Who gets to decide

A customer explains a problem. The employee understands it, knows what would put it right, and still has to ask a manager for permission.

Giving that employee a better answer does not remove the wait. The obstacle is the organisation’s decision about who may act on the answer.

This distinction matters for AI because knowledge and authority can move in opposite directions. A worker might gain access to expertise that once took years to acquire while losing discretion over how to use it. The same technology that helps someone handle a difficult case can help a distant manager prescribe how every case should be handled.

The consequential question is how work changes when expertise stops being a good reason to reserve decisions for a small group of people. Organisations could distribute responsibility more widely. They could also use their new ability to supervise people to concentrate it further.

Two ways to make a decision cheaper

Suppose a company wants customer problems resolved with fewer mistakes. It can give employees better information and let them settle more cases themselves. Or it can make it easier for a central team to review their proposed decisions. Both arrangements may improve service. They create very different jobs.

There is evidence for this distinction from before generative AI. Nicholas Bloom, Luis Garicano, Raffaella Sadun and John Van Reenen studied firms in the United States and seven European countries. Technologies that helped people access information were associated with greater autonomy. Technologies that made communication easier were associated with less. Their explanation was that accessible knowledge lets people solve more problems locally, while cheap communication makes it easier to refer problems upward. This was an observational study of earlier technologies, not an experiment on today’s AI. It nevertheless gives us a reason to avoid treating every improvement in computing as a force for decentralisation. The study

AI can serve either purpose. Put an assistant beside the employee and it can explain an unfamiliar policy or suggest a way to investigate a fault. Put it between the employee and the manager and it can summarise cases, flag departures from policy and prepare recommendations for approval. The second arrangement reduces the effort required to oversee people. Nothing in the software’s intelligence requires the company to give those people more freedom.

There is evidence that assistance can spread expertise. A study of 5,179 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by about 14%, with larger gains among less experienced and lower-skilled workers. The researchers found suggestive evidence that the tool spread practices used by stronger workers. The study does not establish that employers consequently handed workers more authority. Generative AI at Work

The organisation inside the software

Consider a service employee handling a request outside the usual rules. A useful system might show the relevant history, explain the trade-offs and let the employee authorise a remedy within an agreed limit. Another system might gather exactly the same information, propose exactly the same remedy and disable the final action until someone at headquarters approves it.

These are two possible designs, not findings from the support study. Their difference is worth making explicit because a demonstration of the AI could look equally impressive in both. The employee’s working life would not.

In the first design, the employee has to exercise judgment. The organisation needs to explain the purpose of the rules, tolerate some variation and review outcomes. In the second, the employee’s main responsibility is to submit the case in a form that the approval system can process. Local knowledge matters only to the extent that it survives that process.

Central control has legitimate uses. One branch should not be able to make a promise that another must keep without knowing about it. A decision that creates a large, irreversible loss deserves more scrutiny than a routine remedy. A company may also discover that what employees experience as useful discretion is producing arbitrary treatment of customers.

But checking a decision has a cost too. The customer waits. Context gets compressed. An employee who is accountable for the outcome may be unable to change the action that produces it. As automated review gets cheaper, an organisation can become better at enforcing a rule without becoming better at noticing when the rule is wrong.

That is why the design cannot end with a threshold for approval. Someone must also be able to challenge the rule, correct the information and get an exception considered on its merits. Otherwise a system introduced to support judgment gradually becomes a system for preventing it.

A different bargain at work

My expectation is that the most valuable change will occur where firms combine broader access to expertise with broader, bounded authority. People close to a problem can act; the organisation can inspect what happened and improve its rules. This is a choice about how to organise work, not a property that arrives with a model subscription.

It also creates a practical test. After introducing AI, can more employees resolve a case without escalation? Can they depart from a recommendation and explain why? Are decisions corrected more quickly when new facts emerge? If output rises but every meaningful exception still travels upward, the company has improved productivity without distributing authority.

The strongest reason this expectation could be wrong is that local judgment may remain expensive to trust. If errors are difficult to detect, if employees face conflicting incentives, or if a decision affects many other parts of the business, central review can remain the better arrangement. More capable assistance alone does not settle those problems.

What it does change is the argument for excluding people from decisions because they cannot acquire the necessary knowledge. Where that argument weakens, the remaining limits become easier to see. Some protect customers and colleagues. Others protect a hierarchy.

For the person doing the work, the distinction is concrete. They can either understand more and be trusted with more, or understand more while spending their day seeking permission. Both futures can run on the same AI.