AI Training for Employees: What to Teach and in What Order

AI training for employees is the instruction that turns access to AI tools into competent use: what the tools can and cannot do, what information may go into them, how to check the output, and how to use them for the specific tasks of a role. As of Gallup's May 2026 reading, 52% of US employees use AI at work at least a few times a year and 15% daily, but only 25% say their organization has communicated a clear plan for using it, and employees who received training were 1.7 times as likely to use AI frequently as those who did not. Access has arrived ahead of instruction.

The training content on the market is mostly prompt-writing courses, which teach the part of the skill that matters least. The field experiments with real workers show that gains depend on knowing which tasks the tool is good at and which it is not, and that a short overview does not protect people from trusting a wrong answer. This guide sets out a four-module program for all staff, a module for managers and a module for HR, each tied to the evidence, with a free plan template in Word, Excel and PDF.

Download the plan template

The training gap in numbers

The shape is consistent across sources: half of employees use the tools, a quarter to two fifths have been trained, and the training that exists is more often informal than designed. Our AI in the workplace statistics page tracks the adoption series quarter by quarter.

What the experiments say training must cover

Four field experiments with real workers tell you what to teach. None of them is about prompt wording.

• Knowing the boundary. Dell'Acqua and colleagues' experiment with 758 Boston Consulting Group consultants, published September 21, 2023, gave GPT-4 to consultants on two kinds of task. On tasks inside the model's capability, users completed 12.2% more tasks, 25.1% faster, at 40% higher quality. On a task designed to sit just outside the model's capability, users were 19 percentage points less likely to reach the right answer than consultants without AI. A group given a short overview of how to use the tool did no better on the outside task, and in some measures worse. The lesson is that the skill to teach is judging which tasks the model is reliable for, and that a thirty-minute overview does not transfer it.
• Novices gain most. Brynjolfsson, Li and Raymond's study of 5,179 customer support agents, published as an NBER working paper in April 2023 and in the Quarterly Journal of Economics in 2025, found an AI assistant raised issues resolved per hour by 14% on average, 34% for the least experienced agents and roughly nothing for the most experienced. Training effort should be weighted toward newer staff, and experienced staff should be told the gain for them is smaller so they do not abandon their own judgment.
• Individuals with AI match teams without it. Dell'Acqua and colleagues' experiment with 776 Procter and Gamble professionals, published as NBER working paper 33641 in March 2025, found individuals using AI produced work comparable in quality to two-person teams without it, and that AI narrowed the gap between people with and without technical expertise. It tested access, not training, so it says what the tools can do for a prepared workforce rather than how to prepare one.
• Output is not the same as outcome. Workday's January 14, 2026 survey of knowledge workers found 85% saying they saved one to seven hours a week with AI, and about 40% of that saving lost to checking and reworking AI output, leaving 14% with a net positive result. It is a self-report survey and we have it from Workday's summary, but it names the failure mode that training must address: unchecked output creates work for someone else.

Module one to four: the core program for all staff

Four sessions of 45 to 60 minutes, in this order, each with a hands-on exercise in the company's approved tool. The Word template has a session plan for each.

Module 1: What the tool is and what it is not

How a language model produces text, why it can be fluent and wrong at the same time, what hallucination looks like in practice, and the difference between the consumer product and the company plan. Exercise: ask the approved tool three questions whose answers the participant knows, one of which is designed to produce a confident wrong answer, and discuss how to tell.

Module 2: The data rules

The company's data tiers, which tools are approved for which tier, what the vendor does with inputs on the company plan, and why removing a name does not make personal data public. This module is the AI use policy taught rather than signed. Exercise: sort twenty realistic snippets of company information into tiers and decide which could go into the approved tool.

Module 3: The boundary for your role

The BCG finding applied to each team. For a role, list the recurring tasks, try the tool on each, and classify them: reliable (summarizing a document the person has read, first drafts, reformatting, code scaffolding), reliable with checking (research, analysis, anything with numbers, anything with citations), and not for this tool (final legal or financial judgment, anything about an identifiable employee or customer, anything the person cannot verify). The output of the session is a one-page task map for the team that goes in the shared workspace.

Module 4: Verifying and owning the output

How to check a figure, a citation, a quotation and a piece of code; what the review step costs and why skipping it moves the cost to a colleague; and the rule that the person who sends it owns it. Exercise: each participant takes an AI-produced document with three planted errors and finds them.

The manager module

Gallup's finding that employees whose managers actively support AI use are far more likely to use it well, and the CHRO finding that half of HR leaders are not confident their managers can lead with AI, make the manager session the highest-return hour in the program. It covers: how to set the task map with the team rather than for it; what the manager may and may not do with AI in performance reviews and people decisions, which our AI in performance reviews guide covers in full; how to talk about AI and job security without either dismissing the worry or feeding it; and how to recognize when a report is handing in unchecked output. Resume Builder's June 2025 survey found two thirds of AI-using managers had informal or no training on using it in people decisions, which is the gap this module closes.

The HR module

HR has the one set of tasks where AI use is regulated. The module covers the state and local rules on AI in hiring, promotion, discipline and dismissal, which decisions require notice and audits, what the Colorado law will require from January 1, 2027, and the data rules for employee and candidate information. It then works through the approved uses with exercises: drafting a job description, summarizing exit interview themes, and drafting a policy, each using our ChatGPT prompts for HR with their data and legal markers. It ends with the vendor questions from our AI resume screening guide, because HR will be asked to approve tools that screen people.

The legal reasons to train

No US law requires general AI training for employees. Three things make it a compliance matter anyway. First, the state AI hiring laws in Illinois, California, New York City, Colorado and Connecticut place duties on the people who use the tools, and a duty nobody has been told about will be breached. Second, Colorado's definition of meaningful human review, which applies from January 1, 2027, requires a reviewer who considers primary evidence and does not default to the system output; a reviewer who has not been trained to do that cannot be shown to have done it. Third, for employers with staff in the European Union, Article 4 of the EU AI Act has required providers and deployers to ensure "a sufficient level of AI literacy" among staff operating AI systems since February 2, 2025. Our AI hiring laws by state table has the US duties with dates.

How to measure whether it worked

1. Boundary accuracy. Before and after module 3, give participants ten tasks and ask which the tool is reliable for. The BCG study says this is the skill; measure it directly.
2. Error detection. The module 4 exercise, scored. Repeat at six months.
3. Data incidents. Count pastes of restricted data blocked or flagged by the data loss prevention tool per hundred users, before and after. It should fall.
4. Frequent use. Gallup's 1.7 times figure is the benchmark: the share of trained staff using the approved tool weekly or more should pull away from the untrained share.
5. Rework. Ask reviewers and managers, quarterly, how often they received AI output that had to be redone. Workday's 40% is the baseline to beat.
6. Who was trained. The Excel tracker records completion by team, by tenure band and by role. Brynjolfsson's result says the newest staff should be first, and the tracker will show whether they were.

Our employee training statistics page has the benchmarks for training hours and spend against which to set the budget.

Key takeaways

• 52% of US employees use AI at work and 15% daily, but 25% say their organization has a clear AI plan. Trained employees are 1.7 times as likely to be frequent users (Gallup, May 2026).
• Half of CHROs are not confident their managers can lead teams using AI; 57% say they train managers to (Gallup, 102 CHROs, early 2026).
• The BCG experiment with 758 consultants found 40% higher quality inside the model's capability and 19 points worse accuracy just outside it, and a short overview did not protect people. The skill to teach is judging the boundary.
• Novices gain most: a 14% average productivity gain among 5,179 support agents was 34% for the least experienced and near zero for the most experienced.
• About 40% of self-reported time savings are lost to checking and reworking AI output (Workday, January 2026, self-report), which makes verification the fourth core module.
• The core program is four sessions: what the tool is, the data rules, the boundary for your role, and verifying and owning output. Managers and HR each get a further module.
• No US law requires general AI training, but state AI hiring laws place duties on users, Colorado's human review standard requires a trained reviewer from January 1, 2027, and the EU AI Act's Article 4 has required AI literacy since February 2, 2025.
• Measure boundary accuracy, error detection, data incidents, frequent use, rework and who was trained, using the free Excel tracker.