The most cited finding from the Harvard Business School and Boston Consulting Group field experiment with GPT-4 is the productivity number: 758 consultants completed 12.2% more tasks and 25.1% faster, with quality rated 40% higher, when working on tasks inside the model’s capability frontier. The less-cited finding is the one safety-conscious training programs have to grapple with. On tasks outside the model’s frontier, the AI-assisted group did worse than the control. They were more confident, faster to commit to an answer, and more wrong. The AI made the failure mode more dangerous, not less.
Most corporate AI training programs do not teach this. They teach prompts. They show off the wins. They skip the governance, the data hygiene, and the failure modes that determine whether the rollout creates value or liability. Below is a framework for the missing layer, organised around three categories of risk: data, hallucination, and judgment.
Data: what employees should never paste into a public model
Public AI tools are improving on data handling, but the safe assumption inside any training program is that anything pasted into a free-tier consumer chat interface may be used for training, retained for support investigation, or seen by a vendor employee responding to a support ticket. Employers signing enterprise agreements typically negotiate stricter data terms, but the consumer tier most employees use does not have those guarantees.
A useful training rule, kept short enough to remember:
- No regulated data. Patient records, financial account details, anything covered by HIPAA, GDPR, GLBA, or sector-specific privacy regimes. Use sanctioned enterprise tools or do not use AI for that task.
- No identifying customer data. Names, contact details, contract numbers, anything that ties an output to a real person.
- No internal strategy documents marked confidential. Roadmaps, M&A material, unannounced product information. Either redact or do not use.
- No code from proprietary repositories without an enterprise tool. GitHub Copilot for Business, ChatGPT Enterprise, and Claude for Work all offer different data handling guarantees. Pick one and standardise.
The training should include a 60-second decision rule that an employee can run before pasting. If you cannot answer “yes” to “would I be comfortable with this text appearing in a vendor support log,” the answer is route the task to the sanctioned tool or do it without AI.
Hallucination: the failure mode the BCG study quantified
The Harvard and BCG study is the cleanest evidence that AI degrades performance on out-of-frontier tasks. Hallucination is one mechanism. The model produces a plausible-sounding answer, the employee accepts it, and the error compounds because the AI confidence cue (fluent prose, internal consistency) overrides the human’s usual skepticism.
The countermeasure is not “be careful.” Be-careful is not a skill. The countermeasure is a structured evaluation habit, taught and reinforced.
A copy-paste prompt every employee should learn to use after generating a substantive output:
Take the response you just produced. Identify every factual claim, every named entity, and every number. For each, rate your confidence on a scale of 1 to 5 and explain what evidence supports it. Then list the three claims most likely to be wrong if any of them are wrong.
This is a self-critique pass. It does not eliminate hallucination. It surfaces the highest-risk claims so the human knows what to verify. Pair it with a simple rule: any claim under 4 confidence with no cited source must be checked against an authoritative source before the output ships. Perplexity and NotebookLM are both useful for this verification step, the first for live web sources, the second for grounding answers in documents the team uploads.
Judgment: the governance layer that most training skips
The McKinsey State of AI 2025 found that only 7% of organisations have scaled AI across the org, and only 39% can attribute any EBIT impact to it. The single biggest differentiator the report identifies between high performers and the rest is governance, defined operationally as named owners for AI risk, defined approval workflows for new use cases, and measurable outcomes tied to specific projects.
For a training program, the governance layer translates into three concrete artefacts every employee should know exist and how to use:
An approved-tool list. A short document, maintained by IT or the AI working group, that names which AI tools are approved for which kinds of work. Employees should not have to guess. The list should be referenced in training, not assumed.
A use-case approval flow for higher-risk work. Any AI use case that touches customer-facing communication, regulated data, or external decisions (hiring, lending, clinical) should require a sign-off from a named owner. The flow should be lightweight (a one-page template, a 24-hour SLA) but real.
An incident reporting channel. When an AI output produces a problem (a wrong customer answer, a leaked piece of data, a hallucinated citation in a report), the employee should know exactly where to report it. The reporting should be blameless. The point is to learn the failure modes faster than the next team can repeat them.
Training that includes these three artefacts is materially different from training that just teaches prompts. It treats AI as a regulated workflow component, not a magic productivity wand.
What the rollout should look like in practice
Combine the three layers and the training program has a defensible shape: one hour on data hygiene with the four-rule decision tool, one hour on hallucination with the self-critique prompt and a verification workflow, and one hour on governance with the approved-tool list and the use-case flow. Three hours of structured training. Two refreshers a year. A monthly incident review where the team walks through one real failure (not a hypothetical) and what the prevention would look like next time.
The companies that do this will have employees who use AI to produce more value. The companies that skip it will have employees who, per the BCG study, are more confident on out-of-frontier tasks and quietly making expensive mistakes. The training is not optional. It is the difference between the productivity gain and the liability event.