Measuring changes the system: Goodhart in the wild

When a measure becomes a target, it ceases to be a good measure. That is Goodhart's law, and it means any KPI you attach a reward to will gradually stop describing reality, because people optimise the number rather than the work the number was meant to represent. You cannot switch this off — but you can design around it by pairing every volume metric with a quality metric, putting targets on outcomes rather than proxies, and auditing what your best-looking numbers have started to hide.

Where does the law come from?

Charles Goodhart, a British economist, made the original observation in 1975 about monetary policy: any observed statistical regularity tends to collapse once pressure is placed on it for control purposes. The Bank of England would find a money-supply measure that correlated with inflation, target it, and watch the correlation dissolve.

The systems-thinking reading is the useful one. A business is not a machine with gauges; it is a system full of adaptive agents, and the people inside it respond to whatever you measure them on. Donella Meadows made the general point in Thinking in Systems: the act of intervening changes the system's behaviour, and information flows are among the strongest leverage points a founder has. A published target is an information flow with an incentive bolted on. It will change behaviour. The only question is whether it changes the behaviour you wanted.

What does Goodhart look like in a B2B firm?

You have almost certainly seen it, even if you did not have a name for it.

• Measure salespeople on calls made, and call volume rises while call quality falls — short calls, wrong numbers redialled, voicemails logged as conversations.
• Measure marketing on leads generated, and the definition of a lead quietly loosens until sales stops trusting the list.
• Measure support on ticket closure time, and tickets get closed prematurely, then reopened — the customer's problem now takes longer to solve, while the dashboard improves.
• Measure cold outbound on open rates, and you get subject lines engineered to trick, inflated further by mail scanners that register opens no human made.

That last one is why we treat opens as a diagnostic at best. The number we hold outbound campaigns to is positive replies — around 4% is a reasonable expectation, and below 3% means fix the campaign — because a genuine "yes, tell me more" from a real prospect is far harder to game than an open pixel.

What is the mechanism that corrupts a metric?

It runs the same way every time. When you publish a target, people study which behaviours move it. When the metric is a proxy — and almost every metric is a proxy for something you actually want, like revenue or retention — there is a gap between moving the proxy and doing the work. When effort discovers that gap, it flows into it, because gaming the proxy is cheaper than improving the underlying thing. When enough effort has moved, the correlation the metric relied on collapses. Then the dashboard says the business is improving while the business stands still, and management doubles down on the target because the target appears to be working.

Nobody in that chain needs to be dishonest. Ordinary people responding rationally to the scoreboard they were given is sufficient.

How do you measure without wrecking the measure?

Four design rules, all boring, all effective.

1. Pair every volume metric with a quality metric. Emails sent with positive-reply rate. Calls made with meetings booked. Tickets closed with reopen rate. Gaming one side usually shows up on the other.
2. Put targets on outcomes; report proxies as diagnostics. Reward booked meetings that actually happened, revenue won, clients retained. Watch opens, clicks and activity counts to debug — never to pay bonuses.
3. Audit samples, not just totals. Read five random "closed" tickets a month. Listen to two logged calls. The gap between proxy and reality is visible the moment anyone looks at the raw material.
4. Leave headroom. Goodhart bites hardest when a target is only just reachable, because then gaming becomes the rational route to hitting it. Systems need slack — the capacity you keep on purpose — and targets are no exception.

There is a broader principle behind all four: a handful of honest, hard-to-game rules beats an elaborate KPI architecture. I have written separately about why simple rules beat complex plans; metric design is the clearest case of it.

Which numbers should a founder actually watch?

The ones closest to money and hardest to fake: qualified conversations created, proposals out, win rate, revenue per client. A useful decision rule from our own work — when win rate sits above 60%, raise prices by 15% — only works because win rate is an outcome, not a proxy anyone is paid to inflate. The practical structure is a dashboard that separates targets from diagnostics and compiles itself, so nobody is grading their own homework; the MD Dashboard Blueprint sets out how we build that.

Measure less, measure truer, and assume every number you reward will be optimised exactly as written — because it will be.

---

Keep the thinking coming: The Format — a fortnightly letter on business and personal systems. Subscribe by email.