The GenAI Paradox: More Isn’t Better

Introduction

In Eli Goldratt’s The Goal, the introduction of a robot into a manufacturing plant seemed like a breakthrough. It ran constantly, churned out parts, and looked highly productive. But there was a catch: it didn’t actually increase the plant’s overall output. Instead, it created bottlenecks, piled up inventory, and disconnected production from actual demand. The key insight? Local efficiency does not equal system effectiveness.

Now, apply that to the modern workplace - except the robot is a large language model (LLM).

LLMs Are Today’s Super-Robots. They produce content, draft emails, write code, answer questions, and generate summaries. They never sleep, never slow down. On paper, they look incredibly efficient and productive. But just like Goldratt’s robot, an LLM’s productivity is meaningless unless it contributes to what actually matters.

An AI’s output only matters if it helps humans make better decisions, faster.

And here’s where it gets interesting: the real constraint isn’t the LLM’s capacity. It’s ours.

The NCX-10 Scene: Local Efficiency vs Global Goals

In Eliyahu M. Goldratt's The Goal, there's a pivotal scene where Alex Rogo, the plant manager, evaluates the productivity of a newly implemented robot, the NCX-10. Initially, Alex believes the robot has enhanced efficiency by 30% and reduced costs. However, during a conversation with his former professor, Jonah, Alex is challenged to reconsider this assessment. Jonah questions whether the robot has led to increased shipments, reduced operational expenses, or decreased inventory levels.

“…productivity is the act of bringing a company closer to its goal. Every action that brings a company closer to its goal is productive. Every action that does not bring a company closer to its goal is not productive.” Jonah to Rogo, The Goal

The robot converts 25 parts per hour, but the upstream process doesn’t always deliver 25 parts—sometimes it's fewer. The result? The robot ends up waiting or producing unneeded parts. Through multiple dimensions—parts completed, time to completion, production steps—we witness the impact of dependent events and statistical fluctuations.

Alex realizes that, despite the robot's high utilization, there hasn't been a corresponding improvement in the plant's overall performance. This insight underscores a central theme of the book: local efficiencies, such as maximizing machine utilization, don't necessarily translate to global improvements in throughput or profitability. The scene illustrates the importance of aligning all processes with the organization's primary goal - in this case, making money, rather than focusing solely on isolated performance metrics

Cognitive Control: The Human Bottleneck

Cognitive control is the brain’s ability to focus, prioritize, and manage mental resources to focus on relevant information while suppressing distractions. It’s what allows an engineer to monitor live data streams or a strategist to make split-second decisions. But this control operates on a shockingly low bandwidth: about 3 to 4 bits per second.

To understand how little that is, consider that 1 bit of information is enough to answer a single binary question - a dilemma that can be resolved with a simple yes or no, or option A or option B. That’s it.

So even if your LLM can generate a 2,000-word report in under a second, you - its human partner, can only process that report in narrow, focused slices. And if the information isn’t clear, relevant, or aligned with your goal? Those bits are burned in confusion.

The Driving Analogy: Why One Bit Matters

Picture yourself driving in a foreign country. One of the first things you need to know is: Which side of the road should I drive on?

If the rule is clear - "left side in the UK," "right side in Germany", you don’t even think about it. But if it’s ambiguous or switches suddenly, your brain has to use its precious bandwidth just to resolve that question.

That’s one bit gone.

Now extend this to digital tools, AI assistants, workflows. Imagine your GPS, dashboard, and road signs all use different units, different symbols, or different conventions. More bits lost. Before you even get to the driving itself - navigating turns, avoiding pedestrians, responding to other drivers, you’ve already spent a chunk of your mental capacity just figuring out how to interpret the environment.

The same thing happens at work. If you have to constantly re-interpret what a tool is telling you, search for hidden settings, or decode inconsistent outputs, your cognitive control is being drained. And that’s before you even start the task that matters.

The LLM Paradox: More Output, More Overload

An LLM that gives you 100 suggestions might feel powerful, but in truth it’s flooding your 4-bit narrow cognitive channel. Productivity doesn’t come from more possibilities - it comes from sharper focus.

True AI productivity isn’t about more information. A good cognitive assistant should act like a compression algorithm for your brain:

Reduce ambiguity
Highlight the relevant
Frame the decision
Make the next action obvious

If the assistant isn’t doing that, then it’s just another NCX-10 robot - working hard but moving you nowhere.

The Cognitive Productivity Metric

The true productivity of an LLM - or any cognitive tool, should be measured not by how much it produces, but by how well it aligns with the limits of human cognitive control.

We need to rethink what productivity means in a world of human-AI collaboration. It’s no longer about how many documents are produced or lines of code are written.

Instead, we should ask:

How many bits of cognitive load did this tool save?
How many important decisions did it help clarify?
Did it compress complexity—or add to it?

If it overwhelms the human with 100 possibilities instead of sharpening focus down to one or two relevant distinctions - it’s no different than the robot overproducing inventory and clogging the line.

Final Thought

In Eli Goldratt’s The Goal, the real constraint in a factory wasn’t a machine—it was how well the entire system aligned with the goal. Now, as we integrate AI like LLMs into knowledge work, the bottleneck isn’t output. It’s cognitive control: our brain’s limited ability to filter, prioritize, and act.

The true productivity of an LLM - or any cognitive tool, should be measured not by how much it produces, but by how well it aligns with the limits of human cognitive control.

Helping a person choose this over that - with clarity, confidence, and speed, is worth far more than a thousand generated options.

The future isn’t about maximizing machine output. It’s about aligning that output with the real bottleneck: the human mind.

Because when every bit counts, the best tools are those that help us spend fewer bits to get to what matters.

Reference

1. Wu, T., Dufford, A. J., Mackie, M. A., Egan, L. J., & Fan, J. (2016). The Capacity of Cognitive Control Estimated from a Perceptual Decision Making Task. Scientific Reports, 6, 34025.

2. Zheng, J., & Meister, M. (2024). The unbearable slowness of being: Why do we live at 10 bits/s?. Neuron.