It's Time for Management to Re-Learn Goodhart's Law

I came up professionally as a Software Developer. Since then, I’ve transitioned into technical education and leadership. I’ve held executive-level positions and founded and run three companies. Today, I’m fortunate to be able to deliver work for my clients as an individual contributor and also advise executive teams. I feel like this keeps me somewhat balanced in my perspectives.

The current AI push is simultaneously fascinating, hopeful, and dystopian. As a society, we don’t currently have the final answers on the pros and cons of these tools and whether they’re a net positive ROI. So, as is natural, we want to measure it. However, I know, from the mistakes I’ve made as a leader, that we have to be very careful about what we measure, because what you measure, you get more of, for good or ill.

The Rush for Data

Across industries, organizations are mandating the use of AI tools and tracking employee engagement with these systems. Sales teams must demonstrate AI-assisted call preparation. Legal departments report percentages of documents reviewed with AI assistance. Software developers log their usage of coding assistants. The intention is good: companies want to ensure they're getting value from expensive AI investments.

But measuring AI usage creates incentives that optimize for tool engagement rather than actual value creation. I believe a lot more companies are going to get “enshittified”, the term the cool kids are using on social media.

Impacts on Software Developers

We learned decades ago that measuring software developer productivity by the number of lines of code produced is a horrible metric. In programming, much like prose, there are many “correct” ways to write something. Some outputs are more readable, others are more performant, and it’s absolutely possible (and sadly common) to produce bloated, inefficient programs. Some developers gamed the system by writing verbose, unnecessarily complex code to hit their targets.

I’m already directly seeing the impact of this and hearing increasing complaints from my technical peers. AI coding assistants are genuinely useful. They can often generate quick, functional snippets. With the decline in usability of Google “Search”, AI tools are pretty great at doing some initial research or quick lookups.

I put “Search” in quotes ironically, because search is becoming increasingly unreliable due to inaccurate AI summaries and SEO slop appearing at the top of the results.

Currently, AI tools struggle with broader system architecture and existing codebase patterns. When developers are pressured to hit AI usage targets, the early results are unsurprising. Here are just a few of the things being discussed in my peer groups:

Code duplication increases because AI tools lack the context of existing functions within the system. This is especially true when code is spread across multiple projects. One lesson we learn quickly when growing from student to professional is that the vast majority of the time and expense associated with software is incurred after initial creation, in maintenance. These tools also produce a significant amount of technical debt when not managed carefully.

Security risks increase dramatically. AI tools are often unaware of new exploits and package versions, so they sometimes recommend outdated, vulnerable approaches. Developers who rely heavily on AI without a deep understanding ("vibe coders") won't catch these issues. In fact, malware authors are creating fake packages that represent common AI hallucinations to trap these careless developers.

Burnout is increasing among the more senior and competent engineers. They’re being forced to review more code than ever because code is easier to generate. However, much of the code generated by their less skilled team members is code that they are either too lazy or too incompetent to review properly. It’s turning the senior job into more of a “playing goalie” position than a “creative leader and mentor” position. They hate this, and I’m hearing more often that they’re considering quitting.

The irony of this isn’t lost on more senior professionals like me. We abandoned lines of code as a metric because it encouraged quantity over quality. Now, with measuring AI tool usage, we’re encouraging the same problematic behaviors.

Other Fields Feel it Too…

The problem is already extended beyond software development. In high-stakes sales situations, buyers need to feel connected and understood by their representatives. When purchasing enterprise software worth hundreds of thousands of dollars or considering major consulting engagements, prospects evaluate the salesperson's expertise and genuine understanding of their challenges.

AI-generated responses and templated solutions risk breaking the human connection. A sales rep optimizing for AI usage metrics might rely on algorithmic insights for discovery questions or proposal writing, but sophisticated buyers can detect this artificiality. The very metric designed to improve sales efficiency can be systematically destroying close rates on the most valuable deals.

In fact, most of my peers with purchasing power have told me in no uncertain terms that if they detect AI usage during the initial outreach and discovery, it kills the potential deal. They’re engaging with providers for solutions to big problems, and if all the vendor has to offer is $20/month GPT plans, why should they pay a premium for it?

Legal services face an even more serious risk. The profession exists to provide expert judgment and advocacy, yet measuring AI adoption encourages shortcuts that undermine these core responsibilities. Recent cases of judges making decisions based on AI hallucinations and lawyers getting sanctioned for citing fabricated precedents are early examples of the consequences of this trend.

We don’t have the tools in place to measure these impacts yet. How many cases have been weakened by surface-level AI research that missed crucial precedents? How many clients have received subpar representation because their attorney optimized for tool usage rather than deep case understanding? The appeal of maintaining billable hours while reducing effort is strong, but this thought process is at odds with providing quality legal representation.

Activity vs Outcomes

When executives are busy, distracted, and don’t understand the reality of the work below them, it can lead to conflating activity with outcomes. Activity is relatively easy to measure, but outcomes are not. AI usage is easily monitored through software logs and self-reporting. Actual improvements in decision quality, client satisfaction, or code maintainability require more sophisticated measurement approaches.

This isn’t a new problem. There are volumes of business case studies that demonstrate how proxy metrics can seem reasonable but miss the point entirely. Customer service departments that optimize for call duration create frustrated customers who can't get their problems solved. Marketing teams that focus on lead quantity rather than quality waste sales resources on unqualified prospects.

AI usage metrics represent the latest iteration of Goodhart’s Law. When tool adoption becomes the primary goal, the incentive it creates is in opposition to quality and outcomes.

It’s disturbing to me that with all the complaints about Windows 11, Microsoft Teams, and the god-awful customer service that users are reporting out of Microsoft, they’re publicly mandating and measuring AI tool use instead of, you know… customers being delighted by the quality and usefulness of their software.

There are Better Approaches

I’m not advocating for abandoning AI tools or avoiding measurement. AI is genuinely useful when applied carefully. However, companies should be focusing on outcome-based metrics that capture actual value creation.

For example, the Duolingo CEO recently boasted that they launched 148 new AI-written courses that would have taken human workers 12 years to develop. As an educator, my response was “so what?” Are the learning outcomes better? Are your customers happier? I can assure you, as someone who crafts curriculum for a living, that the quality of this output is likely subpar. And, if you poke around, you’ll see a lot of users are complaining about a substantial drop in quality. If I were on the board at Duolingo, I would see this as an existential threat to the business and react accordingly.

It’s an interesting idea I’ve been working through. If AI can generate quality courseware with minimal human intervention, why does anyone with a $20 subscription to a frontier model need Duolingo at all? What’s their value-add? It certainly isn’t accuracy and quality curation!

The legal folks have it easier. They can measure case outcomes, client feedback, and the thoroughness of research rather than the percentage of documents processed with AI assistance. These metrics are more difficult to game and more directly related to the value these professionals provide.

Sales and marketing? Well, I’m not the right person to weigh in on effective measurement. Hats off to sales and marketing experts, because what you do is basically voodoo magic to me! Please leave some comments if you'd like to suggest thoughts for me to consider.

Think Carefully about What You Measure

Goodhart's Law reminds us that measurement shapes behavior. When we measure the wrong things, we get more of the wrong things. The software industry learned this lesson with lines of code, but apparently, we need to learn it again with AI adoption.

The goal should be better outcomes, not more tool usage. Executives need to resist the temptation to measure what's easy rather than what matters. To do otherwise is to optimize for activity while undermining outcomes.