The AI Iteration Tax: Why 80% Right Still Costs You
Updated
Knowledge on this page was mainly distilled from the following articles: The Tokenmaxxing Equation: AI ROI Was Never About the Tokens, Promised 10x, Got 2x. Why, and How to Fix It.
The real cost of AI assistance is not bad output. Often the first attempt is impressive. The cost is the loop between "impressive" and "what I actually wanted."
You ask AI to write something. It returns something at 80%: good structure, reasonable arguments, but it does not sound like you. So you iterate. "Less formal." "More specific here." "That's not what I meant." Each correction costs minutes. Compound those micro-corrections across a full workflow and 10x dissolves into 2x.
The Perception Gap
METR ran a randomized controlled trial in early 2025 with experienced open-source developers. They were 19% slower with AI tools but perceived a 20% speedup, a 39-percentage-point gap between feeling fast and being fast. The iteration loop does not just cost time. It costs time while feeling like a shortcut.
At scale, Faros AI tracked 10,000+ developers and found they produced 21% more tasks, but code reviews took 91% longer and bugs increased 9%. Speed shifted to a different part of the pipeline rather than disappearing.
The Iteration Tax in the Multiplicative ROI Framework
The iteration tax is not just a time cost. In the broader AI ROI equation (V = H × A × L × C), the correction loop is where human capability (H) interacts with AI task fit (A). A high iteration tax signals that one or both terms are weak for the given task. Because the equation multiplies, a poor H-A interaction does not merely reduce returns; it can collapse them entirely, regardless of how much leverage or capture potential exists.
This reframing explains why the iteration tax hits harder than simple time accounting suggests. It is not just that corrections take 5 minutes instead of 1. It is that the correction loop degrades the human capability term by consuming attention on editing rather than on judgment, problem selection, or value capture.
Q&A
What is the AI iteration tax?
It is the cumulative cost of correcting AI output from its first attempt to what you actually wanted. Each individual correction is small, but they compound across a full workflow. A task that would take one minute with a collaborator who knows you takes five minutes with one who does not, turning a potential 10x into roughly 2x.
Why does AI feel faster than it actually is?
METR's 2025 controlled trial found a 39-percentage-point gap between perceived and actual speed: developers felt 20% faster but were 19% slower. The iteration loop creates a sense of momentum because each correction step is small and feels productive, even though the total time exceeds doing the work directly.
Does the iteration tax compound at the team level?
Yes. Faros AI tracked over 10,000 developers and found 21% more tasks completed but 91% longer code reviews and 9% more bugs. Individual speed gains can shift costs downstream to review, debugging, and maintenance rather than producing net organizational productivity.
Why does the iteration tax repeat every session?
Because current models do not retain corrections across sessions. Tomorrow you will give the same corrections you gave today. The model did not absorb your preferences. A memory system might record surface-level preferences, but it cannot capture the full set of micro-preferences that define your judgment and style.
How do top performers minimize the iteration tax?
They have internalized which problems AI handles well and which ones it fumbles. They decompose requests into chunks the model can execute cleanly on the first attempt. They also recognize when the AI is solving a translated version of their intent rather than their actual intent, and they reframe before the loop begins. This skill transfers across tools.
How does the iteration tax relate to the AI ROI equation?
The iteration tax sits at the intersection of human capability (H) and AI task fit (A) in the multiplicative ROI framework. When correction loops are long, it means the H-A product is low for that task. Because the equation multiplies all terms, a weak H-A interaction can zero out returns even when problem leverage and value capture are strong.
Does the iteration tax affect human capability beyond time spent?
Yes. Every minute spent correcting AI output is a minute not spent on higher-leverage activities like problem selection, validation, or value capture. The iteration loop consumes the scarcest resource in the equation: human judgment and attention. Over time, this can degrade the effective H term across an entire workflow, not just the task being corrected.