The Company That Learns Every Day

VISPAICO Journal, Issue 024

---

Before 1788, running a steam engine at a steady, safe speed required a person to stand near it, more or less continuously, watching the pressure and adjusting the throttle by hand. Left unattended, an engine under a heavy load would slow dangerously; relieved of that load suddenly, it would race, sometimes catastrophically, faster than its own machinery could safely tolerate. The engine had no way of knowing anything about its own condition. It simply did whatever the steam pressure and the valve setting told it to do, moment to moment, indifferent to whether the result was safe.

James Watt's solution to this problem is one of the more elegant inventions in the history of engineering, and it is worth understanding in some detail, because the logic behind it applies with unusual precision to a mistake most companies are still quietly making. Watt attached a pair of weighted balls to a spinning vertical shaft connected to the engine itself. As the engine sped up, the spinning weights flew outward under centrifugal force, and this outward motion was mechanically linked to the steam valve, closing it slightly. As the engine slowed, the weights fell inward, and the valve opened again. No person needed to watch a gauge. No person needed to intervene at all. The engine had become capable, for the first time, of noticing its own deviation from the correct speed and correcting for it, continuously, in real time, without waiting for anyone to check.

This single mechanism, the centrifugal governor, did something that no amount of careful periodic supervision had ever managed to do reliably. It closed the gap between deviation and correction down to almost nothing, and in doing so, it made the steam engine trustworthy enough to power an industrial economy that could never have scaled on manual supervision alone.

Most companies today are still running on the supervision model Watt's governor replaced. They have simply dressed it up in more modern language.

---

What a KPI Actually Is

A key performance indicator, examined honestly, is functionally very close to the person standing beside the pre-governor engine, checking the pressure gauge at scheduled intervals and adjusting the valve by hand. A monthly sales report. A quarterly review of customer churn. An annual assessment of operational efficiency. Each of these represents a deliberate, useful, and entirely reasonable attempt to notice when something has drifted away from where it should be, and each of them, by design, only looks at the moment someone remembers to check.

This is not a criticism of the practice, which has served organisations reasonably well for a very long time, in the same sense that manual supervision kept early steam engines from destroying themselves reasonably well, most of the time, when the person watching happened to be paying close attention. The limitation was never a failure of diligence. It was structural. A gap always exists between the moment a deviation begins and the moment someone with the authority to notice it actually looks, and during that gap, whatever is drifting continues drifting, entirely unattended, regardless of how conscientious the eventual review turns out to be.

Companies have absorbed the cost of this gap for so long that it rarely registers as a cost at all. A customer satisfaction metric that only gets reviewed quarterly means three months can pass, in the worst case, between the moment something genuinely starts going wrong and the moment anyone with the ability to fix it actually finds out. A pricing anomaly that only surfaces in a monthly report can quietly cost real revenue for weeks before anyone notices the pattern. None of this looks like negligence from the inside. It looks like a perfectly normal management rhythm, because it is one, the same rhythm every organisation has operated under for as long as management reporting has existed. It simply happens to leave a gap that Watt's governor was specifically invented to close.

---

The Difference Between Correcting and Learning

It would be a mistake, however, to conclude that the solution is simply a faster, more continuous version of the same KPI dashboard, real-time metrics instead of quarterly ones, watched, in effect, by a tireless observer who never blinks. This gets partway to the right answer, but it stops short of the more important distinction, and Watt's own governor is useful for illustrating exactly where it falls short.

The centrifugal governor was a genuine breakthrough in continuous correction, but it was not, in any meaningful sense, a learning system. It corrected the same deviation, using the same fixed mechanical rule, for as long as the engine ran, indifferent to whether conditions had changed in ways that made the original rule less appropriate than it once was. A governor calibrated for one kind of load would apply exactly the same correction to a very different kind of load, because it had no capacity to notice that the situation itself had changed, only that the speed had drifted from a fixed target. It could keep an engine steady. It could never make the engine, or itself, any wiser about the conditions it was actually operating in.

This is precisely the distinction worth drawing between mere continuous monitoring and something that actually deserves to be called a learning loop. Continuous monitoring closes the gap between deviation and correction, which is valuable and considerably better than periodic review, but it applies the same correction rule indefinitely, regardless of whether that rule remains the right one. A genuine learning loop does something the governor structurally could not: it uses each correction as an opportunity to refine the rule itself, so that the response to next month's deviation is measurably better calibrated than the response to this month's, because the system has actually accumulated something from having encountered the pattern before.

---

What This Looks Like Away from the Abstraction

It is worth grounding this distinction in something concrete, because the difference between correction and learning sounds subtle in the abstract and is anything but subtle in practice. Consider a company monitoring customer complaints about a particular product feature. A continuously monitored system, in the governor's sense, would notice a spike in complaints quickly and route it to the right team faster than a monthly report ever could, a genuine improvement, closing the gap Watt's engine once suffered from.

A genuinely learning system does something further. It notices, over time, which categories of complaint tend to precede a larger, more serious pattern, and which tend to resolve themselves without further escalation, refining, gradually, its own sense of which early signals actually deserve urgent attention and which don't. The correction itself gets better calibrated with every cycle, the way an experienced operations manager's instincts get sharper with every quarter of accumulated pattern-recognition, except available consistently, without depending entirely on that one manager remaining in the role long enough to build the instinct personally.

This is the meaningful difference between an organisation that has simply made its dashboards faster and one that has genuinely built a learning loop. The first notices problems sooner. The second gets measurably better, cycle after cycle, at knowing which problems actually matter and how to respond to them, an accumulating capability, rather than a faster version of the same fixed response, applied indefinitely without ever improving.

---

Why This Requires Rethinking the Review Cycle Itself

None of this is achievable by simply asking existing teams to check their dashboards more frequently, and it's worth being direct about why, because the instinct to solve this with more frequent meetings is both common and largely unproductive. A weekly review is still, structurally, a periodic review, a person, checking a gauge at a slightly shorter interval, still bound by the fundamental limitation the governor was invented to eliminate entirely: a gap, however narrow, between deviation and detection, and a static rule applied identically regardless of what has actually been learned since the last time anyone looked.

The organisations that build genuine learning loops are doing something categorically different from scheduling more frequent meetings. They are building the organisational equivalent of Watt's mechanical linkage, a structural connection between deviation and correction that doesn't wait for anyone to schedule a review at all, paired with something Watt's engine never had: a mechanism for the correction itself to improve, cycle after cycle, based on everything the organisation has already encountered. This is a genuinely different kind of management discipline than reviewing KPIs, however diligently, and it requires leadership to think about performance management less as a calendar of scheduled check-ins and more as an ongoing structural property of how the business actually operates, continuously, whether or not anyone happens to be in the room.

---

What Trust Actually Requires

It is worth remembering what Watt's governor actually made possible, beyond the narrow technical achievement of keeping an engine at a steady speed. It made the steam engine trustworthy enough to be left running without constant supervision, which meant, in practical terms, that industry could finally scale past what a limited supply of vigilant human attendants could ever have supported on their own. The governor didn't just solve a mechanical problem. It solved a trust problem, and trust, once genuinely earned, is what allowed the steam engine to become the foundation of an entire industrial economy rather than remaining a powerful but perpetually supervised curiosity.

Organisations face a similar threshold now, and the trust they are trying to earn is structurally the same kind Watt's governor once earned for the steam engine. A company that can only notice and correct its own deviations on a quarterly rhythm will always be, in some meaningful sense, running unsupervised for most of the time that actually matters, trustworthy in the narrow windows someone happens to be watching, and simply drifting, unremarked, everywhere else. A company that has genuinely built a continuous, improving learning loop earns something considerably more valuable than a faster dashboard. It earns the right to be trusted at a scale and a pace no quarterly review, however carefully conducted, could ever have supported on its own.