AI-generated editorial illustration.

THE SIGNAL

A useful expert explains what would make them change their mind. Capture that alongside the answer.

“Apprenticeship has never been more important,” Dan Swan says in the WIRED–McKinsey interview. He also argues that business owners must own the transformation: “It has to be their initiative, their effort, not something that’s happening to them.” Read the interview.

Those ideas belong in the same implementation plan. Ask experienced employees to help design the system, and preserve the reasoning that allows a newer colleague to challenge its recommendation. A searchable archive of final answers alone cannot do that.

WHY IT MATTERS

A field study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined an AI assistant introduced to 5,172 customer-support agents. Productivity, measured as issues resolved per hour, rose 15% on average. Less experienced and lower-skilled workers improved speed and quality; the most experienced and highest-skilled workers saw small speed gains and small quality declines. Generative AI at Work, revised November 2024.

That result supports a careful proposition: assistance can spread useful know-how in a particular workflow, while affecting different workers differently. It does not establish a universal productivity multiplier, prove competence in regulated decisions or justify removing apprenticeship.

For enterprise leaders, the measurement consequence is immediate. Separate results by experience level and task difficulty. An average gain can hide an expert's extra correction work or a novice's dependence on suggestions they cannot evaluate.

Ask people where the recommendation failed, which exception they recognized and what source they trusted instead. Their disagreement is valuable operational evidence. Give them paid time to document it and a visible way to get incorrect guidance repaired.

THE DECISION EXAMPLE

Hypothetical decision for a service-operations lead: “Should we approve a four-week onboarding trial by Friday, using 20 reviewed historical cases, with a senior reviewer retaining every customer-impacting decision until new hires can explain the evidence and escalation rule?”

Choose cases that contain disagreement and missing information, not only clean successes. If permission to reuse a case is unclear, remove it from the trial until the data owner resolves the issue. Do not copy private customer details into an unapproved AI tool.

THE CONTROL TEST

Create one decision record per case with seven fields: the question; the facts available at the time; the relevant policy and version; options considered; the reason for the choice; the escalation trigger; and the eventual outcome.

Keep the original rationale distinct from hindsight. An outcome can be lucky even when the reasoning was poor. Equally, a defensible decision can have an unfavourable result. Record who reviewed the case and when its guidance expires.

Make the learning test demanding. Before showing the stored answer, ask a trainee to explain the decision. Then change one important fact and ask again. A trainee who repeats the old answer has retrieved information; they have not yet demonstrated judgment.

Measure time to complete a case, correction rate, inappropriate escalations and missed escalations. Compare performance with assistance and on separate supervised cases without it. These are proposed local evaluation measures, not claims that the cited study tested this specific programme.

A library cannot establish that all expert intuition has been captured. Sensitive, novel or ambiguous cases still require the designated decision owner. Never promote a generated answer into approved institutional guidance without review.

THE ENTERPRISE MOVE

Run a 45-minute case-review session with one expert and one newer employee. Spend ten minutes reconstructing the original evidence, fifteen comparing possible decisions, ten exploring a counterexample and ten recording the escalation rule.

Start with a small, permission-checked set. Assign a steward to each policy-linked record. Review corrections weekly and retire outdated guidance instead of quietly overwriting its history.

After four weeks, ask whether new hires explain their choices better and whether experts spend less time fixing avoidable errors. If use is high but those measures do not improve, revise the material and workflow before adding more documents. Training completion is an input; competent decisions are the outcome.

QUESTION FOR YOUR TEAM

If your most experienced colleague were unavailable tomorrow, which exception would the team be least prepared to recognize?

WHERE WISDOMTWIN FITS

WisdomTwin builds sovereign, governed AI Judgment Twins that let regulated enterprises make high-consequence decisions without waiting for the next meeting.

Bring one permission-cleared decision record to a WisdomTwin demonstration. Ask how changing evidence and approval boundaries are handled. Any benefit needs validation against your own people, cases and controls.

Founder disclosure: Roman Bodnarchuk is Co-Founder and CEO of WisdomTwin.ai. Operational examples are hypothetical; cited research and company reports are identified separately. Demonstrations use synthetic scenarios.

Read the latest briefings ↗