Lag, Gain, and Damping in a Business
This part surprised me. The same four parameters describe a person deciding under pressure, a manager running a team, and a whole company steering itself. Not by analogy either. Same mechanics, because they're all regulators with sensors, targets and lag.
If you've read Part 1 you already have the vocabulary. If you haven't, you need six words: set point, sensor, error, gain, damping, delay.
Why do good marketing changes look like failures?
Because the loop gets run faster than the system can answer. A change goes in Monday and results land in four to six weeks. The owner checks Wednesday, sees nothing, and changes it again. By the time the first change would have surfaced, three more are layered on top and no result can be attributed to anything. The campaign didn't fail. It was replaced before it could report.
Go back to the person holding the thermostat dial in Part 1. They aren't stupid and they aren't over-emotional. They're acting on a system whose response arrives later than their expectation of it, and they don't know that yet.
That's the single most common failure I see in businesses.
The fix isn't patience as a virtue. It's damping as an engineering decision. Find out what the system's response time actually is, then set your review interval longer than that.
A change reviewed on a shorter cycle than its own feedback delay can't be evaluated. It can only be replaced.
What are the two ways a leader "overreacts"?
They're different failures with the same complaint attached, and they need opposite repairs. An accurate read with too much gain means the perception is correct and the response is oversized. An inaccurate read held with high confidence means the response is proportionate to a picture that's wrong. From outside, both look like somebody overreacting, and both get the same unhelpful advice.
Take them separately, because the intervention flips.
An accurate read with too much gain. A leader correctly notices a real problem and responds at a magnitude the problem doesn't warrant. The estimate is right. Telling them they're overreacting to nothing is both wrong and expensive, because they aren't reacting to nothing.
What needs to come down is the amplitude, not the perception. And reducing controller gain doesn't mean the response stops. Keith Richards is 82 with arthritis and books a residency instead of a world tour. The output continues. The amplitude comes down.
An inaccurate read held with high confidence. Someone certain about a market, a competitor or a person, on thin evidence, who doesn't update when contradicted. Calming them down does nothing here. What moves it is a test with a written-down prediction.
There's a reason that works and pep talks don't. Behrens and colleagues showed people adjust how fast they learn to how volatile the world is, which means the update machinery is evidence-driven by design (Behrens et al., 2007). Give it evidence and it moves. Give it encouragement and it has nothing to work with.
Same complaint from a colleague, he overreacts, two different repairs. Ask which one you're looking at before you intervene.
Why does micromanagement make output worse?
Because checking every two hours runs the loop far above the rate at which the work produces signal. Every check generates an error, every error generates a correction, and the corrections arrive faster than the work can absorb them. Output starts oscillating, the manager reads that as unreliability, and the checking rate goes up. The loop is the problem rather than either person in it.
Lengthen the interval and the oscillation damps out on its own. Nothing about the employee has to change.
That's hard to do, because it means acting less at the exact moment the system looks worst. Which is the same thing that makes response prevention hard for a client, and it fails for the same reason: the correction feels necessary precisely when it's most destabilizing.
Why are most KPIs set wrong?
Because a KPI is a set point, and nobody treats it like one. A target chosen because it sounds ambitious still gets defended as though it described something real, and the number reporting on it is almost always a stand-in for what actually matters. Get those two wrong and every correction downstream is aimed at the wrong thing, confidently, on a schedule.
None of that is a metaphor. Carver and Scheier made the case forty years ago that feedback control describes self-regulation at the level of a person (Carver & Scheier, 1982), and an organization steering by numbers is running the same architecture with slower sensors and worse reporting.
Five failure modes, and they compound.
The set point is arbitrary. A number picked because it sounds ambitious. The system will defend it anyway.
The sensor measures a proxy. Plasma osmolality is not thirst, and monthly revenue is not business health. Both get defended as though they were the thing.
The delay is unacknowledged. Reporting lag means every correction is aimed at where the business was, not where it is.
The gain is too high. One bad month triggers a response sized for a bad year.
There's no damping. Strategy changes every quarter, and no strategy runs long enough to produce evidence about itself.
A company with those five settings looks chaotic, and it usually gets described in terms of culture and personality.
It's a loop with the parameters set wrong. Parameters you can adjust. Culture you mostly can't.
Does the self-confirming loop show up in business too?
Exactly as it does in a session, and it's expensive. A belief that a segment won't buy produces reduced effort with that segment, which produces poor results, which confirms the belief. The action prevented the evidence. Nobody was irrational at any point, and the belief is now backed by data the belief itself generated.
Part 4 walks the clinical version of this at a dinner table. The commercial twin has identical structure and a larger invoice.
The test is the same in both places. State the prediction in advance, run the experiment without the safety behavior, and check the result against what you wrote down rather than against how it felt.
That last clause is doing more work than it looks like. Checking against how it felt is how a loop stays closed for a decade.
What should you do differently in a hard decision?
Three things, and all of them get set before the pressure arrives. Separate what you think is happening from how hard you're going to act on it, and say them as two sentences. Decide your review interval in advance, because under stress you'll set it too short. And write down what would change your mind before you have the answer, which converts a belief into a test.
None of that is new advice. Every one of those three shows up in some form in books that never mention control theory.
What the model adds is why they work, and that matters at the moment you least want to follow them.
The written-down prediction in particular is the same instrument a clinician uses in a behavioral experiment, and it works for the same reason: expectancy violation, not repetition, is what updates a belief (Craske et al., 2014).
Most bad decisions under pressure are an accurate read with a disproportionate response, and they get defended by pointing at the accuracy of the read. I was right about the problem is true and is not an argument for the size of what you did about it.
The last piece in this series is the one I'd read most skeptically if somebody handed me all seven. It lists what would show the whole framework is wrong.
References
Behrens, T. E. J., Woolrich, M. W., Walton, M. E., & Rushworth, M. F. S. (2007). Learning the value of information in an uncertain world. Nature Neuroscience, 10(9), 1214–1221. https://doi.org/10.1038/nn1954
Carver, C. S., & Scheier, M. F. (1982). Control theory: A useful conceptual framework for personality–social, clinical, and health psychology. Psychological Bulletin, 92(1), 111–135. https://doi.org/10.1037/0033-2909.92.1.111
Craske, M. G., Treanor, M., Conway, C. C., Zbozinek, T., & Vervliet, B. (2014). Maximizing exposure therapy: An inhibitory learning approach. Behaviour Research and Therapy, 58, 10–23. https://doi.org/10.1016/j.brat.2014.04.006
The full documents, free and ungated