Systems Thinking for AI Projects: Why Fixes Come Back and What to Do Instead

You fix a problem, and it comes back. You hit a target, and the thing the target stood for gets worse. If you run a team, a service or a rollout, you have seen this. Systems thinking gives you a few simple tools to understand why. AI makes these tools more useful. It speeds up the fast parts of a system, but the slow parts stay slow.

Organizations are not machines

A machine does the same thing each time you pull the same lever. An organization is made of people. They respond, adapt and remember. A quick fix aimed at a symptom often moves the pressure somewhere else, and the first symptom returns. The better question is not "Who caused this?" It is "What structure keeps producing this?"

Stocks and flows

A stock is anything that builds up, and you could take a snapshot of it. Examples are cash, trained staff, open tickets and customer goodwill. A flow is how fast the stock fills or drains. You cannot change a stock directly. You can only change its flows, and then you wait.

Picture a bathtub. Turning up the tap does little if the drain is open. Closing the drain raises the water even if the tap stays the same. Managers are often given targets for flows, like hires this quarter. But they are judged on stocks, like whether the team can deliver a year from now.

Reinforcing and balancing loops

A reinforcing loop makes a change bigger. Happy customers tell others, which pays for better service. The same loop can run the other way. Two people quit, the rest work longer hours, and more people quit. A balancing loop pulls a system back toward a set point. That is why systems often resist well-meant fixes. When a problem keeps coming back, look for the loop that grows it and the loop that fights your fix.

Delays cause overshoot

Say a new hire takes months to become useful. Then the effect of your decision shows up long after you make it. Leaders who expect fast results often add more of the fix, and then they overshoot. Write the delay into your plan. Then judge the change by the time it truly needs.

When a measure becomes a target

The economist Charles Goodhart made a point that people usually sum up this way: when a measure becomes a target, it stops being a good measure. This hits AI projects hard. AI tools are very good at hitting whatever number you give them. They do not care about what the number left out. The fix is to pair each speed or volume number with a quality number. Report them together, to the same people. Make sure the quality number can change real decisions. Closing more tickets means little without the repeat-contact rate beside it.

Ask "and then what?" three times

Before you approve an automated rule or a new tool, trace three ripples. What changes? What does that cause? What will people do in response? Assume they will adapt. The people with the most tries and the most to gain learn an automated rule fastest. Put a review date on the rule.

A first exercise

Pick one problem that has come back more than once. Sketch its numbers over the past year. Name the stock behind it. Write down the loops and the delays. Then decide where you would step in. A change to a rule or to who sees what information often moves a system more than a change to a budget.