The Choice Architecture Tax
There’s a number buried in this week’s research that I can’t stop thinking about.
Knowledge workers now spend 23% more time on what researchers are calling “choice architecture” — evaluating and deciding between AI-generated options — than they spent on original creation two years ago.
Not 23% more productive. Not 23% more output. 23% more time making meta-decisions about what the AI gave them.
That number is the mechanism nobody’s talking about.
What’s actually happening
You adopted AI tools. Tasks got faster. Tasks that used to take twenty minutes now take four. That part is real.
But here’s what replaced the twenty minutes: a steady stream of small decisions.
Which of these three email drafts to send. Which of these four research summaries to trust. Which of these two code implementations to build on. Which of these suggested next steps actually matters for your specific situation.
Each one feels trivial. Each one takes thirty seconds. Multiply by a full workday, across every workflow you’ve AI-enabled, and you’ve created a new category of work that didn’t exist before.
This is the choice architecture tax.
The AI didn’t take work off your plate. It converted task work — making the thing — into decision work — evaluating what the AI made. And decision work is more cognitively expensive than task work. Every option assessment requires loading context, comparing against criteria, and committing to a direction. It’s not the same muscle as executing. It’s harder.
The data that should have been obvious in hindsight
BCG surveyed 1,488 full-time workers on AI use, cognitive load, and performance. The headline finding got coverage: 14% more cognitive effort, 19% more information overload for heavy AI users.
What didn’t get as much coverage: only 8% of the time savings from AI tools are being reinvested in ways that benefit the worker.
The other 92% is absorbed by the organization as increased output demands — or consumed by the overhead of managing AI-generated options themselves.
This is why 67% of workers who adopted AI tools in 2025 reported working more hours, not fewer, by end of year. Not because AI failed. Because the choice architecture tax is invisible. Nobody budgeted for it. Nobody built infrastructure to reduce it.
The UC Berkeley longitudinal study found the same pattern. You adopt the tools. Your per-task output goes up. Your total cognitive load goes up faster. You end the year more productive on paper and more exhausted in practice.
The constraint-reveal
This is the part people get wrong.
The common diagnosis is: too many AI tools, overwhelming volume of output. The prescription is: be more selective about which AI tools you use, or learn better prompting to get better first drafts.
That’s not wrong. It’s just addressing the wrong layer.
The real constraint isn’t the quality of AI output. It’s the absence of a system that knows what matters enough to pre-decide most of these choices on your behalf.
Every time you answer “which of these three emails should I send,” you’re doing coordination work your infrastructure should handle. Not because the AI can’t draft emails — it can — but because without context about what’s open, what you’ve committed to, what your actual priorities are this week, the system genuinely cannot tell you which option is right. It can generate options. It cannot select between them without knowing what matters.
So it generates. You decide. The tax compounds.
The difference between intelligence and coordination
Here’s the distinction that matters.
AI tools have gotten very good at generating. They can produce options, drafts, summaries, code, analysis — faster and better than most humans on most tasks.
What they haven’t gotten good at — what nobody has built — is the system that knows enough about your open loops, your commitments, your current priorities, to reduce the decision surface rather than expand it.
Not “here are three email drafts.” But “here’s the one draft that’s consistent with what you told Sarah last Tuesday, accounts for the thread that’s still open with Marcus, and matches the tone you use with this client.”
That’s not a model capability problem. Gemini, GPT-4o, Claude — they can all synthesize that context if you give it to them explicitly. The problem is nobody has built the layer that maintains it implicitly. The system that tracks your open threads, your commitments, your priorities, and feeds that context automatically so the AI can decide rather than generating options for you to evaluate.
Intelligence without coordination generates options. Intelligence with coordination resolves them.
Right now, you have intelligence. You are the coordination layer.
What 23% means at scale
Put the number back in.
A knowledge worker today spends roughly 8 hours in productive cognitive work. 23% of that is now choice architecture — evaluating AI output, selecting between options, making meta-decisions the system couldn’t make without more context.
That’s roughly two hours a day. Every day. Spent not on the work, but on managing the interface between the work and the system that was supposed to make the work easier.
That’s not a productivity gain with a side effect. That’s a productivity gain with a hidden cost that’s eating the gain.
The question isn’t whether to use AI tools. They’re faster, and the baseline benefit is real. The question is whether you build the coordination infrastructure that reduces the choice architecture tax — or whether you keep absorbing it.
Two hours a day is 500 hours a year. Per knowledge worker.
That’s the market for what comes after the AI tools.
The pattern
The tools got faster.
The overhead got bigger.
The gap between intelligence and coordination — the gap between “AI generates options” and “AI knows what matters” — is where the 23% lives.
Nobody is closing that gap with a better model. You close it by building the layer that knows what’s open, what’s committed, and what should resolve — so the AI can select rather than generate.
That’s the infrastructure play. Not another tool. A different kind of system entirely.
The choice architecture tax is evidence of the missing layer.
It’s also a map for where to build next.
