The bottleneck AI can't fix

Artificial intelligence (AI) keeps getting faster, but work doesn't feel any easier. That gap is the real story, and it starts with a question most offices rarely consider: faster at what, exactly?

Eliyahu Goldratt tackled a similar question in his 1984 business book, "The Goal." His main idea, which he called the theory of constraints, is straightforward. Every system moves only as fast as its slowest part. If you improve everything except that slow point, the whole system still moves at the same speed. The improvements are real but the results don't change.

Goldratt explained using this example: A product goes through framing, roofing and wiring. Roofing is the slowest step. If you add more framers or give them better tools, the framing pile just grows faster than the roofers can keep up with. The total output doesn't increase.

Today, offices are doing something similar but with software instead of framers. AI writes emails, makes slides and summarizes meetings. But these tasks were rarely the slowest part. The real bottleneck has always been judgment — deciding whether to trust a number, how to word a tough message or when to wait before making a decision. None of that speeds up just because the email was written faster.

Instead, the type of work people do changes while the amount doesn't really go down. Studies on knowledge workers' use of AI show a clear trend. As people trust the tools more, they spend less time generating ideas and more time checking them. The focus shifts from creating to verifying. This isn't free time gained but a different kind of work. It's quieter but just as demanding and often goes unnoticed.

This shift comes with a cost that isn't about output but attention. When your mind jumps between a draft, a chat window and several open tabs, each one leaves a trace. Psychologists call this "attention residue." It doesn't go away just because you switch quickly. It builds up, which is why a day of AI-assisted work can still leave you feeling exhausted for no clear reason.

This is where the bottleneck metaphor stops working, because attention isn't just another step in a process. You can't fix it by adding more people or faster software. Attention is a limited resource that needs to be protected, often by choosing not to open another tab. No tool can make that choice. Only you can.

This isn't an argument against using AI. The tools themselves aren't the issue and blaming them misses the real point. The real problem is mistaking activity for progress and thinking that if the easy parts go faster, the hard parts must be speeding up, too. But that's not how it works. Goldratt understood this about factories long before anyone invented language models.

What if organizations measured not how quickly people worked, but where their attention went? Instead of counting emails sent, they could look at decisions made with care. Instead of tracking slides produced, they could focus on judgment calls that were given enough time to be right — not just fast.

The real question isn't whether AI can save time. AI clearly does, but only for certain tasks that were never the main problem. The tougher question is what happens to the time that's saved: Does it become space for the slow, deep thinking that really matters? Or does it just get filled up by a faster version of the same distractions?

A roof still takes as long to build as it always did. The question to consider in your own work this week is: What was never the bottleneck, and what have you been avoiding?

Choi Hee-jin is an educator and practical theologian, Yale Divinity School fellow (2025–2026), exploring medical humanities at Duke. She writes about medicine, technology, and human flourishing, with imagination and hope.