I hated today's most important book for leaders. I was forced to read The Goal, by Eli Golderatt, and I hated it. I think it may be the most valuable possible book for leaders right now. The Goal is about optimizing factory throughput. My boss, coming from running an aircraft lighting plant, wanted me to read it. As a software engineer, I thought reading about factory operations was a complete waste of time. I was wrong. I was wrong then, and more wrong today, because AI is a nearly perfect example of the problem the book addresses. The book teaches a simple idea - the "theory of constraints" - that an assembly line can only produce at the speed of its slowest step. The idea here is simple. Henry Ford famously made the Model T, his early best seller, only in black. Why? Black paint dried the fastest. By using only black, he could paint more cars more quickly. Perhaps Ford could have sped up engine or chassis assembly. But if the bottleneck was at the paint booth, all that would do is lengthen the line waiting to be painted. AI makes certain tasks much faster. But companies have not yet converted AI investment into actual business results. This book explains why. AI allows fewer engineers to crank out more code. It also accelerates some parts of testing, marketing copy generation, product prototyping, and many other things. If AI writes software much faster, but all the other parts of your company remain the same, the bottlenecks elsewhere will still hold things up. To get more results, you have to redesign the whole process, always focusing on the bottleneck steps. The CEOs counting token use are missing this point. Getting their teams to use AI does matter. But for the next few years the challenge for leaders is also going to be finding the bottlenecks where either AI does not apply at all or no one has figured out how to use it, and focusing there. Faster coding won't magically create better software products. They also need both architecture and UX design. But beyond that, they then also need sales, marketing, support, documentation, and possibly all kinds of messy physical systems where AI is currently no help at all. Take Amazon as an example. AI doesn't create more trucks or delivery drivers. It might make their routes slightly more efficient, but that isn't a big change. So all the AI in the world on the website, helping people shop and buy is good, but it won't matter unless some leader figures out trucks and warehouses to support the new scale. Even if AI helps the leader come up with the plan, you have to buy the trucks and hire the drivers. For big leaders over the next 2 to 3 years the lessons of "The Goal" can help you solve the real problem - not "AI adoption" but "more actual market results because of AI." Where do you see fast AI prototypes hitting a wall of other necessary steps, slowing the features back down to "regular" speed?
I read The Goal in B-school and was thinking about it just this week. One of my team's goals is to figure out how effective AI coding agents actually are, and the deeper I go, the more I keep thinking about Goldratt's book: pick the right measures, measure the process, find the bottleneck, fix it, then watch the constraint move somewhere else. The piece people skip is building slack into the system so it doesn't snap when you speed one step up. Slack also matters for measurement: if every step is running flat out, you've got no room to instrument the process and capture the data that tells you where the constraint actually is. On the tech side, the bottleneck I see is how far the agent can run on its own, writing tests, running them against the code, and fixing what it finds with no human in the loop. That's more than TDD. It drags in questions like whether you've got the resources to spin up clean environments and test against production-grade data. On the non-tech side, it's the upstream approval process. Prototyping a feature is fast now. Getting it launched, or getting the resources to actually build it, still runs through human judgment and sign-offs. That's where the speed leaks back out, exactly the wall you're describing
This reminds me of the Jeff Bezos and Jeff Wilke conversation. Jeff Wilke, coming from manufacturing background, perfectly knew that adding "more" to the queue simply slows down or even halts the production line. He famously said to Bezos "You have enough ideas to kill Amazon"! Fortunately Bezos listened. Will the current CEOS follow the same advice in the age of AI?
In my top 5 (out of 1100+) of non-fiction books I have read.
At some level, the most common mistake is failing to distinguish between friction and resistance. Friction is a limitation that reduces speed, destroys materials, etc., for no benefit, whereas resistance is a desirable slowdown that helps avoid catastrophic errors and limits the system's throughput to keep it running smoothly. The next level is knowing how to embed a self-throttling mechanism with feedback loops. This is why the current version of AI remains mediocre: LLMs are not designed to understand feedback loops.
I think you’re out of touch. The AI progress re enables the idea of one person as a craftsman, versus needing scale. All the sub functions merge. Progress is at least 10x more. Where you might be correct is if Amazon senior leaders are slowing progress to their own intellectual pace. Probably true for any large company. And I do agree counting tokens is dumb. Output matters most.
Love this framing. My instinct is that for many organizations, the constraint isn't the tools, it's the moment where someone has to stop and truly understand what it means. AI can surface more potential insight than ever, but insight only becomes a decision when someone actually absorbs it. More data without the capacity to absorb it doesn't compound, it overwhelms and confuses. The bottleneck isn't information anymore, it's comprehension. Leaders who recognize that and create space for real synthesis are going to unlock something genuinely powerful. The Goal has never felt more relevant.
This resonates deeply with me. I spent ~15 years in supply-chain software working with organizations such as Motorola, Flex, and sustainability-focused supply-chain startups. One lesson that keeps repeating is that local optimization rarely produces system-wide gains. AI is accelerating coding, research, prototyping, and testing, but the Theory of Constraints still applies. The bottleneck simply moves. After validation comes implementation, change management, customer onboarding, support, operations, governance, maintenance, and scaling. These second- and third-order effects often determine whether a product succeeds. In supply chains, improving one workstation doesn’t improve throughput if another station remains constrained. AI adoption appears to be following the same pattern.
The goal and its associated movie were required reading at i2 technologies. The TOC always gets me even though it’s so intuitive. Making the slowest trooper the front of the line - to me this is like - If the bottleneck is data quality, don’t route around it with a RAG demo on clean data; point the AI at the dirty data and make the constraint visible.
I've been pondering how much users/customers can absorb change as a bottleneck. Eventually if AI agents themselves are a large percentage of the user base for software perhaps it doesn't matter but for now remains relevant. Maybe not so much as a bottleneck but a brewing issue I am trying to get ahead of is the over-focus on speed of shipping new changes and the assumption that once you ship it you are "done". That the initial act of building was the hard part. I am concerned that without adequate frameworks for managing the growing complexity in our systems and knowledge gaps those rapid fire changes (authored mostly by AI) create...all the "ilities" seem at risk (maintainability, extensibility, etc). Again, if I pull the thread further I can see how one day perhaps AI handles self-healing, managing complexity down, etc. But we are in the dangerous liminal space it seems where none of that has caught up. P.s. Ordered The Goal! Thanks for the recommendation.
The big misunderstanding, in a lot of companies, is that “software development” was the bottleneck. And that’s because the Lead Time - the full value chain, how long it takes for an idea to be turned into something that a customer pay for - has been rarely measured. Many reasons, the biggest one for me are the individual department KPIs which foster and encourage sub-optimization. And, when you sub-optimize your system, the performance overall is not improving. It stay the same, if not degrading.