New to control systems, or need a refresher? Start with Control Systems: A Primer.
My day job is controls engineering. Over the last year I noticed I was using the same mental model on everything else too. The outreach pipeline, a voice agent, my own finances and the AI agents that run my businesses all ended up drawn as the same diagram: a sensor, a setpoint, a controller and a plant, with a wire running from the output back to the input.
A process without that wire is open loop. You set the input and hope. Most of what goes wrong in a business looks like that: a campaign launched and never measured, money moved on a feeling, an agent given permissions because the demo looked good. Closing the loop isn't complicated. But once I started doing it everywhere, four rules kept deciding things for me.
1. Measure before you tune
You don't tune a loop you can't see. When I built a voice agent for screening phone calls, the first version felt slow, and I could have guessed at the cause. Instead I instrumented every stage of a response first: speech-to-text, end-of-turn detection, the language model's first token, and the first audio out. With the numbers in front of me, the fixes were obvious, and the median response time dropped by about 30 percent. Guessing would have had me tuning the wrong stage.
2. Know your delays
Dead time is what makes loops ring. Ad platforms report conversions days after the spend. Savings take days to turn back into cash. A prospect replies a week after the email. Every one of those delays limits how hard you can react. If you scale ad spend on last week's numbers without waiting out the attribution lag, you overshoot, then overcorrect, and you have built an oscillator. I wrote out the math in Money as a Control System. Each period of delay tightens the stable range of the gain.
3. Raise the gain in steps
On the plant floor, a new loop comes online in manual. Then it runs in auto with conservative tuning, and it tightens only once the trend data shows it holding. I now treat capital going into a business the same way: first a fixed test budget to identify the plant, then more only while the measured return stays positive. I treat AI agents the same way too. Each kind of action they take starts out needing my approval and earns autonomy on its own track record. The details are in Agent Autonomy as Loop Gain.
4. Keep the hard limits outside the controller
Safety interlocks don't live in the program the operator can edit. Some limits in my systems sit where the controller can't touch them: no debt or retirement money goes into ads, an agent never gets to spend money or contact a customer on its own, and agents can't edit their own rules. A controller that can rewrite its own limits doesn't have limits.
None of this is new. It is first-semester controls, applied outside the plant. What surprised me is how often the hard part of a business decision turns out to be a measurement problem in disguise. Once the loop is closed, the decision usually makes itself.
Yours Truly