The cleanest single piece of evidence on what AI does to professional work is the Harvard Business School and Boston Consulting Group field experiment with GPT-4. 758 consultants were given a real consulting task and access to GPT-4. The treatment group completed 12.2% more tasks, did them 25.1% faster, and produced output rated 40% higher in quality. The catch is the part most summaries skip. The performance gain only held on tasks inside the model’s capability frontier. On out-of-frontier tasks, the AI-assisted group performed worse than the control. Faster, more confident, and more wrong.
That finding reframes what “teaching employees to work with AI” actually means. It is not teaching them to generate. It is teaching them to evaluate, to recognise the boundary of what the tool can do well, and to keep the failure modes from becoming expensive.
The skill the training has to teach
Most AI courses are 80% prompt engineering. Prompt engineering is necessary but not sufficient. The binding constraint on usable AI productivity in 2026 is the human’s ability to spot a confidently wrong output fast. The Anthropic Economic Index shows augmentation (52%) has overtaken automation (45%) as the dominant interaction pattern, which means the productivity gain depends almost entirely on the human in the loop catching what the AI gets wrong.
A useful framing for a training program is the three-skill stack:
- Generation. Writing prompts that produce useful outputs. This is the skill most courses teach. It is also the easiest of the three.
- Evaluation. Reading an AI output and identifying what is likely wrong before shipping it. This is the skill most courses skip. It is the binding constraint on real productivity gain.
- Routing. Deciding which parts of a task to give to AI and which to keep. The HBS and BCG study makes this an explicit skill. Workers who can route well perform 12% more tasks, 25% faster. Workers who cannot, or who route poorly, end up in the out-of-frontier failure zone.
The training plan should give roughly equal time to all three. Most do not.
A worked example: a customer support escalation summary
The classic “AI for support” use case is summarising a long support thread before escalating to a senior engineer. Here is what each layer of the skill stack looks like in practice.
Generation. The prompt:
Summarise the support thread below for a senior engineer who has not seen it. Cover, in this order: the problem the customer reports, what they have tried, what we have tried, the current state, and the open questions. Keep the summary under 200 words. Use bullet points only for the lists of attempts. Plain prose for the rest.
That prompt, used in ChatGPT or Claude, produces a usable first draft on most threads. This is the easy part.
Evaluation. Before sending the summary on, the support agent runs a self-critique pass:
Read the summary above. Identify every factual claim drawn from the thread. For each, point to the line in the thread that supports it. Then list the three claims most likely to be wrong if any of them are wrong, and the questions I should answer before I trust them.
This step is where the skill compounds. Over 50 summaries, the agent learns the model’s recurring failure modes on this specific task: it sometimes attributes a quote to the wrong party, conflates two attempted fixes, or misses a recent escalation status change. Once those patterns are known, the evaluation pass takes 30 seconds.
Routing. The agent decides which parts of the escalation work to keep doing manually. Often the summary is fine for AI to produce. The judgment call about whether the issue actually warrants escalation, the politics of which senior engineer to route it to, and the framing of the customer’s history with the company are tasks the agent keeps. The AI compresses the documentation. The human keeps the decisions.
That is what teaching employees to work with AI looks like at the task level. It is not “use ChatGPT.” It is generation, evaluation, and routing, taught on the actual work.
Building the program around the three skills
Three concrete components make the skill stack teachable inside a real organisation.
A shared prompt library, organised by task and role. The library is populated by the team itself, not by a vendor. Run a three-week pilot with 15 to 20 enthusiastic early adopters. Collect the prompts that produced good outputs. Annotate each with the task and the failure mode the user found. The result is a 50 to 100 entry library that becomes the spine of formal training. New hires copy from it. Skeptics see it and find a starting point. The library is the artefact that survives the rollout.
A weekly evaluation drill. A 30-minute session, four to six people, where each person brings one AI output and the group walks through what could be wrong with it. This is the single most useful hour in the training calendar. It builds the evaluation skill durably. The format works for Microsoft Copilot, Gemini, and any other tool the team uses, the skill is tool-agnostic.
A monthly routing review. Once a month, the team picks one workflow and asks: which parts of this should AI be doing, which parts should we keep, and where is the line between the two right now. Workflows shift as model capability improves. A task that was out-of-frontier in June 2024 is often in-frontier by January 2026. The routing review is how the team keeps the work allocation current.
What the data actually says about returns to this kind of training
Two findings worth taking seriously.
The McKinsey State of AI 2025 high-performer analysis found that organisations getting EBIT impact from AI are three times more likely to have senior leaders demonstrably owning adoption, and consistently more disciplined about training, with named cohorts, named owners, and measured outcomes. The skill stack does not work without the organisational support around it.
The WEF Future of Jobs Report 2025 projects 39% of workers’ skill sets transformed or outdated by 2030. AI literacy is among the top five fastest-growing skill needs. The window for building the capability is now, not after a competitor has already done it.
The headline takeaway is not that AI is replacing employees. It is that the employees who get good at evaluating and routing are now significantly more productive than the ones who only learn to generate. Training that misses the second and third skill is training that produces enthusiasm without measurable output. The companies that get this right will look very different in 24 months from the ones that did not.