The New Bottleneck is Me
In my post last week I explored research showing that AI isn’t necessarily reducing work. If anything, it is intensifying it. People work faster, take on broader scope and fill their breaks with prompts, not because they’re told to, but because AI makes doing more feel possible (and in many cases, enjoyable). The productivity metrics look good. The experience of work gets denser.
This got me thinking further into the topic and into the human-plus-machine relationship. What happens when AI accelerates everything except the part that matters most?
The Shift People Aren’t Talking About (Yet)
In my efforts to learn how to work with and practically apply AI over the last year or more, the shift in what I can produce is significant. Frameworks, strategic documents, presentations, research, even functional web applications. Things that previously required specialist input or simply weren’t possible for a solo practitioner with limited capacity and technical skills.
But the nature of the work has fundamentally changed. I still start with a blank page but very quickly engage the machine. I brief, I review, I challenge, I iterate. The bulk of the work has moved from creation to orchestration. And orchestration carries its own cognitive weight.
Before AI, thinking through a problem, structuring a response and building a deliverable happened across hours. The cognitive load was distributed across the working day. With AI, a first draft appears in seconds and a framework takes shape in minutes. But every output still requires the same human judgement.
The effort hasn’t disappeared. It has just been concentrated to specific parts of the workflow. More quality judgement compressed into shorter windows. The cognitive load is at least the same but applied in a fundamentally different rhythm.
The Work of Working with AI
As adoption grows, there is a category of effort in AI-assisted work beginning to emerge: the work of working with AI itself. Structuring prompts effectively. Deciding when to push back on an output versus when to iterate. Learning when AI is likely to be helpful and when it won’t. Getting AI to explain the best approach to a problem before attempting to solve it.
This meta-work is invisible in productivity metrics. Nobody counts the time spent optimising the interaction itself. But it is real, it is cognitively demanding and it grows as you get more sophisticated. Early on, you learn to prompt. Later, you learn to orchestrate multi-step workflows, maintain context across long conversations and build projects that accumulate knowledge over time. Each stage requires more from you.
In research last year, Microsoft observed AI shifting effort from what they call “material production” to “critical integration.” Workers spend less time creating initial outputs but more time verifying, correcting and contextualising what AI produces. As creation burden falls, judgement burden rises.
I recognise that pattern in my own work every day. The production side has never been faster. The review side has never been more demanding.
Where the Work Has Moved
AI is moving the bottleneck from production to the human side of the interaction. It generates content, code, analysis and options at a pace no individual could match. All of this output still flows through the same human checkpoint. Your experience, your context and your quality standard.
I run multiple content projects and application development in parallel through AI. Drafts, frameworks, features, all advancing simultaneously. The production capability is extraordinary. But what it creates is a queue and an increasing orchestration challenge born out of my own enthusiasm for what AI makes possible. The bottleneck isn’t that I’m slow. It’s that I’m still the singular human player in the workflow.
A big risk this creates is “close enough” acceptance. Not as a deliberate quality choice, but as a capacity consequence. With a growing backlog as production accelerates, the pressure to move things through without proper engagement grows quietly. When I resist that pressure and engage properly, the output improves. Often substantially.
This pattern shows up at scale in software development. Faros AI analysed data from over 10,000 developers across more than 1,200 teams and found that developers using AI write more code and complete more tasks, but the code is getting larger and carrying more defects. The bottleneck has shifted to code review.
The UC Berkeley research I mentioned last week documented the same dynamic from a different angle. When product managers started writing code and researchers took on engineering tasks, the experts who reviewed that work absorbed a hidden oversight load that never appeared in any productivity dashboard.
What This Means for Measurement
If the real work has moved to human judgement, then measuring AI success by output volume or time saved feels misleading. A more honest approach would reflect where the work has actually moved.
First, decision quality. Is AI helping your teams make better decisions that create better outcomes? Measured by whatever matters to the business: forecast accuracy, campaign performance, customer retention, operational efficiency. This is a hard metric to implement because it requires a clear baseline and discipline in attribution. It could also be the most meaningful. Flat or declining outcomes mean something is wrong.
Second, rework rate. This is the most direct signal of how well AI is being used. A high rework rate does not mean AI is failing, rather that AI application needs attention. It might signal poor prompting, insufficient context, or mismatched use cases.
Third, review burden. If senior staff are spending more time checking work than doing their own, the production pipeline is being loaded from below in ways that may not be visible to a rudimentary “hours-saved” measure.
Ultimately you want a framework that helps you understand whether AI is genuinely improving the quality of what your teams are able to produce or quietly degrading it while increasing volume.
Two Approaches Are Emerging
AI makes production easier, faster and available at greater scale. But production has never been the hard part of knowledge work. The hard part was always judgement: knowing what to build, whether it is good enough and when to stop or change course. AI has made this human contribution more concentrated, more constant and more consequential.
Two responses to this reality are emerging. The first bolts AI onto existing workflows and hopes for efficiency. Production accelerates, the human role stays the same and the pressure on the review layer intensifies. This is where a number of organisations find themselves today. It is also where plenty of individuals are, particularly solo practitioners, including me.
The second reimagines how work is done. Rather than asking humans to absorb more output at the same pace, it redesigns the process around what AI does well and what humans do well. It re-engineers the workflow so that human capacity is protected and applied where it creates the most value, rather than spread across an ever-growing queue of AI-generated output.
Most of us can see the difference between these two approaches. Fewer of us have made the shift. Recognising that the bottleneck is structural rather than personal is the first step. Perhaps if we stop trying to review faster and start asking whether the workflow itself needs to change, we can move the conversation to somewhere more productive.
The bottleneck is you. That is not a weakness. It is where the value lives. For 2026 I think the question is whether your organisation is building around that or just pushing more through it.
Sources
Ranganathan, A. and Ye, X.M. (2026). “AI Doesn’t Reduce Work — It Intensifies It.” Harvard Business Review, February 2026.
Faros AI (2025). “The AI Productivity Paradox Report.” July 2025.
Lee, M. et al. (2025). “The Impact of Generative AI on Critical Thinking.” Microsoft Research, presented at CHI Conference, April 2025.
Gartner (2026). “Top Future of Work Trends for CHROs in 2026.” January 2026.