Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Reinforcement learning and organizational management

The pitfalls plaguing our algorithms plague our companies, too.

[Photo: Getty Images]

Profile photo of Neel Somani

Founder · Eclipse Labs

I took Berkeley’s CS 188 course in artificial intelligence years ago, and like most students, I left with only a basic understanding of reinforcement learning. When I began coding and working in the industry, I realized many gaps in my education. I became curious, and that curiosity took me in an unexpected direction.

Exploring the gap between theory and practice, I observed that reinforcement learning’s mechanisms apply to everyday life, including how organizations adapt and learn.

Reinforcement learning is a branch of machine learning, but it’s general. Here, I’m talking specifically about:

  • Reward signals: Knowing what behavior to encourage
  • Feedback: Connecting decisions to their actual consequences
  • Exploration vs exploitation: Staying curious instead of just optimizing what already works

Almost all problems can be framed in this way.

The same forces that make reinforcement learning difficult for machines also make organizational adaptation difficult. For example, in a chronic drift of rewards away from what the system intended to optimize.

Think of this as a thought experiment. I’ll be borrowing a few tools from reinforcement learning to better understand how groups of people learn and adapt.

What gets rewarded is what gets learned

In reinforcement learning, the reward function is everything. An algorithm doesn’t understand goals the way humans do. It just knows that a certain action produced a specified reward, so more of that. If you define the reward badly, the algorithm performs exactly as you told it to—not what you meant.

It’s analogous to how many companies want to reward top performers, but end up encouraging the wrong things. For example, leaders might emphasize innovation, collaboration, and psychological safety, yet reward individual output in terms of quantity rather than quality. Status accrues to whoever ships fastest, not whoever solves the hardest problems.

People aren’t stupid—they optimize for what actually affects their career, not what the leadership deck says matters. The system then learns what it rewards, reinforcing the mistake. Let’s look at it through both lenses:

Mechanical reward signals

You set a reward function, and the algorithm finds the shortest path to maximize it, even if that path misses the point entirely.

 A classic example is rewarding a cleaning robot for picking up dirt from the ground. The robot then increases the amount of dirt on the ground so it can pick it up. This is called “reward hacking.”

Organizational reward signals

When incentives contradict stated goals, people learn to game the system rather than solve problems. For example:

  • Sales teams close deals that churn immediately because revenue counts, but retention doesn’t.
  • Engineering teams ship features that create technical debt because velocity is measured, but maintainability isn’t.
  • Support teams close tickets without resolving underlying problems because resolution time is tracked, but customer satisfaction isn’t.

The incentive landscape is mostly invisible, making change difficult.  

In reinforcement learning, the fix is straightforward: Redesign the reward function to measure what you care about.

For organizations, the equivalent is auditing what gets rewarded. What behavior leads to promotion? Which metrics drive compensation? Whose work gets celebrated in all-hands? If those answers don’t match your stated priorities, your stated priorities just don’t matter.

Learning requires usable feedback

In reinforcement learning, there’s a problem called credit assignment. When an algorithm takes multiple actions and receives a reward signal much later, it must determine which prior action actually mattered. When feedback is delayed or noisy, the algorithm ends up guessing.

Organizations face the same problem. For example, annual performance reviews try to connect January’s work to December’s results. Or strategy decisions get evaluated after three reorganizations and two budget cycles.

Without clear feedback loops, teams optimize the visible and immediate. Using additional dashboards or metrics to fix the issue tends to only add noise. The more complex the system, the harder it is to connect action to outcome.

In RL, the answer is shortening the feedback loop. I won’t tell leaders how to do that in your organizations, because it’s not my expertise. However, you’ll need to consider opportunities to identify high-quality work early within your company.

That said, this risks manually rewarding intermediate results rather than the final, desired result. For example, within a venture capital firm, associate rewards could be based on the number of new founders they meet, since meeting more founders generally means closing more deals. However, what you really want to reward is the venture capital firm’s profitable earnings.

When exploration ceases

Every reinforcement learning algorithm faces the exploration-exploitation tradeoff. Exploitation maximizes reward based on current knowledge, while exploration attempts new approaches that might perform better, despite uncertain outcomes.

Pure exploitation appears productive, so algorithms naturally converge on it. A better strategy might exist, but the algorithm stopped looking. Sound familiar?

Organizations naturally drift toward exploitation, but when the market changes or a new technology emerges, they’re stuck with the playbook they spent years perfecting. Let’s break down the two approaches.

  • Mechanical: An exploration-oriented RL algorithm will sometimes take actions that yield lower immediate rewards. That’s the tradeoff. Most algorithms solve this with epsilon-greedy strategies—exploit most of the time, but reserve a small percentage of actions for random exploration.
  • Organizational: Companies can identify mechanisms for exploration, such as hack weeks, role rotations, and experimental budgets. However, these are always cut first when efficiency is prioritized.

In both machine and human decision-making, exploration often appears wasteful before it yields returns.

The system underneath

Artificial reinforcement learning is just one lens to evaluate organizations. However, this thought experiment taught me that the organizational psychology of incentives and motivation requires structural attention. If AI is any guide, the organizations that design better learning conditions will outpace the ones that don’t.

Neel Somani is the founder of Eclipse Labs.