AI Literacy That Sticks: Building a Workforce Program That Meets EU AI Act Article 4
AI literacy is the practical knowledge a person needs to use AI safely in their own job: what the tool is good at, where it goes wrong, and what the house rules are for the data they put into it. Most organizations buy it as a 20-minute video and a completion report. This lesson shows you how to build an AI literacy program for employees that changes what people do on Monday morning, and that leaves behind the record EU AI Act Article 4 expects you to be able to show.
This is Lesson 4 of Level 1 (AI Aware) on the five-level ladder that runs from L1 AI Aware to L5 AI Native. Lesson 1 on finding and governing shadow AI gave you a Shadow AI Register and a one-page Data-to-Tool Rule. Lesson 2 gave you an AI Data Inventory. Lesson 3 on choosing your first AI use case gave you a First Use Case Charter that names real users who are about to start working with an approved tool on real data. Now you make sure those users, their managers and the people checking the output know what they are doing. Lesson 5 on your first AI acceptable use policy and readiness assessment then writes the rules down and takes an honest readiness reading, using everything you have built as evidence.
The AI Literacy Problem and What It Costs
The email arrives late on a Thursday, from the chair of the works council, copied to HR. Before the AI assistant is rolled out to the rest of customer service, the council would like three things in writing: what AI training staff have had so far, broken down by role; whether the agency workers in the warehouse are included; and who staff should go to when the tool gets something wrong. HR forwards it to you with one line: "Do we have this?"
You pull the usage report for the approved tool from Lesson 1, six weeks after it went live. It shows three groups of roughly equal size.
One third of your staff have never opened the tool. They are not resisting. Nobody showed them what it is for, and they have a queue to clear. One third use it now and then. The final third use it every day, paste in whatever is in front of them, and send the output without reading it twice.
Then a line manager stops by and asks whether "that EU thing" means everyone needs a certificate by Friday. You do not know. Neither does HR. Legal says they will come back to you.
That afternoon contains three separate costs, and it helps to name them, because each one lands in a different budget.
- Licences bought and unused. You are paying per seat for the first group. That is a line your CFO will find at the next budget cycle.
- Output sent unchecked. The third group is one confident wrong answer away from the Air Canada chatbot ruling (BC Civil Resolution Tribunal, 14 February 2024), which the Level 2 lesson on building a golden test set takes apart in detail.
- A question you cannot answer. The works council asked a fair question, and a national authority could ask the same one. Right now you hold a purchase order for a tool and nothing else.
How big is the AI skills gap?
Large on training, and usage is now running ahead of it. In BCG's AI at Work survey of June 2025 (more than 10,600 respondents in 11 countries and regions), only one-third of employees said they had been properly trained on AI.
The 2025 survey also found a split. More than three-quarters of leaders and managers used generative AI several times a week, while regular use among frontline employees had stalled at 51%. A year later the picture moved. BCG's June 2026 AI at Work survey (11,749 workers across 14 markets, published 3 June 2026) found 74% of frontline employees now use AI daily or several times a week, up from 51% in 2025. The same release reports that 66% of those regular frontline users get limited or no guidance on what to do with the time AI saves them.
Read the 2026 number carefully. BCG defines frontline as "individual white-collar employees with no managerial responsibilities", so the figure says nothing about your warehouse, your drivers or your night shift. What it does tell you is that usage is outrunning guidance. Most regular users are working out the rules alone. Check your own usage report by site and shift, not the survey.
Treat vendor figures with care. The Cisco AI Readiness Index 2025 (October 2025) reports that 75% of its top "Pacesetter" group say their staff are proficient in AI, against 16% of other organizations. It is a self-reported vendor survey: a direction, not a measurement.
The duty that changed its verb
Now the story at the centre of this lesson. It has no villain and no breach. It is about one verb.
Article 4 of the EU AI Act, the AI literacy article, has applied since 2 February 2025. As first adopted in 2024, it told providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff. A provider is the organization that develops an AI system or puts it on the market. A deployer is the organization that uses one in its work. If you bought a general-purpose assistant and gave it to your staff, you are a deployer.
"Ensure" is a heavy word. Many organizations bought a course, pushed it to every inbox and chased 100% completion. Dashboards went green.
In May 2025 the European Commission published a Q&A on AI literacy. As summarized by Covington's Inside Privacy blog (27 May 2025), it said no certificate is needed and that internal records of training and other initiatives suffice. It prescribed no fixed content. It asked for something harder than a video: measures tailored to people's technical knowledge, experience and the context in which they use the system.
Then the verb itself changed. The Digital Omnibus on AI was adopted as Regulation (EU) 2026/1744, dated 8 July 2026, published in the Official Journal on 24 July 2026 and in force from 27 July 2026 (confirmed by White & Case). The amended Article 4 now reads: "Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf." It adds a sentence that matters to every HR director: the obligation "does not require providers or deployers to guarantee any specific level of AI literacy of any individual."
So the duty went from ensure a sufficient level to take measures to support the development, and the word "sufficient" is gone. Think about who that left where.
The organizations that built a certificate machine had prepared for a test that was never set. They hold a completion report, and their staff still paste customer data into the wrong box. The organizations that asked "which roles use AI, for what, at what risk, and what does each need?" and wrote down what they did are in the same good position under either wording.
Build for the question "what did you do, for whom, and why was that proportionate?" and you are covered whichever verb the law uses.
A working way to hold Article 4 in your head: reasonable, documented effort, proportionate to role and risk. That phrase is our reading of the sources, not legal text, and this lesson is not legal advice. Have legal or compliance check the current consolidated text before you quote it in a policy.
How AI Literacy Works: A Driving Licence, Not an Engineering Degree
Three ideas before the workshop. Each one changes a design decision you will make later.
Idea one: literacy is tiered by role and risk
Think about how a country handles driving. Nobody asks you to understand how a gearbox is machined. You need to know the rules of the road, to have spent hours behind the wheel with an instructor beside you, and to know what to do when something goes wrong. And the licence is not one size: a car, a truck and a coach full of passengers each need a different one, because the harm from a mistake is different.
AI literacy works the same way. A warehouse operative who never opens the tool, an agent who drafts forty replies a day with it, and the team lead who approves those drafts need different training. One module for all is one leaflet for a pedestrian, a commuter and a coach driver.
So this lesson uses three tiers. They are our synthesis of the sources, not an official scheme:
| Tier | Who | What they need |
|---|---|---|
| Everyone | All staff and contractors | What AI is, its limits (especially hallucination), and the data rules |
| Regular users | People who use the approved tool in their daily work | Hands-on practice with the approved tool on their own real tasks |
| Overseers | People who check or approve AI output, or oversee higher-risk systems | Human-oversight training: when to reject, how to correct, how to log it |
Hallucination, since we will use the word a lot, is when an AI system produces an answer that sounds confident and is wrong or made up.
The Commission's own Q&A gives a small, concrete picture of what "tailored to context" means. Staff who use a generative tool for advertising text or translation "should be informed about the specific risks, for example hallucination", and simply relying on the system's instructions for use, or asking staff to read them, "might be ineffective". Notice the level of detail. The example is about a person who translates with the tool knowing it can invent things. It is not about how neural networks work.
Idea two: literacy is practice, not information
You did not learn to drive from the handbook. You learned in a car, on a real road, with someone in the passenger seat saying "check your mirror".
BCG's June 2025 survey found the same thing for AI. Regular use was sharply higher among employees who had at least five hours of training and access to in-person training and coaching. BCG did not publish the exact usage figures for that threshold on the page we read, so we will not put a number on the uplift. The shape is clear enough to design to: hours, live practice and a coach matter more than access alone.
The same 2025 survey tells you who the coach should be. Positive sentiment toward AI rose from 15% to 55% when employees had strong leadership support, yet only about one-quarter of frontline staff said they got it. Your leaders are users. They are not yet sponsors. A program that skips them will stall at the frontline. This is the people side of BCG's rule of thumb, in use since at least 2019, that AI value is roughly 10% algorithms, 20% technology and data, and 70% people and processes. The Level 4 lesson on the AI operating model and the 10-20-70 rule builds the full change program on it. At Level 1, the 70% starts with a manager who can coach.
The Samsung source-code leak that opens the shadow AI lesson was, read this way, a literacy failure before it was a security failure: diligent people who had never been told where their text went.
Idea three: the record is part of the program
A second picture: the fire drill. Nobody learns to evacuate from a poster by the lift. Once a year you walk the route and the floor warden ticks names on a sheet. When the inspector visits, the sheet is what they ask for, because it shows the drill happened, on which floor, with whom.
Your training log is that sheet. It records who (or which role), which tier, what measure, when, how many hours, with which version of the material, and who delivered it. The log is how you find the night shift that missed the session. It is also your answer when the works council, or anyone else, asks what you did.
What does Article 4 of the EU AI Act expect in practice?
It expects measures you can explain, tailored to role and risk, and a record that they happened. It does not expect a certificate, a test score or a guaranteed level for any individual. The primary source is the Commission's AI Literacy Questions and Answers, which was revised after the omnibus and, when this lesson was checked on 13 November 2026, showed "last updated 20 October 2026". It is a living page, so open the current version before you rely on any line below.
- The minimum steps, which survived the update: ensure a general understanding of AI in the organization; consider whether you are a provider or a deployer; consider the risks of the systems you provide or deploy; and build your literacy actions on that analysis, allowing for differences in technical knowledge, experience, education and training.
- Who is covered: staff, and also people dealing with the operation and use of AI on your behalf, which the Q&A says can include a contractor, a service provider or a client.
- What counts as evidence: no certificate is needed, and an internal record of trainings and other guiding initiatives is enough.
- Examples to borrow: the Q&A points to a living repository of AI literacy practices with more than 40 examples, and warns that copying one does not by itself give you a presumption of compliance.
- Who checks: national market surveillance authorities have supervised and enforced Article 4 since 2 August 2026, according to the current Q&A. Covington's summary of the May 2025 version gave 3 August 2026; go with the Commission's current page.
One honest limit on penalties. The Q&A says authorities could impose penalties and other enforcement measures under national law, weighed by the gravity of the infringement and whether it was intentional or negligent, but it gives no amounts, and the research behind this lesson found no enforcement action on Article 4 as of November 2026. So we will not quote you a penalty.
Where does Article 4 meet the rest of the AI Act?
In three places your training content has to reflect, and each one has a date.
- Transparency, Article 50. These duties have applied since 2 August 2026. Providers of generative AI systems placed on the market before that date have until 2 December 2026 to add machine-readable marking to their output (new Article 111(4)), but that grace period is for providers only. If your staff make deepfake images, audio or video with a tool, your deployer duty to disclose that the content was artificially generated or manipulated has applied since 2 August 2026. Put that in the "everyone" session.
- Human oversight, Article 26(2). Deployers of high-risk systems must assign human oversight to people with "the necessary competence, training and authority" (Article 26 of the AI Act, unofficial consolidated text). Your overseer tier is the seed of that duty.
- Workers' representatives, Article 26(7). Before putting a high-risk system into use at the workplace, an employer must inform "workers' representatives and the affected workers". Under the omnibus, Annex III high-risk duties apply from 2 December 2027 and Annex I product-embedded duties from 2 August 2028.
A general-purpose drafting assistant is not a high-risk system, so Article 26 does not bite at Level 1. National employment law may already apply: in Germany, for example, the Works Constitution Act requires timely information on planned AI use (section 90) and usually co-determination for tools that could monitor performance (section 87(1) no. 6), as CMS Germany explains (18 July 2024). Check with employment counsel before rollout, whatever the AI Act timetable. Under the AI Act itself, the works council email is a preview of December 2027. Answer it now with a plan and a log, and you will have the habit before the duty arrives.
The Workshop: Build Your AI Literacy Plan and Training Log
You will design the program in one 2-hour session and deliver it over the following four weeks. The artifact you leave with is the AI Literacy Plan and Training Log: one document with a plan, a log and a one-paragraph rationale.
Who should be in the room?
Training goes wrong when its designers never meet the people who sit through it, so this room is weighted toward the floor:
- The head of HR or learning and development, who owns the program and the log.
- The AI governance owner (your CIO or CDO), who co-owns it and brings the register and the data rule.
- Legal or compliance, for the first 30 minutes only, to agree the reading of Article 4 and the wording of the rationale.
- The owner of the first use case from your Lesson 3 charter, who knows the real task and has the pilot outputs.
- A line manager from the first cohort, because managers will do the coaching.
- One frontline representative, because they will tell you which shift never sees an email.
- A works council or employee representative, where you have one, for the second hour. They asked the question; let them see the answer being built.
Bring four inputs: the Shadow AI Register, the Data-to-Tool Rule, the First Use Case Charter, and headcount by role with contractors included. If HR's headcount report leaves out agency workers, get the number from whoever signs the agency invoices.
The six steps
- Map every role to a tier, and every system to a role and a risk. Sort the headcount list into everyone, regular users and overseers. Include contractors and service providers acting on your behalf. The register tells you who already uses AI; the charter tells you who is about to. For each AI system, write one line: are you the provider or the deployer, and how risky is it? At Level 1 the answer is usually "deployer, general-purpose tool, not high-risk". Write it down anyway: it is one of the Commission's four minimum steps.
- Write three learning outcomes per tier, as things a person can do. "Understands AI risks" does not qualify. "Can say which data class a customer email is and which tool may receive it" does. If you cannot watch someone do it, rewrite it.
- Design the measures to the BCG signal, using real work. Plan at least five hours in total for regular users, with live practice and coaching and not only e-learning. Five hours is a design guide taken from a 2025 survey; the law sets no minimum. Build every exercise from the task in your charter, with data your Data-to-Tool Rule permits, and include one exercise where the tool is confidently wrong and the learner has to catch it. Ask the pilot users for an output that fooled them.
- Put leaders in the first cohort, visibly, and brief every manager. Directors sit in the same lab as the customer service agents. Every manager gets a 30-minute brief on coaching: what to ask, what good checking looks like, and where to send questions. This is your answer to the works council's third question.
- Deliver in order, and log on the day. The "everyone" session first, then regular users, then overseers, because people need the data rule before they touch the tool. Schedule by site and shift, not by email list. The trainer fills in the log the same day: name or role, employment type, tier, date, hours, format, material version. A log rebuilt from memory three months later will not survive a question.
- Spot check at 30 days, close the gaps, and write the rationale. Send five questions to a sample from every site: name the approved tool, name the data classes, spot the hallucination in a sample output, say who to tell when the tool is wrong, and say what must be disclosed about AI-generated images. Split the result by team and shift, book repeat sessions where it is weak, and log both the gap and the fix. Then write the one-paragraph rationale described below.
What do good learning outcomes look like?
A good learning outcome names something a person can do, in their own job, that you could watch them do. Here is a starting set you can edit in the session.
| Tier | After the training, a person can |
|---|---|
| Everyone | Name the approved tool and where to find it. Say which of the four data classes (Public, Internal, Sensitive, Prohibited) a piece of information belongs to and which tool may receive it. Explain in one sentence what a hallucination is and why output must be checked. |
| Regular users | Complete their own core task with the approved tool. Check a draft against the source system before sending it. Catch a confidently wrong output and report it. |
| Overseers | Decide when to reject a draft. Log a correction so the pattern can be seen. Escalate a repeated error to the use case owner. |
A filled-in example
Here is an illustration, not a real company. Take the freight forwarder we have followed since Lesson 1: 420 staff, plus 35 agency and contract workers who mostly work in the warehouse. Their first use case, chosen in Lesson 3, is drafting replies to customer shipment-status emails for a human to check and send.
In the design session, the frontline representative makes the first useful point: the 35 agency workers are not on the company email system, so any plan that starts with "we will send everyone a link" misses them completely. The plan changes to live sessions by site.
| Tier | Who | Count | Measure | Hours | Record kept |
|---|---|---|---|---|---|
| Everyone | All staff and contractors | 455 | Live 60-minute session by site, plus the one-page data rule | 1 | Attendance sheet, slide version |
| Regular users | Customer service, sales support, developers | 96 | Three 90-minute practice labs on their own tasks, plus a weekly coaching drop-in for a month | 5.5 in total | Lab roster, exercises completed |
| Overseers | Customer service team leads who check drafted replies | 8 | Review workshop: when to reject a draft, how to log a correction | 2 extra | Roster, review checklist |
| Leaders | Directors and heads | 14 | Seats in the first lab cohort, plus a sponsor briefing | 5.5 in total | Roster |
Leaders are not a fourth tier: they take the regular-user labs, plus a sponsor briefing.
Check the arithmetic on the regular users, because someone on your steering committee will. One hour in the "everyone" session plus three labs of 90 minutes each is 5.5 hours, which clears the five-hour design guide before you count the coaching drop-ins.
The "confidently wrong" exercise comes from the pilot. The tool had drafted a reply telling a customer that a consignment had cleared customs, when the tracking system showed it was still held. The draft read perfectly. In the lab, every learner gets that draft and the tracking record and is asked one question: would you send this?
Thirty days later, the spot check goes to a sample from every site. Result: 71% can name the approved tool and all four data classes. That is not a failure. It tells them where to look. When they split the result by team, the gap sits almost entirely in the warehouse night shift, which missed the live session because it was held at 10:00. A repeat is booked during their shift, and the log shows both the gap and the fix. The works council gets a one-page reply built from the plan, the log and the rationale.
The template: AI Literacy Plan and Training Log
Copy these fields into a spreadsheet with two tabs and a cover note. At Level 1, do not buy a system for this.
| Part | Fields | Who fills it in |
|---|---|---|
| Plan (one row per role) | Role; tier; provider or deployer; system risk level; learning outcomes; measure (format and hours); owner; schedule by site and shift | HR or learning lead, with the AI governance owner |
| Log (one row per person or role, per session) | Person or role; employment type (staff or contractor); tier; date; hours; format; material version; trainer; spot check result | The trainer, on the day |
| Rationale (one paragraph) | Why these measures fit these roles and these risks; date of the Commission Q&A version you relied on | AI governance owner, reviewed by legal or compliance |
Do not skip the rationale. It is the piece you would put in front of an authority or a works council. For the freight forwarder it might read:
We are a deployer of one general-purpose AI assistant, used mainly to draft customer shipment-status replies that a person checks before sending. It is not a high-risk system. All 455 staff and contractors received a live session on what AI is, its limits and our data rule. The 96 people who use the tool in daily work received hands-on labs on their own tasks with coaching. The 8 team leads who approve drafts received additional review training. Directors and heads attended the first lab cohort. We check retention at 30 days and repeat sessions where gaps appear. Records are held in the training log, owned by the head of learning.
That paragraph answers the three questions that matter: which roles, which measures, and why they are proportionate. One caution on wording: do not tell your board, or let a vendor tell you, that it makes you "Article 4 compliant" or "certified". No course can promise that.
Failure Modes and the Metrics to Track
Literacy programs rarely collapse in one go. They drift into one of six shapes, and each has a tell you can check in under an hour.
| Failure mode | How to spot it |
|---|---|
| Tick-box video | 100% completion and unchanged tool usage; staff cannot name the data classes |
| One size for all | Developers and warehouse staff took the identical module |
| No record | You cannot produce who was trained on what within a day |
| Frontline and contractors missed | The log contains only office staff on permanent contracts |
| Leaders exempt themselves | No director appears in the log |
| Stale wording | Material says the law requires you to "ensure" a "sufficient level" of literacy, or promises a "certificate of compliance" |
Why do AI literacy training programs fail?
Mostly because they teach the technology and skip the job. An hour on "how a large language model works" feels rigorous and changes nothing. Twenty minutes on "how to check a drafted reply against the tracking system before you send it" changes what happens at 09:05 tomorrow. If your course outline would make sense at any company in any industry, it is too general.
The second cause is treating the program as a compliance tick. The tell is the first row of the table: perfect completion alongside a flat usage report. Completion tells you who clicked through; you are after a change in how people work.
The last row deserves a quick audit before your next session. Anything written before July 2026 is likely to carry the old wording. Search your slides, your intranet page and your vendor's course for "ensure", "sufficient level" and "certificate", and correct all three.
Which metrics show the program is working?
Track five. Each comes from something you already have.
| Metric | How to measure it | What a bad reading looks like |
|---|---|---|
| Percentage of AI-using staff trained | People in the log divided by users in the Shadow AI Register | Users in the register who never appear in the log |
| Hours per person by tier | Training log | Regular users sitting at one hour |
| Percentage of managers trained | Training log | Teams trained while their manager was not |
| Share of staff who can name the approved tool and the data rules | Spot check at 30 days, split by site and shift | A low score concentrated in one site or shift |
| Existence and currency of the training log | Yes or no, plus the date of the last entry | Last entry older than your last new joiner |
Report these to the steering committee that approved the First Use Case Charter. Watch the first metric most closely: the register told you who uses AI, and the log tells you whether those same people have been trained.
Frequently Asked Questions
Is AI literacy training mandatory under the EU AI Act?
Yes, in the sense that providers and deployers must take measures to support the AI literacy of their staff and others using AI on their behalf; Article 4 has applied since 2 February 2025. Since Regulation (EU) 2026/1744 came into force on 27 July 2026, the law no longer asks you to ensure a sufficient level, and it does not require you to guarantee any individual's level. Training is the usual measure, but the duty is to act and document, not to hit a score.
Do employees need an AI literacy certificate?
No. The Commission's AI Literacy Q&A (last updated 20 October 2026 when checked) says there is no need for a certificate and that an internal record of trainings and other initiatives is enough. A dated, same-day training log is worth more to you than a wall of certificates.
How many hours of AI literacy training are required?
None are set by law: there is no minimum number of hours and no uniform test. As a design guide only, BCG's June 2025 survey found regular use was sharply higher among employees with at least five hours of training plus in-person coaching, which is why this lesson plans five hours or more for regular users.
Does the AI literacy requirement apply to contractors and agency workers?
Yes. The Commission's Q&A says the duty covers people dealing with the operation and use of AI on your behalf, which can include a contractor, a service provider or a client. In practice, add an "employment type" column to your log and check that agency and contract workers appear in it.
What happens if we ignore Article 4?
You are exposed to national enforcement: market surveillance authorities have supervised Article 4 since 2 August 2026 and, according to the Commission's Q&A, could impose penalties and other measures under national law. The Q&A gives no fine amounts, and no enforcement action on Article 4 had been found as of November 2026. The more immediate cost is practical: you cannot answer a works council, an auditor or a customer who asks what you did.
Key Takeaways
- AI literacy is the practical knowledge to use AI safely in your own job. Treat it like a driving licence, not an engineering degree: rules, supervised practice and a plan for when things go wrong.
- Article 4 has applied since 2 February 2025. Since Regulation (EU) 2026/1744 came into force on 27 July 2026, it is a duty to take measures to support the development of AI literacy, with no guaranteed level for any individual. Retire material that says "ensure", "sufficient level" or "certificate".
- National authorities have supervised Article 4 since 2 August 2026. Read the Commission's Q&A as it stands today (last updated 20 October 2026 when this lesson was checked), not a 2025 summary.
- Usage is outrunning guidance: BCG found white-collar frontline regular use rose from 51% in 2025 to 74% in June 2026. Tier the program by role and risk, and include contractors, night shifts and leaders.
- Design to the BCG signal of at least five hours with live coaching for regular users, built on real tasks, with one exercise where the tool is confidently wrong. Say plainly that five hours is a design guide, not a legal minimum.
- The record is part of the program. Log on the day, write the one-paragraph rationale, spot check at 30 days, and treat a works council's question as practice for the Article 26 duties that arrive with Annex III on 2 December 2027, while checking national works council law, which may apply already.
- You now hold the AI Literacy Plan and Training Log. Lesson 5 turns the rules your people have just learned into a short acceptable use policy with an owner, then uses your artifacts from Lessons 1 to 4 for an honest readiness assessment.
Skill.re