AI can remove routine work while leaving a difficult question unanswered: how will people keep developing the expertise needed to judge its output?
A team may still have experienced reviewers today. That does not explain how a new colleague will learn the work, or how an experienced colleague will stay ready for an unfamiliar case. When automation changes the tasks people perform, leaders should also revisit the opportunities people have to practise.
The leadership decision is which human capabilities must remain available, and how the organisation will keep them usable.
Identify the skills the business still depends on
There is little value in preserving every manual step simply because it used to exist. Start with the responsibilities that remain after automation: recognising an exception, checking evidence, explaining a decision, identifying when an answer is outside the system’s scope, and deciding when to escalate.
For each important AI-assisted workflow, ask:
- Which judgments must a person still be able to make?
- Where will people encounter enough relevant cases to develop that judgment?
- Which skills are now used so rarely that normal work may provide too little practice?
- Who is learning to replace or support the current experts?
This extends role-based AI training into the months after adoption. Initial training and assessment can establish readiness for a role. Continued practice helps maintain that readiness as the work changes.
Protect a learning path for less experienced colleagues
Consider a hypothetical service team using AI to draft responses to customer requests. Previously, a junior colleague examined the policy, assembled the facts, proposed an answer, and received feedback. After automation, the same colleague may mainly approve a finished draft.
The draft can be useful. But the manager should check whether the new workflow still exposes the colleague to the reasoning behind the answer. Seeing a polished result is different from working out which evidence matters.
A practical response is a small, protected set of learning cases. Ask the colleague to identify the relevant policy, missing information, and escalation point before revealing the AI draft. An experienced reviewer then discusses both approaches. The purpose is to practise the judgment the job still requires, rather than to recreate all the work automation removed.
Use fictional or appropriately de-identified cases in an approved environment. Do not turn practice into a reason to copy customer information into an unapproved AI tool.
Design practice around meaningful differences
Repetitive exercises can become another completion metric. Select cases that expose different kinds of reasoning: an incomplete request, a conflicting policy, a plausible but unsupported answer, or a situation where the correct next step is to ask for help.
A useful practice session has four parts:
- Independent assessment. The person records a provisional judgment and the evidence behind it before seeing the AI output.
- Comparison. They examine where their assessment and the AI answer differ, without treating either as automatically correct.
- Expert feedback. A suitably qualified reviewer explains the relevant reasoning, including uncertainty and acceptable alternatives.
- Transfer to work. The team identifies whether the lesson requires a clearer instruction, better reference material, or a change to the workflow.
The example is a suggested operating approach, not a prescribed standard or a claim that one exercise suits every role. Frequency and depth should reflect the consequence of error, the complexity of the work, and the opportunities for supervised practice already available.
Make competence visible without rewarding agreement
Agreement with the AI is a weak assessment target. So is disagreement for its own sake. A capable reviewer should be able to explain what supports an answer, what remains uncertain, and which next action is justified.
Review evidence such as whether people identify missing information, locate the applicable rule, recognise the limits of their authority, and make an appropriate escalation. Keep the assessment proportionate and use it to identify support needs. A single missed case should not be treated as a complete judgment of someone’s ability.
Keep a short record of the capability practised, the type of case, the feedback given, and the next learning need. Attendance can show that a session happened; it cannot, by itself, show that the required judgment remains available.
Give practice an owner and time in the workload
A request to “keep your skills current” is incomplete if the operating model measures only speed and volume. Name the person responsible for selecting cases and arranging feedback. Protect time for the people who practise and the experts who coach them.
Review the arrangement when responsibilities change, a new AI capability removes more of the task, or recurring exceptions reveal a gap. Also check whether expertise is concentrated in one person. A documented procedure is useful, but it does not automatically create another capable practitioner.
Where a skill is required during disruption, connect this work to the AI exit and continuity plan. Continued learning and fallback readiness overlap, but they serve different needs: developing people during normal operations and keeping essential work possible when a system is unavailable.
A question for the next adoption review
Ask the workflow owner: “If our most experienced reviewer left, what would show that someone else could exercise this judgment?”
The answer should point to a learning path, supervised practice, and demonstrated capability. If it points only to a training certificate or an available AI assistant, the team has a useful gap to examine.
This approach extends the capability-building emphasis of the Team & Ownership and Adoption Roadmap phases in VORTA. As an external reference, NIST’s AI RMF Playbook, GOVERN 2.2, discusses role-appropriate AI risk-management training and assessing the skills needed for assigned responsibilities. The practice structure above is a practical synthesis, not a NIST requirement.
For help reviewing the human capabilities your AI-assisted workflows still depend on, schedule a consultation.