Back to Guides
Enterprise Software 15 min read

The Verification Framework: teaching staff to check AI output

The Verification Framework: teaching staff to check AI output

Two thirds of employees use AI output without evaluating its accuracy, and more than half have already made a work mistake because of it. Verification fails as a habit, not as an idea. What works is naming the exact check for each kind of output, tying the standard of proof to what the output touches, and drilling it until the check costs almost nothing.

Every training deck in this category ends on the same slide: always verify AI output. The instruction is correct and changes nothing, because it describes an attitude rather than an act. Nobody skips a check they can name in four words and finish in ninety seconds. People skip the check that was never defined and has to be invented under a deadline.

Why people who know better skip the check

The best numbers come from Trust, attitudes and use of artificial intelligence: A global study 2025, led by Professor Nicole Gillespie and Dr Steve Lockey of Melbourne Business School with KPMG. It surveyed more than 48,000 people across 47 countries between November 2024 and January 2025.

Three findings matter. Sixty-six percent of employees rely on AI output without evaluating its accuracy, 56% say they have made mistakes in their work because of AI, and 57% hide their AI use or present AI-generated work as their own.

Read together, those change the diagnosis. A workforce that conceals its AI use at that rate is not short of information about hallucination. It is short of a defined procedure and a reason to run it in the open.

The mechanism has been studied. In a CHI 2025 paper, Hao-Ping Lee and colleagues at Carnegie Mellon and Microsoft Research surveyed 319 knowledge workers who gave 936 first-hand examples of using AI at work. The clearest signal was confidence: more faith in the tool went with less critical thinking, while more faith in one’s own ability on the task went with more. Confidence in the output arrives at once; the ability to check it has to be built, and almost nobody builds it on purpose.

NIST’s Generative AI Profile (AI 600-1, July 2024) calls the failure mode confabulation, and the word earns its place in a training room. You are checking for content that is confident, well formed and false. It does not look like an error. It looks like finished work.

What verify has to mean before it can be taught

The US Department of Labor’s AI Literacy Framework, released by the Employment and Training Administration on 13 February 2026, is the clearest public statement of the target. Its five foundational content areas are: understand AI principles, explore AI uses, direct AI effectively, evaluate AI outputs, and use AI responsibly. It creates no legal duty, and it is the vocabulary to borrow when a board asks why checking is a named competency. Most AI security training for employees stops at what not to paste, and most AI governance training stops at the policy. Checking is the half left out.

The framework stops where a curriculum starts. “Assess whether outputs are accurate, complete and appropriate” is a competency statement, not a lesson plan. Four things turn it into one, and no national framework can supply them, because none can know your workflows:

  1. An act. What the person does, against which artifact.
  2. A stopping rule. What makes the check finished, so it has an end and a cost.
  3. A trigger. Which outputs get which standard of proof, since no team can check everything.
  4. An owner. Who is answerable when the output is wrong.

The fourth does the most work. When nobody names the owner, the implied answer is the tool, and a tool cannot be held to anything. The line that settles it is blunt: the person who sends it owns it. Put that in the written policy and most of the behavior follows, the wider argument in our case that policy and training are two halves of one control.

Five checks, named by output type

Generic advice fails because the act differs by output type. These five cover most of what a non-technical team produces with an assistant.

Output The named check What it catches Cost
A figure Trace the number to its source system, then re-derive it once A total including a category it should not, a right number for the wrong period 2 to 5 min
A citation Open it. Confirm it exists, the author and date match, and the claim is in it Invented sources, real sources saying something adjacent, stale editions 1 to 3 min
A quoted clause Match the quote against the original, negations and carve-outs included A dropped “not”, two clauses merged, a paraphrase passed off as a quote 3 to 8 min
A long summary Read in full the three passages the summary compresses or omits Omission rather than error: the exception, the deadline, the condition 5 to 15 min
Reasoning or code Run it on one case whose answer you know, then one you expect to break Plausible steps that do not compose, ease on a case it never saw 5 to 20 min

Two things matter more than the contents. Every row carries a time cost, because a check without a cost is a slogan and a check with one can be scheduled. And every check is an action with an end state: “open it” finishes, “be critical” never does. Build your own from the team’s last month of AI-assisted work, keep it to one page, and put it where the work happens.

Tiering by consequence

The hidden reason this advice gets ignored is arithmetic. Applied to everything, those checks cost more than they save, so a busy team applies them to nothing. Tier by what the output touches and the budget reappears.

Tier What it covers Standard of proof
Leaves the building Client deliverables, filings, public copy, numbers going to a lender or auditor Every figure traced, every citation opened, every quote matched, plus a second reader. No deadline exception
Informs a decision Internal analysis, board material, a recommendation someone will act on The named check on each figure and source, by the author, with sources left attached
Internal and reversible Drafting, brainstorms, meeting notes, reformatting the author already knows An author read, unless it is promoted to a higher tier

Most teams have this inverted without noticing. Reversible work draws anxious over-review because it is visible early, and the client deliverable gets a fast read at 6pm on Thursday. Writing the tiers down surfaces that, and the sorting argument is itself the lesson.

One rule keeps it honest: anything promoted to a higher tier is re-checked there. A paragraph drafted in a brainstorm and later pasted into a client memo has never been checked, and the person pasting it rarely wrote it.

The three conditions that decide whether it survives

Three conditions decide whether the habit holds past the second week, and all three sit with management, not the learner.

Time inside the deadline. If the deadline assumes AI made the task instant, the check has nowhere to live. The honest planning line: the tool compresses drafting and adds a review step, so the net gain is real and smaller than the demo suggested.

Access to the source of truth. An analyst who cannot open the system of record cannot trace a figure. A paralegal without the contract cannot match a clause. Checking failures are often access failures wearing a discipline costume, and the fix is a permission.

No penalty for slowing down. If the first person to say “I need an hour to check this” becomes the bottleneck in the retelling, nobody says it twice. The counter-signal has to come from a manager, in public, at least once. Pair it with an escalation route people can use without looking slow, which is also how checking stays separate from monitoring, a tension in our argument for enforcing an AI policy without surveillance.

The ninety-minute drill

What transfers is practice on real material. This fits ninety minutes, and an internal lead can run it after seeing it once.

  1. The planted-error document (25 min). Take a real deliverable from the team’s own work, have an assistant rewrite it, then plant three errors by hand: a wrong figure, an invented source, a dropped negation in a quoted clause. Everyone hunts alone, then compares. The misses teach more than the catches.
  2. Sort the outputs (20 min). Eight AI outputs on screen, labeled by the group as correct, incomplete (true as far as it goes), or misleading (confident, well written, wrong). Then accept, correct, or escalate, with a reason.
  3. Name your own checks (20 min). Each function writes the named check for its two highest-volume AI-assisted outputs, in the format above, with the time cost. This block produces the one-page sheet the team keeps.
  4. Tier the work (15 min). Sort this week’s real deliverables into the three tiers. Disagreements get logged as rulings rather than settled by the most senior voice.
  5. One live escalation (10 min). Someone raises a genuinely uncertain case through the real route while everyone watches. A path walked once gets walked again.

Two notes from running these. Show the tool failing on the team’s own material, because a session where the assistant only succeeds teaches trust. And log the rulings the room produces: within a month or two that log covers nearly every case that comes up, and it teaches better than the policy, because it is made of real decisions.

When the check is genuinely impossible

Some failures are not habit failures. An assistant that answers with no reachable source makes tracing impossible, and asking staff to check what the tool will not show them trains cynicism.

Three responses, in order of preference. For the top two tiers, prefer tools and modes that attach sources to claims. Where that is unavailable, make the person produce the source themselves, which turns the assistant into a drafting aid rather than a research tool. Where neither is possible, the task sits outside sanctioned use and belongs in the record of things you decided not to do, the same discipline as keeping an AI risk register a small company can actually maintain.

Measuring the habit at sixty days

Completion rates measure attendance. These four signals measure behavior, and one person can collect them in an afternoon.

  • Re-run the planted-error document at sixty days with fresh errors and score the catch rate. It is the only direct measure of capability you will get, and the first run is the baseline you get one chance at.
  • Count voluntary near-miss disclosures. Expect this to rise for a quarter if the no-penalty condition is real. Zero across hundreds of licenses is a reporting failure, not a safety record.
  • Count escalations through the route. A route with no traffic does not exist, whatever the policy says.
  • Audit ten shipped top-tier deliverables for attached sources, which stand in for the check having happened and are visible without reading anyone’s prompts.

Whether any of this should be compulsory is a separate decision, argued both ways in the case for and against mandatory AI training.

What this costs to run

Run internally, the cost is small: a morning to write the sheet and the tiers, ninety minutes per group for the drill, and an hour each quarter to refresh the planted errors. Add a review of the sheet against any tool that changed under you. The real cost is management attention, which is why this stalls on whoever owns it as a second job.

Brought in, checking is a module inside hands-on AI enablement work rather than a product of its own. Facilitated LLM fluency sessions in the US market run roughly $2,000 to $15,000 by headcount, sittings, and how much is built against your own documents. A packaged module costs far less and buys a completion record rather than changed judgment. Catalog courses cover the five DOL areas at scale, a field we reviewed in our comparison of literacy programs for whole workforces.

This is the clearest illustration of why license spend and capability are separate lines on one budget, the argument in our case that buying licenses is not the same as building a team that can use them. For a read on where your team sits, we run a free AI-readiness call, and the planted-error exercise is usually the fastest diagnostic in it.

Frequently asked questions

What are AI verification habits?

The specific, repeatable checks a person runs on AI output before using it, defined per output type with a known time cost. A habit names the act: trace the figure, open the citation, match the quote. Unlike general caution it has a trigger, a procedure and a stopping point, so it can be taught and measured.

How do you teach staff to check AI output?

Through practice on their own material. The core exercise is a real deliverable carrying three planted errors (a wrong figure, an invented source, a dropped negation) that people hunt and then discuss. Follow with a sorting exercise on real outputs, have each function write its own named checks, and close with one live escalation.

Why do employees skip checking when they know AI makes mistakes?

Because confidence in the output arrives faster than the ability to check it, and because the check was never defined. The CHI 2025 study by Lee and colleagues found more faith in the tool tracked with less critical thinking. Time pressure, no source access, and a culture that reads checking as slowness finish the picture.

Does an AI use policy need a verification clause?

Yes, and specific enough to act on. “Employees must review AI output for accuracy” is unenforceable and unteachable. A clause naming the tiers, stating that the sender owns the output, and pointing to the one-page sheet gives a manager something to coach against.

What does the Department of Labor framework say about checking AI output?

Evaluating AI outputs is one of five foundational content areas in the AI Literacy Framework released on 13 February 2026 by the Employment and Training Administration. It asks that workers judge whether outputs are accurate, complete and appropriate rather than treating AI as a final authority, and leaves the curriculum to the employer.

How long does checking AI output actually take?

One to twenty minutes per item, by type. Opening a citation takes a minute or two, tracing a figure to its source system two to five, matching a quoted clause three to eight, and testing a chain of reasoning against a known case up to twenty. Publishing those numbers makes the habit plannable.

How do we measure whether this training worked?

Re-run the planted-error exercise at sixty days with fresh errors and compare catch rates with the first run. Then count voluntary near-miss disclosures, escalations through the route, and the share of top-tier deliverables shipping with sources attached. Course completion measures attendance.

Is this training required by law?

Not in the United States. The DOL framework is voluntary, and NIST’s Generative AI Profile is a voluntary risk framework. Article 4 of the EU AI Act, applicable since 2 February 2025, does oblige providers and deployers to take measures toward sufficient AI literacy among staff, so US companies with EU operations should treat it as in scope. Client contracts and insurer questionnaires are arriving faster than legislation.

Key takeaways

  • Two thirds of employees use AI output without evaluating accuracy and 57% conceal their use, so this is a habit problem, not an awareness problem.
  • “Always verify” is an attitude. A teachable check is an act with a trigger, a procedure, a stopping point and a time cost, written per output type.
  • Tier by consequence or the habit is unaffordable: full tracing for what leaves the building, named checks for decision material, an author read for reversible work.
  • Three conditions belong to management: time inside the deadline, access to the source of truth, no penalty for slowing down.
  • Measure the catch rate at sixty days, not course completions, and say out loud who owns the output when it is wrong.

Last Updated: Nov 16, 2026

DJ

Dirk Jan van Veen, PhD

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles