Explainer Human Resources & Organizational Development · Learning and Development · Issue 2026-W46
AI training for HR: why buying the tool is only the beginning
CIPD’s new HR Pulse report raises a practical question about preparing people teams to use artificial intelligence. Reading its findings requires distinguishing two samples within the report, then separating faster task completion from better decisions. We propose training around a specific task, a reviewable comparison and clear limits on delegating decisions that affect employees.
CIPD’s new HR Pulse report raises a practical question about preparing people teams to use artificial intelligence. Reading its findings requires distinguishing two samples within the report, then separating faster task completion from better decisions. We propose training around a specific task, a reviewable comparison and clear limits on delegating decisions that affect employees.
An attention-grabbing number needs the right denominator
An HR director hearing that three quarters of people professionals have received no training to apply AI in their jobs might immediately commission a general course. Before deciding, the director should ask who supplied that answer and when. CIPD released HR Pulse on 26 October 2026 with support from Talking Talent. The report combines its main survey with additional data from a separate survey. Its percentages should not all be attributed to the same respondents or collection period. [1]
The main survey included 883 UK people professionals and collected responses in May–June 2026. The 75% training figure came from a 353-person subgroup of the Good Work Index. The AI chapter’s footnote and the methodology section describe that subgroup’s collection period differently; we therefore identify it here as winter 2025–2026 and record the discrepancy. These are self-reported UK data, not an assessment of your team’s competence or causal evidence of training effects. [2]
Separate three problems before commissioning a programme
We suggest that a learning and development lead begin with a conversation about an actual task, such as drafting a response to a general administrative enquiry. First, does the employee know whether the tool may be used for that task? Second, can the employee produce a useful draft? Third, can the employee detect an error or exception within it? These are different capabilities. Someone may operate a tool quickly without having permission to enter the information or the expertise to review the result.
This distinction translates “we need AI skills” into needs that can be examined separately. The missing element might be an acceptable-use policy, professional knowledge or verification practice, rather than prompting technique. Ask for an example of each obstacle using material without sensitive employee data. If the task and its intended benefit remain undefined, defer a broad training purchase until the brief becomes more specific.
Do not use the survey to label untrained employees as incompetent. A person may avoid AI because the rules are unclear or because their work is unsuitable for the available tool. Frequent use is also an incomplete maturity measure. An occasional sensitive task may deserve more attention than dozens of routine drafts. The aim is to discover what the team needs to learn, rather than turn application-opening frequency into an employee performance score.
This conversation should include the reviewer as well as the person producing the draft. A learning need can sit at either end of the workflow. If a reviewer cannot explain why a response is acceptable, a faster drafting process will not resolve the underlying uncertainty. Write down the standard before using it to judge the learner or the tool, and identify the policy owner who resolves disputed exceptions.
Test learning with two measures, not one impression
We propose a small training experiment. Select a low-risk task, use fabricated or public material and agree a quality standard before the exercise. Complete one example without the tool and a comparable example with it. Record time spent on production, checking and correction together. Avoid comparing a simple case with a complex one, and do not treat a small difference from a single example as proof of a general improvement.
Consider a hypothetical administrative response that normally takes 30 minutes. AI produces a draft in 10 minutes, but checking and correcting it takes another 25. The new workflow takes 35 minutes: five minutes longer in this example. This is not an automatic reason to reject the tool. The final wording might be better or easier to update. But claiming a 20-minute saving would exclude verification costs. Compare accepted outputs rather than the moment the first answer appears on the screen.
A useful quality standard could cover consistency with the approved policy, correct handling of exceptions, clarity and the absence of irrelevant information. Separate substantive errors from stylistic preferences. Where reviewers disagree, record why. The policy itself may be ambiguous. Do not penalize a learner for an unresolved organizational question, or attribute the benefit of clarifying that question entirely to AI.
If the exercise reveals that review takes longer because the draft hides uncertainty, training can focus on exposing that uncertainty. For example, learners can practice identifying which part of a response requires a policy owner’s decision. This remains a proposed learning activity, not an outcome demonstrated by the CIPD survey. Repeat the comparison with further examples before deciding whether it transfers to routine work.
What this exercise does not authorize
A successful exercise does not authorize automated hiring, dismissal or compensation decisions. Producing an acceptable general administrative draft does not establish a model’s suitability for evaluating a person. Privacy, access permissions, challenge routes and human review still require decisions within the organization. A UK survey does not settle how those requirements apply in Saudi Arabia or elsewhere in the Gulf, and it should not be used to bypass local or contractual obligations.
Start training around the task you have actually defined, then compare additional examples before expanding use. Monitor accepted output quality and review time together. Stop expansion when an error affects a right or protected information. The productive team discussion is specific: do we need a better tool, or clearer rules and deeper knowledge of what makes the answer correct?
References
- HR pulse report 2026: Inside the people profession
- HR Pulse report 2026: Inside the people profession — full PDF
References open in a new tab.
Discuss this with practitioners
Create a free account to save articles and join the Trainers community discussion.
Discussion 0 published comments
No published comments yet. Be the first to share your experience.