Introduction: The reality of organizations where "goals become counterproductive," and solutions from AI research.
Many companies frequently use the terms "KPI achievement" and "performance indicators." While these systems are supposed to promote organizational growth, many people are actually facing unexpected situations.
This article explores why "good goals" can lead to "bad behavior," drawing lessons from the large-scale fraud scandal experienced by the major US bank Wells Fargo in 2016. Furthermore, it introduces how to implement four strategies developed through AI machine learning research to revolutionize organizational performance measurement.
Gordon's Law, where A and B are reversed, and the trap of overfitting
One concept that you should first understand is "Gordon's Law." This law describes the phenomenon that "when an indicator becomes a goal, the indicator itself becomes ineffective." For example, if you set "sales revenue" as your goal, it can result in actions that prioritize quantity over quality customers.
In the field of research, this is called "overfitting." It is a phenomenon where a machine learning model performs very well on training data, but fails on new data or in real-world situations.
Examples in organizations:
- Sales team→ Achieving goals by forcibly selling unnecessary products
- school→The focus shifts from genuine academic improvement to test preparation.
- Customer Service→Just behave well when you're being evaluated.
Short-term improvements in indicators can actually reveal or mask more serious problems. This is where important solutions emerge.
Four strategies to prevent overfitting: Practical know-how from AI research
Researchers have long studied "overfitting," and there are four strategies to combat it that can also be applied to organizational design. We will explain each of them in detail.
1. Early termination: Regularly re-evaluate metrics.
Research has shown that a model's performance on new data degrades after a certain point. The solution is "early termination," which involves stopping training once performance on validation data peaks.
Practical implications:
- Regularly re-evaluate metrics to align with true goals and prevent overfitting.
- Amazon's example: In 2004, they switched their primary metric from profit margin to free cash flow per share. This decision was made to avoid over-fitting to a specific metric and to align with changing business needs.
- Setting minimum standards can also be effective. For example, a customer service goal like "resolve inquiry 85% within 24 hours" allows you to focus on other important tasks without being forced to improve beyond that.
2. Noise injection: Incorporating random variations.
Systems benefit from intentional randomness. If a model relies too heavily on a particular pattern, it simply memorizes quirks.
Practical implications:
- Conducting random audits→Checks by mystery shoppers at unpredictable times can curb the behavior of sales staff only providing good service when they are being evaluated.
- Team composition rotation→Sales teams presenting to the same manager may end up optimizing for the manager's preferences rather than customer value. Noise injection can be implemented through changes in the work environment.
- Replacing decision-makers→Reduces the risk of always applying the same decision-making criteria.
The U.S. Transportation Security Administration (TSA) rotates different functions, such as baggage screening, metal detectors, and document checks, among its personnel to prevent security gaps created by predictable work assignments. By making it impossible to predict who will perform the inspection and when, it prevents malicious actors from abusing the system.
3. Capacity Alignment: Matching the complexity of metrics to capabilities.
Researchers know how to adjust the complexity of a model based on the amount of data and validation capabilities. A team with millions of training data points and powerful validation methods can safely run a complex model with many parameters, while a team with limited data needs to adopt a simpler model to avoid overfitting.
Practical implications:
- Startup companies→By focusing on simple metrics suitable for the growth stage, such as customer acquisition cost, monthly recurring revenue, and customer retention rate, we are effectively integrating these metrics.
- large company→This involves conducting more advanced analysis to track predictive customer health scores, improvements in experimental outcomes, model performance and fairness, and risk-adjusted ROI for the entire portfolio. These are often integrated into a balanced scorecard that links financial, customer, and investment metrics.
- Important PrinciplesCompanies with advanced analytical infrastructure and skilled data scientists can implement complex, multidimensional metrics while maintaining consistency. On the other hand, organizations with limited oversight should set simpler, more robust metrics.
4. Regularization: Incentives for simplicity
Regularization in research is a technique that penalizes excessive complexity to prevent models from overfitting to specific training data. The same principle can be applied to organizations.
Practical implications:
- Adding constraints→This prevents complex gamified behaviors resulting from excessive optimization and promotes behaviors that lead to essential results.
- Balance of incentive structure→We will build a system that rewards genuine improvement while curbing extreme behavior.
- Aligning optimization with the leader's true goalsThese constraints make achieving superior performance more difficult and costly than gamifying the system, and regularization guides AI models towards simpler patterns, while organizational constraints also discourage complex gamification strategies and promote actions that lead to essential results.
Key points and precautions: Conditions for success
These approaches need to work together, not individually. Here are some of the most important points:
Introduction of measurement intelligence:
- Measurement should not be viewed as a "set it and forget it" system, but rather as a process for continuous improvement.
- It is essential that the people being measured are involved in the design of the measurement system. This leads to the creation of more appropriate metrics and gains support for new initiatives.
Cultural/human issues:
- Leaders need patience and the ability to resist the pressure to deliver short-term results. This is similar to the decision to delay deployment until the AI team has fully trained and tested its models.
- Cultural resistance to new metrics is similar to distrust of decision-making based on black-box algorithms.
- The key to success lies in competition. Just as new algorithms only become effective when they compete with existing software and operations, metrics also need to be integrated across organizational silos.
A balance between structure and flexibility:
- It is necessary to apply metrics that set sufficient friction to deter gamification, without stifling spontaneity or leading to data overload.
- Not everything important can be quantified. Because organizations are comprised of individuals with complex motivations, management decisions remain crucial.
How to utilize the insights gained from training A:
- This will be useful not only for evaluating human performance, but also for setting goals and evaluating AI workers.
- The importance of robust performance metrics will only increase.
If measurement strategies are not improved, organizations risk repeating the mistake of optimizing only a narrow range of features.
Summary: Important insights that can be applied to organizational management starting tomorrow.
We use numerous metrics every day to "measure results." However, the measurement itself often leads us astray from our true goals.
As you should have understood through this article, the insights gained from AI machine learning research are not merely a matter of technology, but are also deeply related to human organizational management. The four strategies for preventing overfitting—early termination, noise injection, capacity alignment, and regularization—are all important guidelines that directly suggest insights into organizational design.
The important thing is not to "set metrics and be done with it," but to think of measurement as a process for continuous improvement. And above all, it is necessary to create an environment where people involved in the measurement system itself can speak up and give their opinions.
It's not easy to resist the pressure to achieve short-term results. However, that effort is a step towards the future that the organization truly aims for. Let's strive for a balanced approach, while also valuing the things that cannot be quantified.
Small changes that can be started tomorrow morning can yield significant results in the long run. Leverage these insights within your organization to develop more effective goal-setting and measurement systems.
