AI Adopted the Way the Garbage Can Model Predicts

I was talking to a strategy director a few weeks ago and he said something that has not stopped bothering me. "We have seventeen AI pilots running and I cannot tell you what problem any of them solves." He was not complaining. He was describing, with the precision of someone who had sat through every meeting, exactly how the decisions had been made. A vendor showed up with a demo. A department head heard about a competitor using the same tool. A board member asked whether the organization had an AI strategy yet. Each of these was a choice opportunity. A meeting. A budget cycle. A quarterly review where a decision was expected. And into each one walked a solution that was fully formed before anyone had asked whether it fit a problem.

Cohen, March, and Olsen (1972) published the garbage can model of organizational choice in Administrative Science Quarterly. They were studying university governance, which they characterized as an organized anarchy defined by three conditions. Problematic preferences means goals are unclear and discovered through action rather than specified in advance. Unclear technology means the organization does not fully understand its own processes or how inputs translate into outputs. Fluid participation means the people involved in decisions change from moment to moment, so the same problem is discussed by a different group every time it comes up. In an organization with all three conditions, rational choice theory is not a description of how decisions happen. It is a story told after the fact.

The garbage can model proposes four independent streams that flow through the organization simultaneously. Problems are unresolved concerns looking for attention. Solutions are ideas, tools, and capabilities looking for problems to attach to. Participants are the people available to be involved at any given moment. Choice opportunities are the occasions when an organization is expected to make a decision. These four streams do not connect through rational deliberation. They collide through timing. A decision happens when a solution, a problem, a set of participants, and a choice opportunity happen to be in the same place at the same time. The collision is temporal, not logical. The choice opportunity is the garbage can itself. Whatever is in the can when it opens determines what comes out.

I think the garbage can model is embarrassingly accurate for describing how most organizations decided to adopt AI between 2023 and 2025. The solution, large language models and generative AI tools, became publicly available in late 2022. By early 2023 it was everywhere. Every conference. Every vendor pitch. Every board agenda. The solution was in the organizational environment before most organizations had named a problem they wanted to solve with it. The typical sequence was not: we have a customer service problem, we evaluated AI solutions, and we chose the best fit. The typical sequence was: ChatGPT exists, competitors are announcing AI initiatives, the board is asking about our AI strategy, so we need one. Form a task force.

The task force is where the garbage can process becomes visible. An AI strategy task force is not a rational decision-making body. It is a choice opportunity. Participants arrive from different departments, each carrying pre-existing solutions that need problems to attach to. IT wants to deploy Copilot because Microsoft is pushing it through the enterprise agreement. Marketing wants a content generation tool because the team saw a demo and wants to be first. Customer service wants a chatbot because the competitor has one. Each participant brings a solution that already exists, funded, vendor-selected, ready to go, and the problem gets written around it. The problems that get discussed, what should AI actually do for this organization, are real but they are not driving the process. The solutions are driving the process. The problems get attached to whichever solution has the most organizational energy behind it when the choice opportunity closes.

Holmstrom, Ketokivi, and Hameri (2009) describe design science as a process of matching existing solutions with new problems, and they ground that claim in Cohen, March, and Olsen (1972). This is the exact mechanism of what I saw happen in organization after organization. The solution was matched to whatever problem happened to be in the room. The match quality depended on timing, not analysis.

This explains why so many organizations ended up with seventeen pilots and zero strategy. The organized anarchy conditions were satisfied on all three dimensions. Problematic preferences: nobody knew what AI should do for the organization, so goals were discovered through action, through the pilots themselves, rather than set in advance. Unclear technology: most organizations did not understand how LLMs actually work, what their failure modes are, or why a model's behavior changes between versions. The technology was opaque to the people deploying it. Fluid participation: different people showed up to different AI task force meetings. The CTO attended the first one. Department heads attended the second. An external consultant ran the third. The fourth was a town hall. With all three conditions present, the process was guaranteed to produce garbage can outcomes. Decisions driven by what happened to be in the can when it opened, not by strategic alignment.

I do not think this is dysfunction in the negative sense. Cohen, March, and Olsen were not diagnosing failure. They were describing how organizations actually function under ambiguity. The garbage can process is an adaptive response to conditions where rational choice is not available because the information and agreement needed for rational choice do not exist. When you do not know what problems matter most, when the technology is too new for anyone to understand its limits, when a different group shows up every time a decision needs to be made, temporal sorting is the only real mechanism. You grab whatever solution is in the room and attach whatever problem fits best. You call it a strategy.

But the model also predicts which organizations would adopt AI coherently. Organizations that do not meet the organized anarchy conditions, those with clear strategic priorities, stable decision-making bodies, and well-understood operational processes, should make AI adoption decisions differently. And they did. The organizations that had a defined strategic planning cycle, a stable IT governance committee with consistent membership, and a shared understanding of their core operational problems did not form seventeen-pilot task forces. They asked whether AI solved a problem they had already named. They either adopted it or did not based on that fit. The process looked like rational choice not because those organizations were smarter but because their structure made rational choice possible. The garbage can model operates only when the conditions for organized anarchy are present. Remove the conditions and you remove the mechanism.

I wrote about this pattern before in my post on why everyone copies everyone else's AI strategy, and the connection to the garbage can model is direct. Mimetic isomorphism, organizations copying each other under uncertainty, is what delivers solutions into the garbage can before problems are identified. When every firm in your industry announces an AI strategy, the solution is already circulating. It arrives at choice opportunities ready to attach to whatever problems are present. The mimetic pressure ensures the can is full of solutions before anyone has checked whether those solutions match the problems. The isomorphism post explained why organizations copy each other. The garbage can model explains what happens after the copy arrives. It enters a decision process where temporal sorting replaces strategic evaluation.

I keep returning to the same conclusion. The garbage can model is not cynical. It is accurate. And fifty-four years after Cohen, March, and Olsen published the mechanism, the last three years have demonstrated that they did not get it wrong. If you form an AI task force in an organization with unclear preferences, unclear technology, and fluid participation, the outcome will be a garbage can process. You can pretend it will be different. But the model has been right for five decades, and it is right about this.