Thinking in Systems: A Primer

Thinking in Systems is not a book about AI. Meadows drafted it in 1993 and died in 2001, seven years before it was published.

It is a short book with an unusually wide lens. Most management writing hands you a team or a process to improve. Meadows hands you the whole company, the industry it sits in, and the reasons both behave in ways nobody designed and nobody wants.

We open leadership sessions by asking who has read it. Almost nobody ever has, which is surprising for a book you get through in an afternoon.

What is Thinking in Systems about?

Donella Meadows was a systems scientist at MIT and Dartmouth, and lead author of The Limits to Growth, the 1972 report that first modelled exponential growth against a finite planet. She spent the rest of her career pulling systems dynamics out of the academy and into language a decision-maker uses.

Her argument is that a system produces its own behaviour. Not the people inside it. The structure.

That claim is uncomfortable in proportion to how senior you are, because it means the recurring problem you keep assigning to different executives is a property of the design rather than a property of them. Meadows puts it plainly: if the same failure survives three changes of leadership, stop replacing the leader.

The book builds from a small vocabulary to that conclusion, then finishes with the part most worth your time, which is where to intervene.

What are stocks, flows and feedback loops?

Four ideas carry the whole book.

1. Stocks are the things that accumulate. Cash, inventory, headcount, technical debt, customer goodwill, unresolved decisions. You measure a stock at a moment in time.
2. Flows are the rates at which stocks fill and drain. Hiring and attrition. Sales and churn. Meadows uses a bath: the water level is the stock, the taps and the drain are the flows.
3. Feedback loops connect the two. A reinforcing loop accelerates whatever is already happening, in either direction. A balancing loop pulls the system back toward a goal and holds it there.
4. Delays sit between an action and its visible effect, and they are the reason systems oscillate. When feedback arrives late, you overcorrect, then overcorrect the correction.

From that she explains why systems overshoot, why they resist policy, and why they return to old behaviour after every intervention. She catalogues the traps by name: policy resistance, the tragedy of the commons, drift to low performance, escalation, success to the successful. Each one comes with a way out.

Then the part that matters most to a decision-maker. Meadows ranks the places you intervene in a system by how much difference intervening there makes. Twelve of them, weakest to strongest. Near the bottom sit constants, parameters and numbers: budgets, targets, prices, headcount. Near the top sit the rules of the system, its goals, and the mindset the whole thing grew out of.

Her warning is sharper than the list. People locate the right leverage point with good instinct, and then push it in the wrong direction.

Why this matters more in 2026 than it did in 2008

Meadows was writing when changing a large system meant changing what people did. Slow, expensive, and mostly attempted through reorganisation.

You now have a way to change what people know, and when they know it, at a speed the book never anticipated. That lands directly on one of her higher leverage points, and almost nobody is spending there deliberately.

Where AI sits in the system

Most AI budget lands at the bottom of her list

Run a typical programme against Meadows' ranking and it clusters at the weakest end.

Licences purchased is a parameter. Adoption targets are parameters. Seats allocated per function is a buffer. Rolling a tool into an existing process is a change to a stock-and-flow structure. All of these sit in the bottom third of her twelve, and she is blunt that intervening there produces very little movement for a great deal of effort.

That is a fair description of most AI programmes running today, and it explains the pattern you keep seeing: real effort, real spend, no change in how the business behaves.

Information flows are the exception

Meadows ranks the structure of information flows, who has access to what and when, well above every parameter, buffer and stock. Her point is that changing who sees what changes what people do, without anyone being persuaded of anything.

This is the one place AI reaches high on her list, and it is the argument for building at company level rather than handing out assistants. When a manager sees the churn signal in week one instead of quarter three, the decision changes without a single conversation about behaviour.

It also connects to her idea of bounded rationality. People act sensibly given the information available at their position in the system. Everyone behaves rationally and the whole still performs badly. Change what each position sees and behaviour follows.

The delay problem nobody plans for

Here is the counterintuitive one, and it is the reason some AI deployments make things worse.

AI collapses the delay between question and answer. It does nothing to the delay between decision and approval, or between approval and action. You have shortened one delay in a chain and left the others untouched.

Meadows would predict exactly what follows. Mismatched delays make a system oscillate harder, because you are now reacting faster to signals the organisation still cannot act on. Speed applied to one link of a chain that still has slow links produces frustration, not throughput.

Three questions for the executive team

For the CEO. Which of the twelve leverage points is your AI programme actually pulling? If the honest answer is budgets, targets and licence counts, you are working at the bottom of the list and should expect what the bottom of the list returns.

For the CFO. What stock is this investment meant to change, and over what period? Meadows' discipline is that you name the accumulation you are moving. An investment that cannot name its stock is buying activity.

For the COO. Where are your delays, and which one did you just shorten? If the answer is "time to insight" and the approval chain is untouched, you have built pressure rather than throughput.

Where the book is weaker

Two things to know before you buy it.

The examples are ecological, not commercial. Meadows came from The Limits to Growth, and her illustrations run to fisheries, populations, thermostats and oil reserves. The reasoning transfers cleanly to a business. The pictures do not, and you will be doing that translation yourself on every page.

It describes rather than prescribes. The book is superb on how systems behave and thin on what you do about it on Monday morning. Her leverage points essay is honest that the higher up the list you go, the harder change becomes, and she offers little on how to actually move a goal or a paradigm inside a real organisation with real politics.

Who should read it

CEOs and COOs, and anyone who keeps solving the same problem.

Read it if you have watched a well-run pilot succeed and change nothing at company level. Meadows explains that outcome better than any book written since about AI.

What you will take away

• A way to draw your company, and the industry around it, as stocks, flows, delays and loops
• A ranked list of where intervention pays, and a way to check where your own spending sits on it
• The reason automating a broken process makes it fail faster
• Language for the recurring problems your business has stopped noticing

The line worth remembering

Every system is perfectly designed to produce the results it currently gets. Adding AI to a system you have not redesigned buys you the same results, sooner.