--- title: "Goodhart’s Law Is Coming for AI" date: "2026-08-14" summary: "AI success isn’t measured by how much you use it, but by the value it creates. Good metrics drive better decisions; bad metrics drive better scores." canonical: "https://thepragmaticcio.net/articles/goodharts-law-is-coming-for-ai/" tags: "CIO, Artificial Intelligence, Leadership, Digital Transformation, Enterprise IT, Governance, Metrics, AI" --- # Goodhart’s Law Is Coming for AI AI success isn’t measured by how much you use it, but by the value it creates. Good metrics drive better decisions; bad metrics drive better scores. ![The wrong KPI can quietly undermine even the best AI strategy. CIOs should measure business outcomes—not just AI usage.](ai-goodhart.png "The wrong KPI can quietly undermine even the best AI strategy. CIOs should measure business outcomes—not just AI usage.") **Every metric changes behaviour.** The moment you start rewarding a number, people stop optimising for the outcome and start optimising for the number. AI is simply the latest example. Across enterprises, organisations are introducing AI adoption dashboards, token consumption leaderboards and usage targets in an effort to accelerate adoption. The intention is understandable. If people aren't using the tools, they'll never deliver value. But there's a danger lurking behind these metrics. The moment token consumption becomes the KPI, employees stop optimising for business outcomes and start optimising for token consumption. Economists have known this principle for decades. It's called **Goodhart's Law**: > *"When a measure becomes a target, it ceases to be a good measure."* CIOs have seen this story before. AI is simply the latest chapter. ## We've Been Here Before Long before generative AI arrived, organisations rewarded people for hitting numbers. Developers were measured by lines of code. Service desks by tickets closed. Infrastructure teams by server utilisation. Sales teams by calls made. Procurement teams by negotiated savings. Every one of those metrics started with good intentions. Every one eventually changed behaviour. Developers wrote more code instead of better code. Support teams closed tickets faster instead of solving problems properly. Infrastructure teams delayed upgrades to maximise utilisation. Sales teams optimised activity rather than relationships. The metric survived. The objective quietly disappeared. Token consumption is beginning to follow exactly the same path. ## The Rise of Tokenmaxxing Some organisations have started tracking token usage as a measure of AI adoption. Others have reportedly introduced internal leaderboards or competitions encouraging employees to use AI more frequently. The logic is understandable. If employees never use the tools, they'll never change how they work. But usage is only the beginning. When employees discover they're being measured on token consumption, behaviour inevitably changes. Instead of asking: *"What's the best way to solve this problem?"* they begin asking: *"How do I improve my score?"* That's the difference between measuring adoption and incentivising behaviour. They're not the same thing. ## Boards Don't Invest in Tokens Here's the mistake. **Boards don't invest in tokens.** **They invest in outcomes.** Nobody celebrates consuming 40% more AI tokens. They celebrate: - Faster product delivery. - Better customer experiences. - Lower operating costs. - Better software quality. - Better decisions. - Higher revenue. Token consumption may contribute to those outcomes. It is not the outcome. Confusing the two is no different from rewarding software developers solely for writing more lines of code. You'll almost certainly get more code. Whether you'll get better software is another matter entirely. ## The Pragmatic CIO AI Metrics Framework Rather than focusing on a single number, I believe CIOs should evaluate AI across four dimensions. ### 1. Adoption Are people actually using the tools? Examples include: - Active users - Percentage adoption - Departments using AI - Frequency of use Without adoption, nothing else matters. But adoption alone is not success. ### 2. Efficiency Is work genuinely getting faster? Examples include: - Reduced cycle time - Time saved - Manual effort eliminated - Faster decision-making This is where AI starts moving from novelty to productivity. ### 3. Quality Did AI improve the outcome? Examples include: - Fewer software defects - Better documentation - Improved customer satisfaction - Higher accuracy - Less rework Doing something faster only creates value if the result is at least as good. Ideally, it's better. ### 4. Economics Finally, ask the question every executive eventually asks. Did we create value? Examples include: - ROI - Cost avoided - Cost per task - Model optimisation - Token consumption Notice where token consumption appears. Fourth. Not first. Because cost without context tells you remarkably little. ## Measuring the Easy Things One reason token usage has become popular is because it's incredibly easy to measure. Every AI platform records it. Every dashboard displays it. Every executive understands bigger numbers. Unfortunately, the easiest metrics are rarely the most meaningful. A development team could double its token consumption while delivering exactly the same amount of software. A marketing department could generate thousands more prompts without improving campaign performance. An analyst could ask twenty variations of the same question. Usage increases. Value doesn't. This isn't an AI problem. It's a management problem. ## The CIO's Role The answer isn't to stop measuring AI usage. Adoption still matters. Understanding how employees interact with AI matters. Monitoring spend matters. The mistake is allowing one easily measured metric to become the definition of success. CIOs should design measurement systems that reinforce the behaviours they actually want. Reward: - Better decisions. - Faster delivery. - Higher quality. - Lower waste. - Greater business value. Not simply higher token counts. Technology changes. Human behaviour doesn't. People optimise for whatever leaders choose to measure. AI is no different. ## The Pragmatic View Every new technology arrives with the temptation to measure what's easiest rather than what's most important. Today it's AI tokens. Tomorrow it will be something else. The lesson remains the same. Metrics should illuminate performance. Not distort it. The organisations that extract the greatest value from AI won't necessarily be those with the highest usage figures. They'll be the ones that align incentives with business outcomes. Because Goodhart's Law doesn't just apply to AI. It applies to every KPI we've ever created. The only question is whether we'll recognise it before we optimise ourselves in the wrong direction. --- **If you could keep only one AI metric on your executive dashboard, which one would genuinely tell you that AI is creating business value?**