Does This Still Make Sense – Shifting the Bottleneck: Why AI Output Is Not Translating into Throughput

Ten times more content. Code delivered in minutes. Administrative work completed in a fraction of the time.

That is the promise. But some of our clients are seeing something rather different. There is more activity, greater use of AI tools, and a growing volume of AI-generated work, without a corresponding increase in what the organisation actually gets done.

They have started discussing a phenomenon I would describe as “productivity evaporation”.

The underlying anxiety is a new take on a classic economic challenge, a modern lens to view the AI revolution. We are seeing the latest iteration of what economists have historically called the Solow Computer Paradox. The phenomenon where massive technology investments fail to instantly move the needle on aggregate organisational output. Some practical examples of this might include:

• An adviser can prepare for client meetings faster, but that does not necessarily mean he makes more sales. He may simply finish earlier and play more golf.
• A claims team may complete parts of the process faster but continue to meet the same targets it always has, except the business is now also paying for the AI tooling.
• A developer may generate code in a fraction of the time. But once review, testing, rework, integration, and the effort required to understand code that nobody on the team actually wrote are taken into account, some senior developers may be delivering more slowly than before.

The tools work, and people often feel faster when using them. The problem is that this does not always translate into greater throughput for the organisation.

“This is not really an AI problem. It is Operations 101.”

As per the Theory of Constraints (Dr. Eliyahu M. Goldratt, 1984), if you accelerate one part of a production line without changing the rest of it, you have not removed the bottleneck. You have moved it, usually to somewhere less visible, with less capacity and less authority to say, “This isn’t good enough yet.”

This is particularly important for work that carries real consequence: contracts, claims decisions, fraud investigations, client communications, project portfolio decisions, or the final deliverable on a consulting engagement.

No matter how quickly the first draft is produced, review and sign-off still happen at human pace. They happen when one of a relatively small number of experienced people has the time and headspace to consider the work properly and is willing to put their name to the decision.

Those people were often already the scarcest resource in the organisation. Generative AI has not created any more of their time.

Some of the slowness in the old process was therefore not waste. A contract taking a week was not necessarily the result of inefficient drafting. Part of that time was spent waiting for someone sufficiently experienced and accountable to review it and own the decision.

Generative AI can make the relatively easy 80% almost immediate while leaving the difficult and scarce 20% largely unchanged. The difference is that the final 20% may now be receiving far more work, from far more people or agents, all expecting the same limited group of senior reviewers to keep pace.

Something eventually has to give. It may be quality, the depth of review, or the capacity of the people doing the reviewing. Much of this happens quietly because “content produced” makes for a better presentation slide than “content that survived scrutiny.”

I managed to demonstrate the same problem to myself recently:

• I wanted to create a quick video in which AI would read a script I had written. It sounded like a task that should take a few minutes. Five hours later, I was still adjusting prompts, trying different tools, and regenerating takes.
• I could have recorded a simple version myself in about two minutes.
• The technology did what it was supposed to do, but the result did not justify the effort. The juice was not worth the squeeze.

That is a small and personal example, but the same problem can occur across an organisation. People spend time learning new tools, refining prompts, checking outputs, correcting mistakes, and finding increasingly sophisticated ways to complete tasks that may not have been especially difficult to begin with.

The fact that the work involves AI makes it feel innovative. That does not necessarily make it valuable.

Nor does saving someone two hours automatically create two hours of additional capacity for the organisation. That only happens if there is more useful work for the person to do, the incentives encourage them to do it, targets change, and the rest of the system can absorb the increased output.

An adviser saving time on preparation does not automatically create another sale. A claims administrator completing one task faster does not automatically settle more claims. A developer producing code more quickly does not necessarily get working software into production any sooner.

Most importantly, an organisation cannot manufacture more senior judgement simply because work is arriving faster. In a world where less used to be more, more is now quite possibly less.

There may also be a longer-term capability cost. If AI increasingly performs the first layer of analysis and decision-making, junior people may lose the opportunities through which judgement is developed. The organisation could become dependent on a shrinking group of senior decision-makers, with fewer experienced people progressing behind them.

Measuring adoption through licences, prompts, hours saved, or content generated tells us whether the tools are being used. It does not tell us whether the organisation is performing better. It also tells us very little about whether the people responsible for reviewing and approving the work can keep up with the volume being created.

This is therefore not only a tooling problem. It is a management question. Leaders decide where AI is introduced, which measures define success, how much additional work the system is allowed to generate, and whether review capacity is redesigned alongside production capacity. If AI adoption creates more activity without more value, leadership cannot place all the blame on the technology or its users.

If you are responsible for signing off on generative AI at scale, the question is not simply how much faster the organisation can produce.

It is whether the organisation is actually achieving more, whether the additional output is worth producing, and whether the people expected to exercise judgement over it still have the capacity to do so properly.

The next phase of AI adoption should not begin with more licences. It should begin with evidence. Show the additional sales. Show the claims settled. Show the software safely released. Show the decisions improved. Show the cost removed. Show what happened to the hours supposedly saved. If none of that can be demonstrated, the organisation does not yet have an AI productivity story. It has an AI activity story. And activity, however innovative it looks, is not value.