Michael Zazon

Engineering leadership · Book notes

Thinking in Systems says headcount is the weakest lever

Donella Meadows ranked twelve leverage points, and the ones transformation programs reach for first are at the bottom.

2 min read

Thinking in Systems is a primer on stocks and feedback loops by environmental scientist Donella Meadows. It was published posthumously by Chelsea Green in 2008, but it continues to make the rounds on engineering and strategy reading lists. The value is in a single ranked list, adapted from her original essay Leverage Points: Places to Intervene in a System. The list ranks twelve places to push on a system, from least to most effective. You might be surprised which familiar levers rank low, and how far at the bottom you’ll find the numbers we usually adjust first. (“99 percent of our attention goes to parameters,” she writes.)

This list makes me uncomfortable when I think about it in terms of an enterprise transformation program. Parameters are headcount and budgets.. the dials leadership can turn from a slide, so that’s what programs turn. Higher up are information flows (who sees cost and incidents), rules (incentives and approval paths), and the goals the system is really optimizing for. I think most enterprise transformations fail at the leverage-point level before they even get to execution.. they change some numbers but leave the structure. Meadows also observes that the real leverage point in a company is often already being pushed on, just in the wrong direction.

Twelve leverage points ranked weakest to strongest, programs at the bottom, AI adoption loop Most leverage Where the leverage is 1 to 3 Paradigms, goals 5 Rules 6 Information flows In between 7 to 10 Loops, delays Where programs push 11 Buffers 12 Headcount, budget, licenses Least leverage AI adoption loop Results Investment reinforces More of what exists

Meadows talks about how cheap wins hide in information flows.. that adding (or restoring) information is often more powerful than rebuilding infrastructure. In my own homelab, I had an auto-restart loop that kept restarting a container with a perfectly fine service because the healthcheck was testing the wrong thing. I fixed it by fixing the information the loop was acting on, not the loop itself. (Autoheal opt-in and healthchecks that actually test something)

This lens also applies to AI. The 2025 DORA report shows how AI amplifies existing state.. the returns track internal platform quality and clear workflows, not how broadly you rolled out the tools. In Meadows’s terms, I’d describe AI adoption as a reinforcing loop that amplifies whatever it starts with. Organizations that already changed their rules and information flows get better results and invest more. Everyone else mostly buys more seats. If you’re leading an AI program, which dial is it turning?