If the Department of Labor now has an official definition of AI literacy, does the training your team finished this year actually meet it? For most small and mid-sized companies the honest answer is "partly," and the parts that are missing are the ones that decide whether AI makes your people better at their jobs or just faster at producing work someone else has to fix.
The Department of Labor's AI Literacy Framework, issued as Training and Employment Notice 07-25 on February 13, 2026, is the first federal attempt to say what AI literacy requires of an ordinary worker. Read closely, it defines literacy as judgment and accountability, not tool fluency, and that is where most training falls short.
Strictly, it requires nothing: the framework is voluntary guidance with no new obligations on employers. What it does is define AI literacy as five content areas and seven delivery principles, and the definition leans hard on evaluating output, protecting information, maintaining accountability, and preparing managers. Training that only teaches prompting meets a small slice of it.
The full TEN 07-25 document calls itself "voluntary guidance to support the evaluation of high-quality AI literacy content," and the Congressional Research Service summary notes that even federally funded workforce grantees are not required to use it. So the framework's value to a private employer is not compliance. It is that a federal agency has now written down what "AI literate" should mean, and your people, your clients, and your next hires will increasingly be measured against that wording.
Our earlier overview of the DOL's AI literacy standard walks through the five areas and seven principles at a high level. This piece reads the fine print, because the fine print is where the gaps show up.
The common belief is that an AI-literate employee is one who uses the tools confidently and writes good prompts. The framework quietly rejects that. Of its five content areas, only one, Direct AI Effectively, is about getting output from a tool. The other four are about knowing what the tool is, where it belongs, whether its answer is any good, and who is accountable when it is not.
Look at the sub-skills the department lists. Understanding AI Principles includes "hallucinations and accuracy limits" and "human design and oversight." Evaluating AI Outputs covers factual accuracy, completeness, logical errors, "alignment with strategic intent, and the need for human judgment." Using AI Responsibly includes "protecting sensitive information, following workplace policies," and "maintaining accountability." That is a definition built around a person who can say no to the tool, not one who can coax more out of it.
This is the same distinction we make on the main blog when we argue that the AI skills gap is not a tool problem. A team can be highly fluent and still not be literate in the framework's sense, and fluency without judgment is how organizations end up with polished, confident, wrong work moving faster than anyone can review it.
The research on how companies actually train suggests the shortfall is structural, not a matter of a few missing topics. In DataCamp's survey of more than 500 U.S. and UK enterprise leaders, run with YouGov, 82 percent of organizations offered some AI training, but only 35 percent had a mature, organization-wide upskilling program, and video-based courses were the most common format. Leaders cited a lack of hands-on projects and training not tailored to specific roles among their top problems.
Employees describe the same gap from the other side. Docebo's AI Readiness Gap report, based on 2,000 respondents across six countries, found that 85 percent of employees say their training does not help them use AI in their role, one in five received no AI training at all, and nearly 60 percent feel learning programs were not designed with people like them in mind. Set those findings beside the framework's text and four specific gaps stand out.
| What the framework asks for | What most programs deliver | The gap |
|---|---|---|
| Evaluating outputs for accuracy, logic, intent, and need for human judgment | A slide warning that AI "can make mistakes" | No practice catching errors in real work |
| Hands-on, experiential learning on actual tasks | Recorded videos and a completion quiz | Knowledge without a working habit |
| Prepared enabling roles: managers, trainers, mentors | Managers take the same course as their teams | Nobody equipped to review AI-assisted work |
| Design for agility with built-in update mechanisms | A fixed course refreshed annually, if ever | Content that lags the tools people use |
Almost every course tells people to double-check AI output. Very few make them do it on material from their own job, with errors planted that they have to find. The framework treats evaluation as a full content area, not a disclaimer, and its delivery examples include side-by-side comparisons of human and AI work. If your training never asked someone to reject an AI answer, it did not teach this area.
The framework's sixth principle says the managers, trainers, and mentors who support workers "are not just secondary learners" and "require tailored approaches." That is the gap we wrote about in why the AI adoption gap is a manager problem. When a manager cannot tell strong AI-assisted work from weak, the review step that the evaluation area depends on does not exist in practice.
The framework is explicit that AI literacy "cannot be treated as a fixed curriculum" and that training needs "built-in mechanisms for adaptation." For a small company, the practical trigger is not the calendar but the software: when a new AI feature appears in a platform your team already uses, the training should change with it.
Not as a legal requirement. It was addressed to state workforce agencies, workforce boards, community colleges, apprenticeship programs, and other parts of the public workforce system, and it imposes no obligations on private companies. The CRS summary does list employers and individual workers among its intended audiences, which tells you how the department expects it to be used.
The practical reach is indirect but real. Community colleges and workforce programs that receive federal funding are the channels most likely to cite the framework when they build or update AI courses, which means it will shape what entry-level and career-changing hires were taught before they reach you. An employer whose internal training is thinner than the federal definition will be onboarding people who were trained to a higher bar than the company itself holds.
You do not need a learning department to close these gaps. A 20 to 300 person company can get most of the way there by changing how training is delivered rather than buying more of it. The changes that map most directly to the framework are these:
That last item is where AI literacy meets governance. A training program that teaches responsible use in the abstract, without reference to the organization's own rules, satisfies neither the framework nor your risk team. If your policy is not something people can follow from memory, start with our guide to writing an AI use policy people will actually follow.
It also changes how leaders should think about their own role. The framework's emphasis on enabling roles is a reminder that literacy is an organizational capability, not a personal achievement certificate, which is the theme of our leadership playbook for AI adoption in SMBs. And because so much AI work now runs semi-unattended, the evaluation habit is the same supervisory skill we described in who is supervising your AI intern.
The framework's timing matters because the tools have stopped arriving as separate products employees choose to open. Writing assistants, meeting summarizers, and copilots now ship inside email, document, CRM, and HR platforms, often switched on by default. That shifts the risk from "who is using ChatGPT" to "who is accepting AI output without noticing it was AI output."
In that environment, the evaluation and responsible-use areas stop being nice extras and become the only real control most organizations have at the point of work. Technical guardrails can block certain data or certain tools. They cannot decide whether a summary is accurate, whether a drafted email misstates a policy, or whether a recommendation fits the business's intent. Only a trained person can, which is why the framework puts so much weight there.
When we assess a client's AI training against the framework, we look at evidence rather than course catalogs: which roles practiced evaluation on real work, whether managers were prepared separately, and whether material changed when the tools did. We pair that with an inventory of the AI features actually active across their platforms, because the training should cover what people encounter, not what was popular when the course was written.
From there, the fix is usually narrow and practical: role-specific exercises built on the client's own documents, a manager session on reviewing AI-assisted output, and a named update trigger tied to tool changes. That turns the framework from a reading assignment into a program you can show working.
Read your current training against the framework's own sub-skills, not its headings. If evaluation, manager preparation, and content updates are missing, those are the three changes that will move your team from fluent to literate.
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.