AI has given a lot of people a new way to misunderstand throughput. The pattern is familiar by now. You automate some step in a process, output jumps 10x, and suddenly everything starts piling up somewhere downstream. Then comes the observation: congratulations, you moved the bottleneck. Maybe. I suspect in a lot of cases the bottleneck did not move at all. You just produced enough upstream capacity to expose the constraint that was already there. Take creative production in a regulated industry. Before AI, maybe the team could make ten viable creatives in a week and compliance could comfortably process them. Now you can generate a hundred. Excellent. You have solved creative production. Unfortunately, compliance remains in possession of linear time. The same review process still exists. The same documentation and approvals still exist. The fact that nobody cared about their capacity before does not mean they suddenly became the constraint when AI showed up. The same thing happens in software. If implementation gets 10x faster but review, testing, security, deployment, or adoption stay slow, faster implementation may mostly give you a larger pile of work waiting to become useful. You have increased inventory more than throughput. Theory of Constraints has been making this point for decades. Find the constraint. Exploit it. Subordinate the rest of the system to it. Elevate it. Then find the next one. The uncomfortable part is the subordinate step. If compliance can review fifteen creatives a week, maybe producing a hundred is not useful. If a team can thoughtfully review twenty PRs, opening two hundred is not productivity. It just makes the queue longer. And this is where AI gets more interesting than the usual “what can we automate?” conversation. Once one part of a process becomes radically cheaper, you get an excuse to interrogate everything around it. Does every item need the same approval path? Why are these steps sequential? Why does this handoff exist? Which parts are actually required, and which parts are organizational sediment? Maybe the answer is to automate the constraint. Maybe it is to redesign it or remove half the process. But before any of that, you need to know what output you actually care about. If you cannot say what the system is trying to produce, “more automation” has no natural stopping condition. You can keep adding agents, schedules, triggers, generation, and orchestration because nothing is telling you whether any of it matters. At that point you are optimizing activity against a void. If your first question is “where can I add AI?”, you may be starting too far downstream. First ask: “what actually constrains this system from producing more of the thing I care about?”
Back to all writing ›