AI EXPOSES THE NEXT BOTTLENECK

Designing the human system around AI

The case for AI at the task level is, we may now reasonably argue, beyond dispute. 

Research reveals that customer support agents resolve 14% more issues per hour. Junior consultants complete tasks 25% faster - and at higher quality. Power users compress what used to be weeks of work into hours.

However, the case at the organisational level is altogether muddier. A National Bureau of Economic Research survey of nearly 6,000 executives across the US, UK, Germany and Australia found that 70% of firms now use AI in some form. Yet 80% reported no impact on productivity or employment over the past three years. PwC's 2026 Global CEO Survey landed in similar territory: just 12% of CEOs said AI had delivered both cost and revenue benefits, while 56% saw no significant financial benefit.

So there is real performance. But it has stubbornly stayed at the task layer. 

What feels fast in the workflow has not yet shown up in the P&L.

This isn't a new pattern. In 1987 Robert Solow remarked that 'you can see the computer age everywhere but in the productivity statistics' - a line that held for two decades. Stanford's Erik Brynjolfsson has argued the AI version may now be lifting, pointing to revised US data showing roughly 2.7% productivity growth in 2025, nearly double the average of the prior decade. But, as mostly everywhere in economics, there is no consensus. Apollo's chief economist Torsten Slok reckons: 'AI is everywhere except in the incoming macroeconomic data.'

And we should not be surprised. The mismatch has a structural explanation. 

A team automates one step in a workflow. The dashboard reports the expected productivity gain. Three months later, the team is no less busy. 

Why? Because the work has moved into checking, escalation, coordination and explanation. Somewhere downstream a constraint that used to not be a constraint has become the new limiting factor.

And the rule of thumb to absorb: a faster task does not automatically deliver a faster organisation. A faster task can simply expose (or even create) a downstream bottleneck in the system: data quality, handoffs, approvals, integration, customer response, or capacity in another team. 

Given intelligence used to be in short supply and only human in nature, creating knowledge-work output was expensive. Human labour was often the constraining factor. A bottleneck. 

But now intelligence is in far greater supply and also digital in nature, it’s easy to do more of that creation. Often at higher quality levels. 

But creation isn’t the whole process. By unbundling and delegating creation, checking the work becomes a bottleneck. Validating the work has been checked becomes a bottleneck. And dozens of others previously hidden bottlenecks are being revealed in the complexity of modern workflows.

Where the bottleneck moves

Not that we don’t have theories to explain this and a body of practice to help solve it.

Eliyahu Goldratt's Theory of Constraints states that improving a step which is not the system's bottleneck does not improve the system. If AI speeds up a non-constraint, the organisation may simply produce more work-in-progress for the real constraint to absorb. If it relieves the true constraint, the next limiting factor becomes more visible. AI improves local steps quickly. But it rarely releases all constraints wholesale.

A deployment that reduces extraction time in a regulated insurance claims workflow only delivers if extraction was the actual constraint. If review and approval was already the bottleneck, extraction automation simply feeds the existing queue faster – more cases per hour to the same number of reviewers. 

If extraction was the bottleneck and AI relieves it, the next constraint may appear in review, escalation, customer response, data reconciliation or another operational handoff.

When the next constraint is human, the load lands in oversight. It lands in judgement at the decision points - where the model's output looks confident but a human must decide whether it is right in this case. And it lands in accountability, where someone has to answer for what the system did.

The solve? Approach the workflow differently. Don’t ask: ‘What is easiest to deliver via AI?’ Instead, identify the constraint slowing the work today, deploy AI to relieve it, watch where the load lands next, and treat that next constraint as the next deployment question. 

That is how gains accumulate. Without it, we may be left with faster steps inside a slower system.

The oversight bottleneck

Berkeley researchers Aruna Ranganathan and Xingqi Maggie Ye, writing in the Harvard Business Review in February 2026, found that AI tools consistently intensified work across an eight-month period. The embedded study of a roughly 200-person US technology firm reported that - as tasks got easier - scope expanded. As scope expanded, multitasking rose. As multitasking rose, the boundary between work and the rest of life dissolved. 

The productivity gain became a quality and retention concern within six months.

A BCG Henderson Institute study of 1,488 US workers points in the same direction: excessive use or constant monitoring of AI tools can increase mental fatigue, errors, decision overload and intent to quit.

In many deployments, AI displaces labour before it removes it. A 2026 survey of 3,200 employees by enterprise software vendor Workday put a number on the drag: workers were giving back close to 40% of the time AI saved them, lost to reviewing, correcting, and verifying its output.

The doing has shrunk. The supervising has grown.

The size of this oversight bottleneck is a design variable. A study of 244 BCG consultants by researchers at Harvard, MIT, and others tracked roughly 5,000 GPT-4 interactions and found three distinct modes of working with AI on the same tool, the same task, with the same access. 

They labelled them Cyborgs (around 60% of users), Centaurs (around 14%), and Self-Automators (roughly a quarter).

Cyborgs argued with the AI, broke tasks into modules, validated outputs, and built fluency. Centaurs kept tight control of analytical work and used AI selectively. Self-Automators delegated workflows in their entirety; their output was fast, polished, and weaker on accuracy.

Cyborgs and Centaurs absorbed oversight into the work itself. Self-Automators pushed it downstream – to whoever had to catch the errors. The same AI deployment produced markedly different oversight burdens depending on how the work was approached. Oversight is part of the cost model, not a free by-product of automation.

When Self-Automators push oversight downstream to a reviewer, they are also outsourcing their own learning. That points to the next constraint.

The apprenticeship bottleneck

Erik Brynjolfsson and colleagues at Stanford's Digital Economy Lab tracked US payroll data through late 2025 and found that young workers in the most AI-exposed occupations saw a 16% relative employment decline against older peers. The codified tasks organisations used to give juniors as their training are the tasks AI can now absorb.

The Citi-Hildebrandt 2026 advisory found 86% of large law firms planning to grow associate ranks overall through 2027, but only around a third planning to grow first-year classes. It is a pyramid quietly turning into a cylinder, where senior roles expand but the rung that used to produce them does not.

No single executive sets out to break the corporate apprenticeship. But a series of defensible, task-level efficiency metrics can do it by accident.

Cyborg and Centaur behaviours build the expertise needed to train the next generation. Self-Automators risk progressing without building the underlying skill. How AI is deployed now will shape whether the apprenticeship pyramid still works in ten years' time. Companies must design the training architecture alongside AI adoption: make AI-augmented work itself a training environment, pair juniors with senior reviewers in deliberate ways, and be explicit about which skills the work is meant to build.

The judgement and accountability bottlenecks

If oversight and apprenticeship are internal costs, judgement and accountability are external risks. In regulated industries, where this pressure lands is already plain to see.

The MHRA's AI Airlock pilot report, published in October 2025, showed how the judgement bottleneck can emerge. During the trial, a pattern of over-reliance emerged at one of two participating hospitals but not the other. The difference was not technical; it was driven by clinician fatigue, time pressure, and experience level. Where clinicians are tired, time-pressed or less experienced, human review can become thinner than the governance model assumes.

When those decisions go wrong, the accountability bottleneck appears. The FCA's AI Live Testing programme is running a 2026 cohort with eight named firms, including Barclays, Experian, UBS, and Lloyds Banking Group. Regulators are looking beyond the models. They are auditing the human systems around them. The design choices being inspected are entirely human: who reviews the output, what evidence is kept, how over-reliance is mitigated, and where accountability sits when the system fails.

The default is a design

If leaders do not actively design this system, someone else will. The defaults are being set by vendors whose incentives favour subscription seats, by deployment defaults that optimise for individual productivity, and by procurement cycles organised around tool budgets rather than work design.

The default is not neutral. It is a design – one whose primary objective is rarely the work itself.

The management task is system design. Identify the bottleneck slowing the work today. Deploy AI to relieve it. Track where the load lands – in another operational task, a data dependency, a handoff, or a human constraint such as oversight, judgement, skill formation or accountability. Treat that next constraint as the next deployment question, and design the human system that receives the gain alongside the model.

We all still need to ask where AI saves time. The harder question is what fills the space it opens: who checks the work, who learns from it, and where accountability sits when the dashboard reports the system is moving faster.

A faster task is only a metric. The advantage sits in the system built around it.