The management of an organization is often a struggle against human nature. Leaders establish metrics to drive performance, yet those metrics frequently produce the opposite of the intended result. This phenomenon is a perverse incentive. It is a structure where the reward mechanism encourages behavior that makes the original problem worse.
In the context of artificial intelligence, organisations are currently repeating historical errors in incentive design. Businesses are measuring "AI adoption" or "prompt volume" without a clear link to business value. These measures are the modern equivalent of a colonial snake bounty. These measures are incentives that exacerbate the problem that companies are trying to solve – keeping costs down while encouraging strategic innovation.
The Cobra Effect: A Lesson from Delhi
The term "Cobra Effect" originates from a story in British India. The government in Delhi wanted to reduce the population of venomous cobras, so they offered a financial bounty for every dead cobra delivered to officials.
Initially, the program is a success. The number of snakes in the city decreases. However, the incentive is too attractive, and resourceful individuals begin to breed cobras to collect the bounty. When the government discovers this fraud, they terminate the program. The breeders, now possessing thousands of worthless snakes, release them into the wild. The result is a cobra population that is significantly higher than before the intervention.
The error is the measurement of a proxy rather than the outcome. The goal is a snake-free city, but the metric is "dead snakes delivered." When the metric becomes the target, the system is gamed. According to economist Horst Siebert (2001), who coined the term, this is the fundamental failure of economic and social policy design.
Modern AI Cobras: The "Tokenmaxxing" Problem
Many enterprises are currently breeding digital cobras. Management is eager to show progress in AI integration. They track the number of employees using Large Language Models (LLMs) or the total count of AI-generated prompts.
The pressure to perform against these metrics leads to "tokenmaxxing." This is a practice where employees run trivial or unnecessary tasks through internal AI agents specifically to inflate usage scores. A recent Financial Times report indicates that even at companies like Amazon, developers are using AI for meaningless tasks to climb internal "Leaderboards.
When an organisation measures prompt count, the result is a massive volume of low-quality text. This is a digital shadow AI risk, and is a waste of compute resources and human time. The output is the AI token version of a "dead cobra" delivered for a bounty, while the actual business problem remains unsolved.
The real objective is improving business outcomes, not simply "using AI". Usage stats are a poor proxy for value, and there is an assumption that "using AI" = "innovation", and this is simply not true.
The Great Emu War: Brute Force vs. Decentralised Reality
The misuse of high-tech tools for complex problems has historical precedent. In 1932, the Australian government declares "war" on emus. Approximately 20,000 birds are destroying wheat crops in Western Australia, and the chosen solution was the deployment of the Royal Australian Artillery, armed with machine guns.
It was assumed that bigger is better, and that strength and machine guns are the answer. However, the emus do not follow traditional military formations, and they split into small, agile, decentralised units. The soldiers fire nearly 10,000 rounds of ammunition but kill fewer than 1,000 birds. The "war" is a humiliating defeat for the military, and the takeaway lesson is that high-tech brute force is ineffective against decentralised, complex problems.
Organisations today are making the same mistake by trying to solve every business challenge with a massive LLM. They are using "machine guns" to hunt "emus." The problem is not a lack of AI power. What is missing? Proper data foundations and organisational data fluency. A centralised, expensive AI tool is frequently the wrong weapon for a nuanced, distributed business process.
Goodhart’s Law and the Ghost of Strategy
The common thread between cobras and emus is Goodhart’s Law. This rule states: "When a measure becomes a target, it ceases to be a good measure."
As discussed in my recent analysis of the $500 Million Ghost, metrics are dangerous without context. If a CEO demands a 50% reduction in customer service response time via AI, the team will implement a bot that closes tickets without solving them. The metric is achieved, but the customer experience is destroyed.
True data fluency is the ability to understand these systemic risks. It is the realisation that data is a reflection of behavior, and humans are experts at optimizing their behavior to capture a reward. If the reward is for "AI activity," you will get activity. If the reward is for "business impact," you might actually get innovation.
Managing means making decisions. Decisions means changing something. Changing means improving something to reach planned goals. How do organisations know if any changes have been achieved if no metrics had been performed at all?
Making decisions under total uncertainty is rather bad practice in management. The PDSA (Planning, Do Study and Act) sometimes will require readapting and replanning. The iterative nature of PDSA enables course corrections, but this feature of the approach is much more effective if there was a clear and reasoned course in the first place. Let's take a look at some practical ways to avoid chasing AI "cobras" and putting forward a better plan that recognises the need to "do" while adapting where and when necessary.
Designing Better Incentives for AI Strategy
To avoid the Cobra Effect in your AI strategy, follow these pragmatic steps:
- Measure Outcomes, Not Proxies. Stop counting prompts. Start measuring the time saved on specific high-value processes or the increase in accurate decision-making.
- Focus on Data Fluency. Ensure that your leadership team understands how to interpret AI outputs. This prevents the "tokenmaxxing" behavior because leaders can identify low-value usage.
- Bridge Technical and Business Perspectives. AI implementation is not an IT project. It is a strategic shift. We help organizations build data foundations that support business goals rather than just inflating tech budgets.
- Empathy-Driven Strategy. Understand the pressure your employees are under. If they are incentivized to use AI, they will use it poorly to protect their jobs. If they are incentivized to solve problems, they will find the most efficient tool: which might not always be an LLM.
The real problem is the disconnect between the boardroom's desire for "innovation" and the ground-level reality of day-to-day operations. AI is a tool, not a trophy. When you stop paying for dead cobras, you stop getting cobra breeders.
Strategic wisdom is the choice to move beyond superficial metrics. It is the commitment to building a resilient, data-fluent organization that knows when to use a "machine gun" and when to simply build a better "fence."
Join the Conversation
Are you tired of digital cobras in your organisation? Book a free slot with me here.
Our new audio show, Jennifer Stirrup: AI Strategy & Wisdom, explores these topics in depth. I discuss the pragmatics of AI, data strategy, and the human factors that drive successful implementation.
