I have seen organizations moving faster than ever: timelines compressed, output doubled, dashboards glowing green. Yet when we look at the time from identifying a problem to delivering an outcome, the room goes quiet.
The bottleneck had moved without being reflected in the productivity measures.
Execution used to be the constraint. Analysis that took a team two weeks now takes an afternoon. Plans that required three rounds of drafting materialize in a single session. The compression is real — controlled experiments show 30-55% speed improvements on scoped tasks.
But the decision architecture governing all that work was designed for slow execution. When it took weeks to build something, the timeline itself was a buffer. Poor decisions had time to surface. Misaligned priorities got caught in review cycles. The slowness was not just friction — it was an immune system.
That immune system is gone.
Some organizations now move from idea to implementation in hours, while the mechanisms that ensure those ideas deserve implementation remain slow, fragmented, and often implicit. The research helps explain why the problem can be difficult to see even for the people closest to the work.
Perception vs. Reality
In 2025, METR ran a randomized controlled trial. Sixteen experienced open-source developers — not novices, not benchmarks, but practitioners deeply familiar with their own codebases — completed 246 tasks, alternating between AI-assisted and unassisted conditions.
The result: AI-assisted developers completed tasks 19% slower.
The perception gap was worse. Developers predicted AI would speed them up by 24%. Even after experiencing the slowdown, they still believed AI had accelerated them by 20%.
Code generation felt faster, and it was. The developers still had to check what came back. Once that work was included, the overall task took longer, even though the experience suggested the opposite.
This is not an isolated finding. Foxit and Sapio Research surveyed 1,000 end users and 400 senior executives and found the same pattern at the leadership level. Executives believed AI saved them 4.6 hours per week. After accounting for validation time — 4 hours and 20 minutes spent checking, correcting, and reworking AI outputs — the net gain was 16 minutes. End users fared worse: a net productivity loss of 14 minutes weekly.
Eighty-nine percent of executives reported feeling more productive. Their own time data contradicted them.
Some of this friction reflects immature tooling, not a permanent structural gap. The METR study used tools that have already been superseded. But even if AI eliminates all verification overhead, the core problem remains. Higher-quality execution of misaligned goals is harder to question, not easier. And McKinsey data shows that 51% of AI-using organizations experienced negative consequences that persist beyond early adopters into mature deployments. The tooling will improve. The direction problem will not solve itself.
When the people making investment decisions about AI cannot accurately perceive whether it is working, the feedback loop that should trigger adaptation breaks. The organization doubles down on speed — the one thing that is not actually the problem.
Speed Produces Volume, Not Quality
For the people doing the work, the cost appears in the time spent checking and correcting outputs while their original responsibilities remain.
Upwork surveyed 2,500 global workers and found a 73-percentage-point perception gap between leadership expectations and employee reality. Ninety-six percent of C-suite leaders expected AI to boost productivity. Seventy-seven percent of employees reported that AI tools had actually increased their workload.
The mechanism is straightforward. Organizations bolted on an AI generation engine without redesigning the validation infrastructure. Employees now perform their original work and manage AI outputs. But forty-seven percent of AI-using employees did not know how to achieve the productivity gains their leaders expected of them. That suggests the workflows and expectations needed redesign, alongside tool adoption.
The hidden tax is quantifiable. Thirty-seven percent of time saved by AI is consumed by rework — correcting, clarifying, or rewriting low-quality AI-generated content.
The Bottleneck Migration Pattern
When one part of a job gets faster, the next person still has to handle what it produces. The bottleneck can move into review, validation, or coordination.
Controlled experiments show 30-55% task speed improvements, but organizational delivery metrics remain flat because the bottleneck relocates downstream — to review, validation, and coordination stages. The pattern follows a predictable sequence. The execution bottleneck dissolves. A validation bottleneck forms. A coordination bottleneck compounds. Total cycle time stays flat or increases — while the dashboard shows every individual stage getting faster.
Organizations have learned to adapt governance to match new capabilities before. Every major technology wave — from the telegraph to ERP systems to cloud computing — initially outpaced decision structures, and governance caught up. Why should this time be different?
Three reasons. First, the speed differential is unprecedented — prior waves accelerated execution by 2-5x over years, while AI compresses it by 10-100x in months, leaving far less time for organic adaptation. Second, the perception gap documented by METR means the feedback signal that would normally trigger governance redesign is broken: organizations believe they are performing better, which suppresses the urgency to change. Third, the compounding cost of misalignment at higher velocity means damage accumulates faster — projects advance further before correction, making rework exponentially more expensive.
As execution compresses, judgment expands. The limiting factor becomes the ability to define problems correctly, evaluate options rigorously, and align stakeholders. AI handles prediction — given a goal, it can generate paths, optimize sequences, and produce outputs at scale. What it cannot do is assign value to those paths. It cannot determine whether the goal itself is correct, whether the trade-offs are acceptable, or whether the outcome aligns with broader constraints. That remains a human function, and it is now under pressure it was never designed to bear.
Gartner predicts that 40% of agentic AI projects will fail by 2027 — not because the technology does not work, but because organizations are automating broken processes instead of redesigning operations. The cognitive load of work has shifted from creation to verification, an inherently senior task. And in the most concerning cases, cascading failures in multi-agent systems propagate at machine speed, amplifying poor decisions before any human can intervene.
Faster Means More Experiments
Faster execution should enable rapid experimentation. More experiments, faster learning, convergence through iteration rather than upfront analysis. The lean startup playbook, applied at AI speed.
In domains with fast, clear feedback loops — A/B testing, code deployment with real-time monitoring, controlled experiments — this argument holds. I have seen it work. When the cost of each experiment is low and feedback is rapid and unambiguous, faster execution genuinely improves decision quality through iteration.
But most high-stakes organizational decisions do not live in those domains. Strategic direction, talent architecture, capital allocation, market positioning — these produce feedback that is slow, ambiguous, and expensive. Faster execution in these contexts does not produce faster learning. It produces faster commitment to paths that take months or years to evaluate.
The data suggests organizations are not using AI speed for experimentation. They are using it for volume — doing more of the same, faster, without the experimental design that would make iteration productive.
Redesigning for Judgment, Not Just Speed
The organizations that adapt will not be the ones that accelerate execution further. They will be the ones that redesign decision systems for the new tempo.
Measure total cycle time, not generation speed. Map your five highest-volume workflows. Compare execution time to decision-and-validation time before and after AI adoption. Flag any workflow where validation time now exceeds execution time — those are your migrated bottlenecks.
Audit your decision rights under speed. Governance that worked when execution took weeks becomes dangerously loose when execution takes hours. The buffer that absorbed poor decisions has disappeared. For each critical decision point, ask: is the decision-maker seeing validated information or AI-generated first drafts? Is there a checkpoint between decision and irreversible execution?
Install direction validation checkpoints. Gates that evaluate whether work is heading in the right direction, not just whether it is progressing. A progress review asks “how much is done?” A direction review asks “should this be done at all?” When a process conflates these two questions, direction drift goes undetected because the work itself looks productive.
Close the perception gap with objective measurement. Instrument three metrics most organizations ignore: validation ratio (hours reviewing AI output versus hours saved generating it), direction rework rate (percentage of AI-accelerated work requiring course correction), and decision latency (time from problem identification to validated decision, not just to first output).
Measure the time from a request to a delivered result. If it has not improved in proportion to execution speed, look at what the people reviewing, validating, and coordinating the work now have to do. Their time belongs in the comparison too.
References
- METR. Randomized controlled trial of AI-assisted software development. 2025. 16 experienced open-source developers, 246 tasks.
- METR. Perception gap analysis from RCT data. 2025. 39-percentage-point gap between perceived and actual productivity impact.
- Upwork. Research Study on AI and Productivity. 2024. Survey of 2,500 global workers including 1,250 C-suite executives.
- Foxit / Sapio Research. AI Productivity Survey. 2024. 1,000 end users and 400 senior executives on time savings vs. validation overhead.
- Writer / Workplace Intelligence. The Hidden Cost of AI Rework. 2024. Analysis of rework costs in enterprise AI adoption.
- Bessen, Denk, Meng. Aggregate productivity meta-analysis. 2025. 371 estimates across multiple studies.
- Vaccaro et al. Human-AI collaboration performance study. 2024. 106 experiments showing combined teams underperforming best solo agent.
- McKinsey. The State of AI. 2024. Survey data on negative consequences in AI-using organizations.
- Multiple controlled experiments on AI-assisted task speed. 2023-2025. 30-55% improvement range on scoped individual tasks.
- Software development meta-analysis. 37 studies on code quality and speed trade-offs in AI-assisted development.
- Gartner. Agentic AI project failure prediction. 2025. 40% failure rate forecast by 2027 due to process automation without redesign.