Securafy AI Lab

AI Literacy Compliance Requirements: Where Training Is Law

Written by Ric Hall | Nov 12, 2026, 1:00:00 PM

"Is AI training actually a legal requirement now, or is that just what training vendors want us to believe?" It is a fair question from any owner or compliance lead who has watched AI rules pile up through 2026. The answer is more specific than either side usually admits, and the specifics change what kind of training you should be paying for.

A handful of rules now write AI literacy compliance requirements directly into their text. Most of the rest never mention training at all, yet treat a trained workforce as the evidence that decides whether you were careful. And the explicit duties share a pattern that a company-wide awareness course does not satisfy.

Which AI rules make literacy a compliance requirement?

Four rules relevant to U.S. businesses spell out training or competence for people who use or oversee AI: the EU AI Act's Article 4 and Article 26, Colorado's replacement AI law, and the NAIC model bulletin for insurers. Others, including Texas's TRAIGA and Illinois's employment amendment, impose no training mandate but make trained, documented oversight your strongest defense.

That split matters because it tells you where training is a checkbox you can fail outright and where it is evidence you will need when something goes wrong. Both are compliance problems. They just fail in different ways, and they reward different kinds of programs.

The take most compliance plans miss: the explicit duties attach to roles, not headcount

The common assumption is that AI literacy compliance means getting everyone through a course. Read the actual text and a different picture emerges. Apart from the EU's general literacy duty, the explicit requirements are aimed at specific people: the person assigned to oversee a high-risk system, the person designated to review an adverse automated decision, the personnel who operate an insurer's AI systems. The law cares less that everyone watched a video and more that the person with their hand on the decision was trained and had the authority to act.

That is why a single all-staff module is the wrong investment for a regulated business. It spreads budget evenly across people with very different exposure, and it leaves the handful of roles that regulators will actually ask about with the same generic material as everyone else. The defensible program starts by naming those roles.

RuleTraining stated in the text?Who it attaches toTiming
EU AI Act, Article 4Yes, measures to support AI literacyStaff and others operating AI on the organization's behalfApplies since Feb. 2025; national enforcement from Aug. 2026
EU AI Act, Article 26(2)Yes, competence, training and authorityPeople assigned human oversight of high-risk systemsEmployment and other Annex III systems from Dec. 2027
Colorado ADMT Act (SB 26-189)Yes, for the designated human reviewerThe person who reviews and can override adverse decisionsJan. 1, 2027, if rulemaking is complete
NAIC AI Model BulletinYes, ongoing training and supervisionInsurer personnel within the AI Systems programIn effect where adopted
Texas TRAIGANoDefense tied to recognized risk frameworksIn effect since Jan. 1, 2026

The EU: a softer general duty and a hard role-specific one

The amended text of Article 4 requires providers and deployers to "take measures to support the development of AI literacy" of staff and others operating AI on their behalf, and adds that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual." The change came through the Digital Omnibus, published as Regulation (EU) 2026/1744 and in force since July 27, 2026. According to the European Commission's AI literacy Q&A, national market surveillance authorities have supervised and enforced the duty since August 2, 2026, no certificate is required, and an internal record of training is the expected evidence.

The role-specific duty is sharper. Article 26(2) says deployers of high-risk systems "shall assign human oversight to natural persons who have the necessary competence, training and authority." The same Omnibus moved standalone high-risk systems, including AI used in employment and credit decisions, to December 2, 2027. A U.S. company is in scope if it deploys AI in the EU or if the output of its AI system is used there, so an EU subsidiary or EU-based staff is enough to start the clock.

Colorado: the risk program disappeared, the trained reviewer stayed

Colorado offers the clearest recent example of which obligations survive political pressure. On May 14, 2026, the governor signed SB 26-189, which repealed and replaced the Colorado AI Act with the narrower Automated Decision-Making Technology Act before the original law ever took effect. The replacement dropped the duty of care against algorithmic discrimination, the annual impact assessments, and the risk management program.

What it kept is telling. When covered technology materially influences a consequential decision in employment, lending, insurance, housing, health care, education, or government services and the consumer gets an adverse outcome, the consumer may request human review. Deployers must designate an individual, train that individual to conduct reviews, and give them authority to override the decision. The law takes effect January 1, 2027, if the attorney general completes rulemaking, with penalties of up to $20,000 per violation. Lawmakers stripped out the risk program and impact assessments and kept the trained human with authority. That is the clearest signal yet of where regulators think the real control sits.

If you have been tracking the earlier version, our analysis of how Illinois, California and Colorado rewrote the rules for AI in hiring covers the employment side in more depth, including Illinois's notice requirement, which applies to employers with even one employee in the state but contains no training mandate of its own.

Insurance: training is a named program element

For insurers and the agencies and vendors that serve them, the NAIC Model Bulletin on the use of AI systems by insurers lists "development and implementation of ongoing training and supervision of personnel" as part of the written AI Systems program. It also tells insurers that regulators may request the program's "policies, procedures, guidance, training materials, and other information" during an investigation or market conduct action. Training is not implied there. It is a document an examiner can ask to see.

About half of U.S. jurisdictions have adopted the bulletin, and examiners have been piloting a structured way to test it, which we covered in the AI exam that exposes paper governance.

Is AI training required if a law never mentions it?

Not by name, but it often decides the outcome. Texas shows how. The enrolled text of TRAIGA gives the attorney general sole enforcement power, a 60-day cure period, and penalties of $80,000 to $200,000 for each violation that cannot be cured. It also protects a defendant who discovers a violation through internal review while substantially complying with NIST's generative AI risk profile or another recognized AI risk framework.

No clause in that defense says "train your staff." But every recognized risk framework assumes people who understand the system they are reviewing, and an internal review conducted by people who cannot explain the tool is not one a regulator will credit. The same logic runs through Illinois's ban on AI that has a discriminatory effect: the statute does not require training, but the person who catches a biased screening result before it becomes a claim is a trained person. This is the shared pattern we described in every new AI rule asking the same question, and the point at which AI without governance becomes a risk you can no longer explain away.

What a defensible AI literacy program looks like

A program that holds up under the explicit duties and the implied ones has a different shape from a typical awareness rollout. It is built around the decisions AI touches, not the org chart. In practice it has five elements:

  1. A role map. List every role that operates, oversees, or reviews AI that influences a decision about a person: hiring, claims, credit, pricing, patient communication, or service eligibility.
  2. Tiered training. A short common foundation for all staff, plus deeper, tool-specific training for the mapped roles, covering what the system weighs, where it fails, and when a human must step in.
  3. Named reviewers with authority. For each decision type, the person who can override the AI is identified in writing and trained for that job, which is what Colorado and Article 26 actually require.
  4. Records an examiner can read. Who was trained, when, on which tool, with what content, kept with the AI system inventory rather than in a separate HR folder.
  5. A change trigger. Training is refreshed when a tool, vendor, or rule changes. Colorado's rewrite and the EU's amendment both show how quickly the ground shifts.

None of this requires tracking every statute line by line. It requires knowing which of your people sit at the point where AI touches a decision, and proving they were ready for it. For operations leaders in professional services, our guide to compliance in the age of AI shows how the same role-first approach applies to client data and confidentiality duties. It also fits inside the broader control set in cybersecurity and compliance for SMBs in 2026.

How Securafy helps regulated teams build this

From a compliance standpoint, the hard part is rarely the training content. It is establishing which AI systems are actually influencing decisions, including features embedded in HR, CRM, and claims platforms that nobody consciously deployed. We start there, building an inventory from the client's tenant and application data, then map each AI-influenced decision to the role that owns it and the rule most likely to apply.

With that map in place, we help clients set up tiered training, document named reviewers and their override authority, and keep training records attached to the system inventory so they can be produced on request. For insurers, agencies, healthcare practices, and financial firms, that turns a scattered set of policies into an evidence file a regulator, examiner, or opposing counsel can follow from tool to person to proof.

Where To Go From Here

Start with the short list of roles that review or override AI-influenced decisions about people. Those are the positions the explicit rules name, and the ones where an undocumented gap carries the most exposure.

If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. AI University gives your people role-based AI training with security built into the material, not bolted on afterward.

If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.