An auditor sits across the table from the head of learning and asks a single question: "Show me your workforce is AI-literate." The head of learning opens the LMS and points at a number: 96 percent completion on the AI-literacy course, 3,840 of 4,000 employees. The auditor does not write it down. Instead she asks: "Which of these people operate the AI tool that screens job applicants, and what can you show me that they can catch it when it ranks the wrong person to the bottom?" The room goes quiet. The completion number was real, and it answered the wrong question. This lesson is about the evidence that answers the right one.

The Gap Between Delivered and Demonstrated

The single most important distinction in evidencing literacy is the line between training delivered and competence demonstrated. Training delivered means a course was assigned and completed: a record exists that a person sat through content. Competence demonstrated means there is evidence the person can actually do the thing the training was about: catch a hallucination, override a biased ranking, ask the right question before trusting an AI recommendation. Why you care: an auditor checking an AI-literacy program is not impressed by the first and is looking for the second, especially for the roles where AI use carries real consequence. A completion record proves attendance. It does not prove capability, and capability is what the duty is reaching for.

This maps directly onto a framework most learning professionals already know. In the Kirkpatrick four levels of evaluation, Level 1 is reaction (did they like it), Level 2 is learning (did they acquire the knowledge), Level 3 is behavior (do they do it differently on the job), and Level 4 is results (did it change an outcome). A bare completion record is barely even Level 1. The evidence an auditor wants lives at Level 2 and Level 3: proof people actually learned the capability, and ideally proof they apply it in their work. The whole craft of evidencing literacy is climbing from "they completed it" toward "we can show they can do it," with the climb steepest and most necessary for high-risk roles.

Why does an auditor care so much about this particular gap, when in plenty of other training contexts a completion record is accepted without complaint? Because the whole point of the AI-literacy duty is a capability, not an exposure. The law does not want people to have heard about AI risks; it wants the person overseeing a high-risk system to actually be able to catch it when it fails. A completion record is evidence that something was shown to someone. It is silent on whether the showing worked. In a low-stakes context, that silence is tolerable. In the context the auditor is examining, where a wrong AI output can harm a person's rights or opportunities, the silence is the whole problem, because the one thing the duty exists to guarantee is exactly the thing a completion record cannot speak to. The auditor is not being pedantic. The auditor is reading the duty correctly.

It is worth being precise about which duty the auditor is reading, because the text is a moving target. Article 4 in force requires employers to ensure a sufficient level of AI literacy, in application since 2 February 2025, with enforcement beginning 2 August 2026. The Digital Omnibus (proposed 19 November 2025, endorsed by the European Parliament 16 June 2026, and not yet published in the Official Journal, so not yet law) would soften that general verb to promote and encourage, while leaving untouched the separate duty to train the operators of high-risk AI systems for human oversight. That is why an auditor's sharpest questions land on the high-risk operators: the competence evidence for that group is the part of the duty that stays load-bearing whichever way the amendment resolves, so it is the evidence you build most rigorously.

A completion record proves someone was present. An auditor is asking whether they are competent. Those are different claims, and only one of them is a defense.

The Three Questions an Auditor Actually Asks

Auditors do not ask "is your workforce AI-literate" as one vague question. They decompose it, and if you know the decomposition you can build evidence for each part in advance. There are three questions underneath.

Coverage: Did the Right People Get the Right Training

Coverage is the question of who was trained, at what depth, relative to who should have been. The auditor wants to see that the people who operate high-risk AI systems received the deep operator training, not just the baseline course, and that nobody who uses AI fell through the gap entirely. Evidence for coverage is the role-to-tier map from your program design, joined to completion data: here are the roles, here is the tier each was assigned, here is who completed which tier, and here are the exceptions and how we are closing them. Coverage is where a flat program collapses, because it can show everyone got the same course but cannot show the high-risk operators got the deeper one they needed.

Records: Can You Produce the Evidence on Demand

Records is the question of whether your evidence actually exists and can be retrieved, dated, and attributed. The auditor wants a record that says who completed what, when, and tied to which role and which AI system, retrievable in one lookup rather than reconstructed from email threads. A record that cannot be produced on demand is, for audit purposes, a record that does not exist. This is unglamorous and decisive: many programs fail not because the training was bad but because nobody can quickly prove who got it. The discipline is to instrument the program so the proof is a query, not an archaeology project.

Competence: Can You Show They Can Actually Do It

Competence is the hardest and most valued question, and the one a completion record cannot touch: can you show the person can actually perform the capability, especially for high-risk operators. Evidence for competence is not "they watched the override module." It is a record that they demonstrated the override, ideally through a scenario or assessment in which they caught a wrong AI output and documented the human decision. This is the Level 3 evidence, and it is the difference between a program that taught oversight and a program that can prove oversight happens. For the high-risk roles, competence evidence is the thing the auditor most wants and the thing a panic build most lacks.

There is a hierarchy inside competence evidence worth understanding, because not all demonstrations are equal. The weakest form is a knowledge check: a quiz that asks the operator to identify, in the abstract, that an AI ranking could be biased. Better is a scenario: a constructed exercise where the operator is shown a skewed ranking and has to catch it, override it, and record their reasoning, captured and graded. Stronger still is evidence from real work: a log entry where the operator, on an actual candidate pool, noticed something off, overrode the tool, and documented why. The progression runs from "they know it could happen" to "they did it in a drill" to "they do it on the job." An auditor will accept the scenario as solid evidence and treasure the real-work log, because the real-work log is the only one that proves the capability survived contact with reality rather than living only in a classroom.

The Evidence an Auditor Accepts

Translate the three questions into the artifacts that actually satisfy them. The table below is the evidence pack worth assembling before anyone asks, mapped to what each piece proves and its weakness if it stands alone.

Read the table as a ladder. The top rows are easy to produce and weak on their own; the bottom rows are harder to produce and far stronger. A defensible evidence pack does not rely on any single row. It pairs coverage (the map plus completions) with competence (scenarios and real decision logs) and the design rationale that explains why each role got what it got. The auditor is not looking for a pile of certificates. She is looking for a coherent story: here is who needed what, here is the proof they got it, and here is the proof the high-risk operators can actually do it.

Instrumenting the Program to Produce Evidence

The mistake that sinks most programs is treating evidence as something you gather after the auditor asks. By then it is too late, because the demonstrations were never captured and the records were never tied to roles. The discipline is to instrument the program so it produces its own evidence as a by-product of running. Build the role-to-tier map as a living document, not a one-time slide. Record completions tied to role and AI system, not as anonymous course-level counts. Design the high-risk operator training to include a graded scenario whose result is captured, so competence is recorded the moment it is demonstrated. Where the work allows it, have operators keep a human-decision log, a short record of when they reviewed or overrode an AI output, which is the strongest evidence of all because it shows oversight happening in real work, not just in training.

Two cautions keep the evidence honest, and an auditor will probe both. First, evidence must be attributable and dated. "We trained everyone last year" is not evidence; "this named operator completed the override scenario on this date and logged these three real interventions since" is. Second, do not manufacture competence evidence you do not have. If you only delivered training and never assessed competence, say so and show your plan to close the gap, rather than dressing a completion record up as a demonstration. An auditor trusts a program that knows the difference between delivered and demonstrated far more than one that blurs it, and blurring it is exactly the kind of confident, unverifiable claim the iron rule of this program exists to prevent.

There is a discipline of evidence freshness that catches even well-run programs off guard. Competence is not a permanent state; it decays, and tools change underneath it. An operator who demonstrated a flawless override eighteen months ago, on a version of the screening tool that has since been retrained, holds evidence that is technically real and practically stale. An alert auditor will ask not just "did they demonstrate competence" but "when, and on which version of the system." This is why the human-decision log is so valuable beyond its strength as a one-time proof: a steady trickle of recent log entries shows the capability is live, current, and exercised against the system as it actually is today. Build the program so competence evidence renews rather than ossifies, through periodic scenarios and ongoing logs, and you answer the freshness question before it is asked. A pile of two-year-old certificates is not a current capability, and an auditor knows it.

Finally, resist the temptation to over-instrument the baseline tier in pursuit of evidence. The same proportionality that governs training depth governs evidence depth. For the analyst who uses AI to draft a summary they will read and check, a completion plus a short knowledge check is proportionate evidence; demanding a graded scenario and a decision log from every casual user would drown the program in low-value records and exhaust the goodwill you need for the operator tier. Spend your evidence-collection effort where the consequence is, exactly as you spend your training effort there. The goal is not maximum evidence everywhere; it is sufficient evidence scaled to risk, which is the same shape as the literacy program itself and, not coincidentally, the same shape the law asks for.

Instrument the program so the evidence is a by-product of running it. If you are gathering proof only after the auditor asks, you are already too late.

A Worked Example: The Screening Tool Question

Return to the auditor's question about the people who operate the applicant-screening tool, and watch two programs answer it.

Before (delivered only). The program ran a company-wide AI course and recorded 96 percent completion. When the auditor asks who operates the screening tool and what shows they can catch it when it is wrong, the head of learning can produce a list of who completed the general course but cannot isolate the screening-tool operators, cannot show any of them learned the bias failure mode specific to ranking people, and has no record of any of them ever demonstrating an override. The 96 percent is real and useless for this question. The program delivered training. It cannot demonstrate competence for the one role the auditor cares about, and the audit note writes itself: literacy program present, but no evidence of competence for high-risk-system operators.

After (delivered and demonstrated). The same question lands differently. The head of learning pulls the role-to-tier map and isolates the eleven screening-tool operators in seconds. For each, she shows the operator-tier completion, the graded scenario in which they caught a skewed ranking and overrode it, the date they did it, and the human-decision log showing the real overrides each has made since. She also shows the program design rationale explaining why these eleven were tiered as operators while the rest of recruiting was baseline. The auditor writes it down. Coverage: the right eleven got the deep training. Records: produced in one lookup, dated and attributed. Competence: demonstrated through scenarios and confirmed in real decision logs. Same workforce, same tool, a completely different audit, because the program was instrumented to produce evidence of competence, not just proof of attendance.

It is worth noticing what the after example cost and what it did not. It did not cost more training; the recruiter in the before version also "took the AI course." What it cost was a small amount of design foresight: tying records to role and tool, building one graded scenario into the operator path, and asking operators to keep a short log. None of that is expensive, and all of it pays out the moment an auditor asks a pointed question. The before program spent the same training budget and got a record that collapses under the first real question. The after program spent a little design attention up front and got a record that answers the question crisply. The difference between the two is not money or even effort; it is whether the program was designed, from the start, to be asked about. A program built to be questioned answers calmly. A program built only to be completed goes quiet at exactly the wrong moment.

The lesson is not that completion records are worthless; coverage is real and they are part of it. The lesson is that completion alone answers a question the auditor is not asking. The question is competence, the evidence is demonstration, and the program that survives the audit is the one built to produce that evidence before anyone asks for it.

Key Takeaways

• The decisive line is between training delivered (a completion record proving attendance) and competence demonstrated (evidence the person can actually do the thing), and an auditor is looking for the second, especially for high-risk roles.
• In Kirkpatrick terms, a bare completion is barely Level 1; the evidence that matters lives at Level 2 (learning) and Level 3 (behavior), so the craft is climbing from "they completed it" to "we can show they can do it."
• Auditors decompose "is your workforce AI-literate" into three questions: coverage (did the right people get the right depth), records (can you produce dated, attributed evidence on demand), and competence (can you show they can actually do it).
• Coverage evidence is the role-to-tier map joined to completion data; a flat program fails coverage because it cannot show the high-risk operators got the deeper training they needed.
• A record that cannot be produced on demand is, for audit purposes, a record that does not exist; instrument the program so the proof is a query, not an archaeology project.
• Competence evidence is demonstration, not attendance: a graded override scenario and, strongest of all, a human-decision log showing oversight happening in real work.
• Instrument the program to produce its own evidence as a by-product of running it; do not wait until the auditor asks, and never dress a completion record up as a demonstration you do not have.
• Evidence must be attributable and dated, and an honest program that names the gap between delivered and demonstrated is more defensible than one that blurs it, which is the iron rule applied to your own records.