 3. The relevant review questions change from factual accuracy to access control, confidentiality and the effect on the individual, even though the underlying model is identical.
A useful internal assessment examines four dimensions.
Figure 4. Use the answers to set an internal review level, then record it alongside the separate legal analysis rather than in place of it.
Use the answers to set an internal review level. Keep that label separate from statutory terms such as high-risk AI or consequential decision. A company may require extensive controls for an application that does not meet a particular law's definition.
Changes Can Alter the Enterprise's Legal Role
Under Article 25 of the EU AI Act, certain changes can cause a deployer or other third party to assume provider responsibilities for a high-risk system. Examples include substantial modifications or changing an intended purpose so that a previously non-high-risk system becomes high-risk.[10]
Include a legal-role reassessment in change control. Fine-tuning, integration or rebranding should trigger questions about the actual change rather than an automatic assumption that every modification has the same legal effect.
Assign Approval Authority to Named Owners
The business owner should be accountable for the workflow's purpose, accepted risks and results. Technical staff should be responsible for implementing and operating the controls within their remit.
A recommended division of responsibilities is:
- Executive sponsor: Sets risk tolerance, funds remediation and resolves cross-functional disputes.
- Business owner: Defines intended use, acceptance criteria, escalation procedures and operational fallback.
- Legal and compliance: Determines applicable obligations and reviews unresolved legal restrictions.
- Privacy and security: Reviews data handling, access, retention, connected systems and incident exposure.
- Technical owner: Maintains configuration, evaluation, logging, monitoring and rollback capability.
- Independent assurance: Tests whether approved controls operate as described.
An AI governance committee should decide exceptions and material risks. Routine uses that satisfy an approved pattern should have a delegated approval route.
Document who can authorize launch, expand permissions and suspend use. A committee charter is incomplete if nobody on call can stop a harmful workflow.
Write Enterprise AI Rules That Employees Can Follow
An acceptable-use policy should translate principles into decisions employees face during ordinary work. The following are recommended policy provisions, to be adjusted for applicable obligations and the organization's risk tolerance.
Approve Tools Together With Their Permitted Data
Specify which applications and configurations can receive public information, internal material, customer data and restricted records. Identify prohibited inputs and an approval route for exceptions.
Separate permission to access a source system from permission to send its contents to an AI service. Require teams to document the intended processing before connecting repositories, CRM records or employee information.
Define Review Requirements by Output and Consequence
Identify which outputs require a qualified reviewer before publication, customer delivery or operational use. State what the reviewer must check: factual support, calculations, commitments, source accuracy or suitability for the decision.
Provide reviewers with the underlying evidence and authority to reject the output. A checkbox without time, information and decision authority is a weak control.
California's ADMT definition illustrates the importance of substance: qualifying human involvement requires a reviewer to understand the output, analyze relevant information and have authority to make or change the decision.[7]
Separate Drafting Authority From Execution Authority
Write distinct permissions for creating a recommendation and carrying it out. Approval to draft an email should not automatically authorize sending it. Approval to analyze accounts should not automatically authorize changing records.
Figure 5. For consequential actions, specify the approving role, permitted transaction scope and required record of authorization.
Make Exceptions Temporary and Reviewable
Each exception should identify its business justification, owner, compensating safeguards and expiry condition. Renew it only after reviewing whether the original constraint still exists.
Include a reporting path for unauthorized use and suspected failures. Employees need to know where to report an issue and what information to preserve.
Evaluate AI Vendors Against the Deployment You Intend to Run
Procurement should request evidence relevant to the specific service, configuration and workflow. Broad assurances about responsible AI provide little basis for an approval decision.
Before signing or enabling a feature, resolve:
- Whether submitted data can be used for training or other secondary purposes.
- Retention, deletion and backup handling for inputs, outputs and logs.
- Processing locations, subprocessors and administrative access.
- Known limitations for the intended task and supported languages.
- Available evaluation results and the populations or conditions they cover.
- Notification of material model, configuration or service changes.
- Access to records needed for investigations, consumer requests or audits.
- Exit arrangements, including data export and removal of credentials.
Treat unanswered questions as limits on the approved use. For example, an enterprise may decide to permit public-data drafting while withholding approval for confidential records until contractual and technical questions are resolved.
Vendor certifications should be reviewed for scope. ISO/IEC 42001 concerns an organization's AI management system; procurement should inspect which organizational activities and services an assurance claim covers rather than treating the label as a performance test for a particular deployment.[2]
Test Failures Before Launch and Retain the Decision Evidence
Set acceptance criteria before running evaluations. Otherwise, a team can rationalize a disappointing result after spending time and budget on implementation.
For each use case, define the task, acceptable output, material failure, evaluation population and conditions that require escalation. Test ordinary cases, ambiguous inputs, missing information and foreseeable misuse.
A proposed evaluation record should explain:
- What was tested and why those cases represent the intended use.
- Which model version, configuration and data sources were involved.
- Which failures occurred and how severe they were.
- Which safeguards reduced the risk and which limitations remain.
- Who accepted the remaining risk and under what restrictions.
Measure performance at the decision level. A document assistant should be tested for unsupported statements and incorrect source references. A workflow that ranks people needs evaluation relevant to that decision and population, alongside its legal review.
Where sample sizes are inadequate, record the limitation. Absence of an observed failure in a small test should not be described as proof that the failure cannot occur.
Build a Compact Evidence Package for Every Approved Use
Figure 6. Use retention schedules appropriate to the information and applicable obligations. Preserve useful audit evidence without indiscriminately storing sensitive prompts or personal data indefinitely.
Keep the inventory record, applicability assessment, vendor documents, evaluation results, approval conditions, training requirements and incident procedure together.
Make each control traceable to an owner and evidence. A requirement that human review is required should point to the review procedure and records demonstrating that it happened.
Govern AI Agents Through Permissions and Transaction Controls
AI agents require review of what they can do in connected systems. Excessive functionality, permissions and autonomy can allow damaging actions when a model produces unexpected or manipulated outputs.[11]
Recommended controls include limiting available tools, granting only necessary permissions, requiring approval for high-impact actions and enforcing authorization in downstream systems.[11]
Apply those principles to the workflow's actual transactions. For a proposed sales agent, separately approve reading account data, drafting outreach, sending messages, updating CRM fields and changing commercial terms. Do not combine those permissions into a single sales assistant approval.
Prompt injection adds another testing requirement. Instructions embedded in external documents or websites can alter a model's behavior; retrieval-augmented generation does not fully eliminate that vulnerability.[12]
Test whether untrusted material can induce disclosure, redirect a task or trigger an unauthorized tool call. Treat retrieved content as information to evaluate, with permissions enforced outside the generated response.
Reassess AI When the Workflow Changes
Approval should specify the conditions that require another review: new data sources, expanded populations, different jurisdictions, changed models, additional tools or increased autonomy.
A recommended operating dashboard should track material incidents, overdue corrective actions, expired exceptions, unreviewed changes and high-impact deployments lacking current evaluations. Define the denominator for each measure so management can distinguish coverage from activity.
Also test shutdown and fallback procedures. Identify who disables access, how ongoing transactions are handled and how affected records or decisions will be reviewed.
Governance will become more demanding as enterprises give AI authority to act. The most durable approval record will connect a defined purpose to enforceable permissions, tested limits and a person empowered to intervene.
FAQ
Who Should Own Enterprise AI Governance?
Accountability belongs to the business owner of the workflow, supported by legal, privacy, security and technical functions. A governance committee decides exceptions and material risks, but a named individual should be able to authorize, restrict or suspend each approved use.
Does Buying an Enterprise AI Subscription Make Its Use Compliant?
No. A subscription establishes commercial terms for a product. Compliance depends on the specific workflow, the data it processes, the decisions it influences and the obligations that apply to the organization in each jurisdiction where it operates.
Do AI Rules Apply to Tools That Are Not Generative AI?
Often yes. Several requirements address automated decision-making technology rather than a particular architecture. Scoring, ranking and eligibility systems can fall within those definitions regardless of whether they generate content.
Must Every AI-Generated Marketing Asset Carry a Disclosure?
Not automatically. Transparency duties depend on the system, the content and the jurisdiction, and Article 50 of the EU AI Act sets specific obligations for certain systems.[13] Separately, existing rules against deceptive conduct continue to apply to claims made in marketing material.[3]
Is NIST AI RMF Mandatory for Every Enterprise?
No. The framework is voluntary.[1] Many organizations adopt it to structure governance, but adopting it neither creates nor discharges a legal obligation.
How Frequently Should an Enterprise Review an AI System?
Tie review to change and consequence rather than to the calendar alone. Reassess when data sources, populations, jurisdictions, models, connected tools or autonomy change, and schedule periodic review for high-impact deployments.
Sources
- National Institute of Standards and Technology, AI RMF Core, AI Risk Management Framework 1.0, 2023; current resource accessed September 30, 2026. airc.nist.gov
- International Organization for Standardization, ISO/IEC 42001:2023 – AI Management Systems, 2023. iso.org
- Federal Trade Commission, FTC Announces Crackdown on Deceptive AI Claims and Schemes, September 25, 2024. ftc.gov
- U.S. Department of Justice, Algorithms, Artificial Intelligence, and Disability Discrimination in Hiring, May 12, 2022. ada.gov
- New York City Department of Consumer and Worker Protection, Automated Employment Decision Tools, Local Law 144 of 2021; current guidance accessed September 30, 2026. nyc.gov
- Colorado General Assembly, SB26-189 – Automated Decision-Making Technology, enacted May 14, 2026. leg.colorado.gov
- California Privacy Protection Agency, Final Regulations Text – CCPA Updates, Cybersecurity Audits, Risk Assessments, ADMT and Insurance, 2025; sections 7001 and 7200 to 7222. cppa.ca.gov
- European Commission AI Act Service Desk, Article 2: Scope, Regulation (EU) 2024/1689; current text accessed September 30, 2026. ai-act-service-desk.ec.europa.eu
- European Commission AI Act Service Desk, Timeline for the Implementation of the EU AI Act, 2026 version incorporating Digital Omnibus amendments; accessed September 30, 2026. ai-act-service-desk.ec.europa.eu
- European Commission AI Act Service Desk, Article 25: Responsibilities Along the AI Value Chain, Regulation (EU) 2024/1689; current text accessed September 30, 2026. ai-act-service-desk.ec.europa.eu
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency, 2025. genai.owasp.org
- OWASP Gen AI Security Project, LLM01:2025 Prompt Injection, 2025. genai.owasp.org
- European Commission AI Act Service Desk, Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems, Regulation (EU) 2024/1689; current text accessed September 30, 2026. ai-act-service-desk.ec.europa.eu
Technology Strategy vs Digital Transformation: What’s the Difference and Why It Matters
Technology Strategy vs Digital Transformation: What’s the Difference and Why It Matters Organizations today rely heavily on technology to compete, innovate, and serve customers. As a result, two concepts frequently appear in business discussions: technology strategy...
What Is a Technology Strategy?
What Is a Technology Strategy? Modern organizations depend on technology not just to operate, but to compete. Customer expectations, remote work, automation, cybersecurity risks, and data-driven decision-making have fundamentally changed how businesses function. As a...
Palantir’s stock is winning over Wall Street. Another analyst just turned bullish.
**Palantir Technologies: Analyzing the Recent Market Selloff and Its Implications for AI Leadership** In recent weeks, Palantir Technologies has experienced a significant and, according to some analysts, “unjustifiable” selloff in its stock price. This downturn has...
The surprising reason why the U.S. economy is growing so fast
**Corporate Spending Set to Propel U.S. Economic Growth for Fifth Consecutive Year** As the U.S. economy continues to navigate the complexities of a post-pandemic landscape, corporate spending is emerging as a significant driver of growth, positioning the nation for...
Miami is not the next Silicon Valley. It’s something much weirder.
Technology’s power players are flocking to Miami once more. Earlier this year, venture capitalist and crypto and AI investor David Sacks declared that Miami is on track to overtake New York City as the financial capital of the United States. Stripe cofounder Patrick...
