AI Governance for Business Professionals

Have you ever used AI to make a life-changing decision? A job to apply for, a class to take, a loan, apartment, school, medical, or legal question.

The first business question is where AI output stops before it becomes action. A system may draft a rejection email, rank job applicants, recommend fraud review, summarize a disciplinary record, or flag a worker for low productivity. Those outputs matter because someone may rely on them. The legal and ethical risk usually appears at the moment the organization lets the output affect a person.

Model confidence does not solve that problem. Confidence is a signal generated by the system or around the system. It is not independent review. It does not tell the business that the output is lawful, fair, accurate, documented, or ready for use.

The core principle we will go over here:

The control has to sit outside the model. That control can be a trained reviewer, a compliance process, a legal review, a technical audit, a manager with defined authority, or a structured workflow with escalation rules. The form can vary, the power cannot. The gate has to stop movement.

The practical test is simple. If the reviewer cannot reject, revise, or escalate the output before it affects someone, the organization has not created a release gate. It has created a ritual.

The risk is that weak controls scale with the same speed. A bad screening rule can reject qualified applicants at volume. A biased dataset can reproduce historical exclusion. A vendor claim can move through procurement without anyone testing what the tool actually does. A monitoring system can turn ordinary worker behavior into discipline before anyone asks whether the signal is valid.

The National Institute of Standards and Technology describes trustworthy AI through characteristics that include validity, reliability, safety, security, resilience, accountability, transparency, explainability, interpretability, privacy, and fairness. Its AI Risk Management Framework organizes AI risk work around 4 functions: govern, map, measure, and manage. That structure matters because business governance is a system of decisions, records, testing, review, and accountability. It is not a slogan.¹

The United States does not have one comprehensive federal AI statute. Businesses operate in a layered environment made of executive action, agency enforcement, state and local law, private litigation, and existing legal regimes such as employment discrimination, consumer protection, privacy, contract, procurement, and discovery. Federal agencies have continued to use existing authority while states move ahead with AI-specific rules.

That means a company does not need to violate an AI law to create AI liability. If a hiring tool creates discriminatory outcomes, employment law may apply. If a consumer scoring tool misleads people or treats them unfairly, consumer protection law may apply. If an AI system uses data in ways the business did not disclose, privacy law may apply. The old laws did not disappear when the new tools arrived.

Federal policy has moved toward national uniformity. Executive Order 14365, issued December 11, 2025, criticized state-by-state regulation, created an AI Litigation Task Force, directed Commerce to identify conditions on certain remaining BEAD non-deployment funds, and directed preparation of a legislative recommendation for a uniform federal framework that would preempt conflicting state AI laws.^3,4

An executive order does not erase state law by itself. Federal preemption usually comes from Congress, or from agency action when Congress has clearly delegated authority to the agency. At the time of this reading, the order itself had not displaced state AI laws, and Congress had not enacted the broad AI preemption requested by the administration.

Congress has also resisted broad state-law displacement. In July 2025, the Senate stripped an AI moratorium from a budget reconciliation bill by a 99 to 1 vote, and a later attempt to block state AI laws through the National Defense Authorization Act was dropped before enactment.^5,6

The business answer is direct. Comply with the state and local AI rules that apply now, keep the program flexible, and do not treat a federal policy announcement as a compliance exemption.

Mobley v. Workday puts the abstract risk into a business setting. Derek Mobley alleged that Workday's algorithmic applicant-screening tools discriminated against applicants based on race, age, and disability. These are allegations. At the pleading stage, the court was deciding whether claims could proceed, not whether Workday violated the law.¹⁵

The important point for business professionals is the vendor theory. Workday argued that it was a software vendor rather than an employer. The court allowed certain federal discrimination claims to proceed under an agency theory, reasoning that anti-discrimination laws reach agents and that employers cannot avoid liability by delegating traditional hiring functions. The complaint plausibly alleged that Workday's customers delegated screening and rejection functions to Workday's tools.

The court dismissed the employment-agency theory, allowed disparate-impact claims to proceed, and dismissed intentional-discrimination claims because the complaint did not adequately plead discriminatory intent. That distinction matters. Disparate impact focuses on outcomes, so a tool can create risk even when nobody set out to discriminate.¹⁵

The case also shows how quickly scale changes the risk profile. In a preliminary collective-certification order, the court noted Workday's representation that roughly 1.1 billion applications were rejected using Workday during the relevant period. That was Workday's account of platform scale, not a finding that its AI independently rejected 1.1 billion people. The same order treated the certified older-applicant group as an opt-in collective, which is different from a Rule 23 class.¹⁶

The case kept moving. In a later order, the court allowed key California FEHA claims and a disability proxy-discrimination theory to go forward while dismissing other theories, including a race-based disparate-impact claim by one plaintiff and the direct-employer theory. The detail matters because AI litigation rarely moves as one clean yes or no. Claims survive, narrow, and change as the court tests the theory against the record.¹⁷

The discovery fight adds the governance lesson that many businesses miss. In May 2026, the court held that certain Workday bias-testing material was protected by attorney-client privilege because lawyers directed and curated the testing for legal advice. The ruling did not say that all bias testing is privileged. It showed that audit structure matters: who runs the test, why the test is run, who controls the records, and how the purpose is documented can change the legal posture of the evidence.¹⁸

The lesson is a design choice the business makes before the audit runs, not after it. A bias audit performed as a routine operational task creates an ordinary business record that an opposing party can request in litigation. The same test, directed by counsel for the purpose of legal advice, may carry privilege. A business cannot relabel the record once a dispute begins, and privilege can still be waived by how the results are circulated. The purpose, the owner, and the record path have to be decided at the start.

The EU AI Act matters to U.S. businesses because of its reach. It applies to certain providers placing AI systems or general-purpose AI models on the EU market, deployers located in the EU, and providers or deployers outside the EU when the AI system's output is used in the European Union. A U.S. company can therefore face EU AI Act obligations even when the company itself sits outside Europe.

For a U.S. company, the practical trigger is not where the company sits but where the output is used. If an AI system's results are used inside the EU, the obligations can apply even when the company has no European office. The company then faces a choice between building one product that meets the stricter standard and maintaining two versions for two markets. Because running two products usually costs more than complying once, the stricter rule tends to become the default, which is how an EU requirement reaches a business that never planned to operate in Europe.

High-risk systems carry heavier obligations, including risk management, technical documentation, transparency, human oversight, accuracy, robustness, cybersecurity, monitoring, and recordkeeping. Employment uses listed in Annex III require Article 6 analysis, including the exceptions for certain narrow procedural or preparatory uses.^19,20

The timing has changed, and precision matters. In June 2026, the Council gave final approval to a Digital Omnibus package that moved certain high-risk AI obligations. Stand-alone high-risk systems move to December 2, 2027, and high-risk systems embedded in regulated products move to August 2, 2028. The same package sets December 2, 2026 for certain synthetic-content marking transition rules.

The delay is a readiness signal, not a reprieve. The high-risk obligations moved because the standards, guidance, and national authorities needed to apply them were not ready in time, not because the requirements themselves got lighter. The core high-risk obligations remain, but the package also changes how parts of the regime apply and interact with sector-specific rules. The practical reading for a business is that the extra time exists for building the compliance framework, and treating the new dates as permission to wait leaves less runway than it looks.^21,22

The Act also imposes transparency duties. People must be informed when they are interacting with certain AI systems, and certain AI-generated or manipulated content must be marked or disclosed.

The Act separately requires providers and deployers to take measures to ensure a sufficient level of AI literacy among staff and others involved in AI operation and use. AI literacy means that the people using the system understand enough about its risks, limits, and proper use to operate it responsibly.

AI literacy is an obligation, not a slogan. It means the people operating an AI system can judge its outputs, recognize when it is likely wrong, and understand where its limits fall, well enough to use it responsibly. That is the same capability this reading builds: the vocabulary to ask what the system did, who it affects, whether anyone tested it, and who can hold release authority. A workforce that cannot answer those questions is not AI literate, whatever the training record says.

The business answer is practical. If an AI tool is placed on the EU market, used in the EU, or produces output used in the EU, the company has to analyze EU obligations. If the tool affects hiring or workforce management, analyze whether the intended use falls within Annex III and build oversight into the workflow now. Retrofitting governance is expensive.^19,20

For business use, explainability does not always require exposing every internal mathematical feature of a model, just enough understanding to govern the system, challenge the result, document the decision, and explain the main drivers in ordinary language when a person is affected. A black box may be acceptable for low-stakes automation. It is much harder to defend when the system rejects, ranks, flags, disciplines, monitors, prices, or deprioritizes people.

The four-fifths rule is an evidentiary screen: a selection rate below 80 percent may indicate adverse impact.²³

The AI-specific lesson is direct. A system can be facially neutral and still produce discriminatory outcomes. The organization does not solve that risk by saying the model did not know the applicant's race, sex, age, disability, or other protected trait. Proxy variables can carry the signal. Zip code, employment gaps, school history, commute patterns, medical-leave patterns, language, scheduling availability, or work history can operate as substitutes for protected traits in practice.

The federal Uniform Guidelines on Employee Selection Procedures define adverse impact in employment selection and require recordkeeping that allows employers and enforcement agencies to assess selection effects. The common rule of thumb is the four-fifths rule: if the selection rate for a race, sex, or ethnic group is less than 80 percent of the selection rate for the group with the highest selection rate, that difference is generally treated as evidence of adverse impact. Smaller differences may still matter where they are statistically or practically significant.²³

A human in the loop is useful only when the person has real authority, enough information, and the power to stop the process. A reviewer who only watches the output move through the system does not create oversight. A manager who can rubber-stamp but cannot reject does not create oversight. A policy that says human review while the workflow auto-rejects people before review does not create oversight.

The EU AI Act's human-oversight provision for high-risk systems points in this direction. It requires oversight measures that allow a natural person to:¹⁹

A clean vendor review should answer 12 questions.

These questions do not slow the business for the sake of caution. They protect the business from adopting a system it cannot explain, defend, or stop.

A responsible AI implementation should produce clear answers before the system goes live.

The checklist is only useful if the business can answer it with evidence. A policy statement is not evidence. A vendor assurance is not evidence. A release gate that cannot stop anything is not evidence.

For business professionals, the point is operational. Do not say human oversight in the abstract. Name the person, the decision, the information available to that person, and the exact action that person can stop.