If the Decision Has Already Been Made, What Are We Discussing?

A hot, airless jury room, a table and twelve men who want to reach a verdict… 12 Angry Men, from 1957. I have seen very few films that show more plainly how we make decisions. When the film begins, the verdict has all but been reached. Eleven jurors are sure the young defendant is guilty. The evidence seems clear. One wants to make it to the evening’s baseball game; another thinks the discussion is dragging on for no reason. Only Juror 8 does not vote guilty. He is not sure the defendant is innocent. He simply wants to talk about whether there is enough reason before sending a man to his death. No new evidence enters the room. Juror 8 only turns what the others see as conclusions back into questions. Everyone looks at the same evidence, but no one sees the same thing. Each looks with a different past, a different anger and a different hurry. Some things they took for knowledge turn out to be assumptions; some thoughts they took for impartial judgment turn out to be personal stories.

What makes the film immortal is not only that we can draw different conclusions from the same information, but that it shows the invisible assumptions our decisions stand on. Good judgment, too, usually begins not with finding new information, but with making visible the assumptions that carry the answer we already have.

The part of the film that most resembles business is not that people can make the wrong decision, but that they behave as if they are still deliberating when the decision has all but been made. We see this often inside organisations. The meeting is called a “decision meeting”. The manager says “I’m open to every view”, while everyone in the room knows which view is safer for their career. There is data, expert opinion and long discussion on the table, yet none of it has any chance of changing the verdict. In moments like these, the decision table stops being a place where we test options and becomes a place where we prepare the defence of a decision already made. In judgment, the verdict can change. If the verdict is fixed from the start, what we are doing is not judgment but justification.

Herbert Simon argued that people and organisations cannot see every alternative or calculate every consequence. That is why, most of the time, we do not find the best option; we stop searching once we find one that looks good enough. Simon received the Nobel Prize in Economics in 1978 for his work on decision-making in organisations. Today, with AI, we process more data than a person could weigh at once, generate hundreds of alternatives and calculate the outcomes of different possibilities. But does having more options really improve our decisions?

In one of the planning problems we worked on at Arya, the first calculations favoured the plan that met the highest possible demand. We were shipping more product, using more capacity and, on paper, protecting more revenue. The system was working correctly with respect to the objective we had given it. Then we saw that a small-looking assumption changed the entire ranking. We had assumed every delay cost the company the same. On the operations screen, both orders appeared as a four-day delay. The sales team, however, knew that one customer could wait, while a delay on the other could stop a production line. Let’s be honest. How many companies record that kind of information in a database on a regular basis? Much of the critical information that shapes a decision lives in the experience and minds of people who have worked with customers for years. We did not discover a new truth. We made visible a truth the company already knew by bringing it into the model. The first model had not given us a wrong answer. We had given the system an incomplete judgment.

There is an expectation we often hear from new customers: “Let Arya look at our data, learn and tell us the answer.” I call this a kind of AI-expectation poisoning. We assume that all the information a decision needs is sitting ready in databases, and that a powerful enough system will process it and produce the right answer on its own. But real decision knowledge does not live only in systems. It is spread across the salesperson who knows the customer, the operator who knows the real limits of the line, the planner, the manager who lived through a similar crisis in the past, and the networks of relationships that form outside official processes. What an organisation knows and what is recorded in its databases are not the same thing. And by AI we do not mean only language models. Forecasting models show what might happen, optimisation shows which options are feasible, and simulation shows what the path we choose would turn into under different conditions. Language models help us talk through the assumptions behind these outputs and compare different explanations. But none of these models produces good judgment on its own. Stronger human–AI judgment emerges when we bring together the knowledge in the data with the knowledge that lives in people’s experience, and when we can test the same verdict with different methods.

So the point is not just to tell a model “Argue with me” or “Show me my blind spots”. An instruction like that can be useful, but it does not mean the antithesis it produces is right. More importantly, no prompt can change the cost of disagreeing with the boss or the incentive to hide failure. In 12 Angry Men, the presence of one dissenting juror was not enough. There was also a setting in which the others had to listen to him. An objection could stop the vote, the evidence could be re-examined and people could change their minds. Telling AI to disagree is a prompt. Building a structure in which objections raised by people, data and different AI tools can be tested against evidence, and can change the verdict, is decision architecture.

We can test our process of judgment with three questions. What assumption are we deciding on? What new information would change our mind? Who at the table, or which system, can actually stop the decision? If these questions have no answers, what we are seeing is most likely not judgment but the corporate staging of a verdict already reached. The first condition of good judgment is not knowing everything; it is keeping our own verdict open to change and building the setting that can test it. We know such a setting is not easy to create. In some companies it takes not just new tools but changes in habits, power relationships and ways of talking that go back years. We can certainly support that transformation with AI. But today we mostly judge AI systems by how quickly and flawlessly they complete the task we give them.

When we do not build a setting for judgment, AI adapts to the existing order too. Instead of testing our verdict, it justifies it; instead of opening up our blind spots, it makes them more convincing. So can we call a system successful if it flawlessly delivers a goal we defined wrongly?