AI training for employees: teach verification, not tools

The company ran AI training in March: which buttons to press, which model picks which task, a prompt cheat sheet. By June the vendor had redesigned the interface, renamed the models, and shipped an agent mode the training never mentioned. The cheat sheet is now a historical document, and the employees are back to guessing.

The problem was the curriculum, not the effort. Feature training depreciates like the software it describes. What does not depreciate is judgment: knowing when an answer needs checking, what a refusal means, and who owns the output. That is what the KPMG numbers say is missing, and it is teachable in hours.

A verification-first curriculum has five units, each one habit.

1. Check an answer: open the source, compare the claim to the passage, note what it does not prove.

2. Read a refusal: an honest "the sources do not contain this" is protection, not failure; distinguish it from a broken one.

3. Shape the question: scope the sources, demand per-claim citations, permit the honest gap, set the cutoff, fix the output shape.

4. Keep authorship: AI proposes, the human approves; nothing consequential applies silently.

5. Verify in your domain: the extra checks your field demands, like jurisdiction and currency in law, or methodology in research.

Each unit is a published guide on this blog, usable as pre-reading. None of them mentions a specific vendor, which is exactly why they will survive the next release cycle.

Unit one converts skeptics fastest, so start there. Bring three AI answers from your team's real work: one correct, one subtly wrong, one fabricated. Have everyone run the 90-second check on all three without saying which is which. The room finds the fabrication together, and the subtle error usually splits the room, which teaches the deeper lesson: fluency tells you nothing.

Close the hour by agreeing on the verification boundary from the team flow: nothing consequential travels unverified. One hour, one habit, and the 58% statistic now has a local answer.

Training that ends at attendance sheets changes nothing. Measure the behavior the training exists to create: in review, ask "which claims did you verify?" as routinely as "is it done?". Celebrate catches out loud, because every AI error found before shipping is the program working. And watch the one number that matters quarterly: how often AI-supplied claims that shipped turn out wrong. KPMG found 57% of employees have already made AI-caused mistakes; your program's success is that number falling locally.

None of this requires policing. It requires the check being cheap and the culture treating catches as wins.

Skills survive where practicing them is effortless. If your team's AI work happens in tools with unanchored output, every skill above fights friction daily. In a grounded tool, the curriculum becomes the interface: Tatsulok answers only from your documents and curated collections, links every claim to its exact passage (unit one is one click), says plainly when sources are silent (unit two is built in), and keeps suggestions as proposals until accepted (unit four is the default).

This week's drill for trainers, no signup needed: before designing anything, survey your team with one question: "describe the last time you checked an AI answer against its source." The blank stares are your baseline, and the handful of good answers are your future co-instructors.