The gap is a management problem, not a model problem

The assumption underneath most AI roadmaps is that the gap between what the technology can do and what an organisation actually gets from it will close as the models improve. Two pieces of evidence from mid-2026 point the other way. A stronger model does not shrink the gap by itself. It can widen it, by making more work look safe to delegate faster than any organisation can build the supervision, judgement and accountability to match.

The hours nobody put in the business case

Glean surveyed workers about what using AI agents actually involves day to day, and coined a word for what it found: botsitting. Respondents reported spending an average of 6.4 hours a week supplying an agent with context it did not have, checking its output before relying on it, and rerunning the cases it got wrong. None of that is a bug in a particular product. It is what delegating work to something that cannot yet be fully trusted actually costs, and it is real labour that a business case built only from the model's capability will not contain.

The direction of travel makes this worse before it makes it better. A more capable model does not reduce botsitting by being more capable; it expands the range of tasks that look ready to hand over, which expands the surface that needs checking. Unless an organisation is deliberately investing in the supervision, review and escalation structures that make delegation safe at the new scale, a capability jump shows up as more hidden labour, not less, however it is reported at the top of the business case.

The variable that actually predicts value

A KPMG quarterly pulse survey found organisations where the CEO is personally accountable for AI outcomes were more than twice as likely to report meaningful business value than organisations without that accountability. This is an association from a single survey wave, not a controlled result, and it does not prove that assigning a CEO the job causes the value; organisations that already take AI seriously enough to name an accountable executive may simply be doing several other things right at the same time. Read cautiously, it is still a useful data point pointing at the same place as the botsitting finding: realised value tracks who owns the gap between capability and outcome, not which model sits underneath the work.

Why the model was never going to close it alone

It is tempting to treat botsitting as a temporary tax that better models will eventually stop collecting. That misreads what the hours are actually paying for. Checking, context and correction are not overhead on the way to a fully autonomous system; they are how an organisation keeps a stake in decisions it is increasingly not making by hand, and they scale with how much is being delegated, not with how good the delegate is. A model that can attempt harder tasks needs judgement applied to harder cases, by someone who understands both the work and the system's limits. That judgement has to be assigned, resourced and reviewed like any other part of the operation. It does not arrive as a side effect of a better release note.

What this changes for a leadership team

The practical implication is not to slow down capability adoption. It is to stop budgeting an AI programme as though the model were the whole cost and the only lever. A roadmap that upgrades the model every quarter and leaves supervision, escalation and accountability exactly where they were is a roadmap for a widening gap, however impressive the underlying benchmark numbers look. The organisations the evidence points to as doing better are not the ones with access to the newest model first. They are the ones where somebody senior owns the space between what the model produces and what the organisation can actually trust, and treats closing it as an ongoing management job rather than a launch-week task.