The expertise AI still needs

One of the more encouraging findings about AI at work emerged when the software failed. In a study of more than 5,000 customer-support staff, AI assistance raised the number of issues resolved per hour by 15%, with larger gains among less experienced employees. During outages, workers still handled chats faster than before receiving the assistant. The outage evidence was less precise, but suggested that the tool had taught them something they could use without it.

The business case for workplace AI should therefore include what happens to human expertise. Faster output is an immediate gain; a workforce better able to understand and direct its work is a longer-term asset. That takes knowledge and discretion. A human signature at the bottom of a machine’s answer provides little assurance if the signatory can neither judge it nor challenge it.

How expertise is made

For a manager, correcting a junior colleague’s first draft may seem an expensive use of time when AI can produce a serviceable version in seconds. Yet working through corrections can teach the colleague why one answer is sound and another merely plausible. The cost of that education appears in today’s budget, its value may emerge years later. When only output counts, teaching can look like an avoidable expense. A firm can enjoy the savings while failing to train the people who will eventually replace its experienced staff.

In a short randomised experiment with 52 programmers learning an unfamiliar software library, those given AI assistance performed worse on a subsequent test of understanding. Reported in a preprint, the study cannot tell us how skills develop over years. It does show why completing an assignment and understanding it require separate measures.

Judging AI output is demanding, too. In a 2023 experiment with 758 consultants at Boston Consulting Group, GPT-4 improved performance on tasks within its capabilities. On a task beyond those capabilities, participants given AI access were less likely to answer correctly. The reviewer’s job is to spot where an answer has become unreliable, however convincing it appears.

Some of the time saved by AI could finance a better apprenticeship. Employees could attempt more demanding work, using AI for explanations and feedback while discussing difficult cases with experienced colleagues. Simply handing a junior colleague more finished work to approve may produce speed without much understanding. The return on the investment would be a growing ability to recognise an unfamiliar problem and explain a decision. That will not show up in a count of completed assignments.

The right to disagree

Knowledge is only half the issue. An employee may understand why an AI recommendation is unsuitable and still find it difficult to depart from it. A manager may demand an explanation for every departure while asking none when the advice is followed. A performance target may reward speed above investigation, leaving little time to question an answer. Formal permission to override then offers limited practical control. Responsibility sits with the employee while the conditions for exercising it are set elsewhere.

If questioning an AI-assisted proposal is taken as criticism of the colleague who produced it, doubts may go unspoken. A leader who expects colleagues to scrutinise AI advice should be willing to have their own reasoning examined, too. Employees also need a say in which problems AI is asked to solve and how success is assessed.

This influence can extend beyond individual teams. An agreement at IBM Germany established an AI Ethics Council with employee and employer representatives alongside AI experts. Its remit includes hearing employees’ objections and reviewing and correcting AI recommendations. Employees have a formal route for disputing what the technology recommends.

AI could make expertise available to people previously excluded from it and open up work they could not undertake alone. Whether that promise lasts depends partly on what employers do with the time saved. An organisation that wants people to take responsibility must invest in their judgement and give them the authority to use it. Those costs belong in the business case as surely as the software subscription.