From ‘I Think’ to ‘I Tried’: Building an Experiment-Centric Team Culture with AI
AI has made trying things cheap, so teams can argue less from opinion and decide more from the facts small experiments produce.
Meetings about a new idea or an innovative project easily sink into endless debate. Everyone speaks with conviction and quotes the authorities (or just an article they saw yesterday), but the discussion seems to go in circles and often ends inconclusively, or with the decision made by the loudest voice, the highest-ranking person, or the most stubborn one. This is especially common in innovation projects.
The Mire of “I Think”
The root of it is that innovation projects deal with the unknown. If you’re building a standard bridge or optimizing a mature system, experience is king: you can find people who have done similar things, their experience counts as reliable “facts,” and reaching consensus on those facts is relatively easy. Innovation is different. When you’re exploring a brand-new AI application or an unprecedented business model, there’s no map, and past experience may be useless or even harmful. The basis for discussion then shifts from experience to opinion: “I think users will like this feature,” “I believe this technical direction has more potential,” “I feel the market isn’t ready yet.” These are opinions, not facts. Opinions are extremely subjective, shaped by personal background, information cocoons, even one’s mood that day. They’re also hard to falsify: you can’t logically “prove” an opinion right or wrong, because it’s usually about predictions of the future and personal beliefs.
So discussions based on opinions easily get stuck. Sometimes they turn into endless debates in which both sides try to win with logic, ignoring that their underlying assumptions may be completely different. Sometimes, when the opinions really can’t be reconciled, the discussion slides into emotional confrontation and escalates into personal attacks or battles of will, which is devastating for team morale. Another common case is that the decision becomes a game of power or trust: the opinion of whoever is more trusted or more powerful wins, and it isn’t necessarily the best idea. Whichever way it goes, it’s extremely inefficient and badly damages the team’s capacity to innovate and its psychological safety, and a great deal of precious time gets spent wallowing in the mire of “I think.”
Don’t Guess, Run It
Fortunately, AI capabilities have exploded in the past few years. Generative AI and AI-assisted development tools in particular have made the cost and barrier of “trying it out” extremely low, which gives us a weapon for getting out of this mire. Validating an idea used to take weeks or even months of engineering effort just to produce a rudimentary prototype. Now, with AI Copilots, low-code platforms, and AI prototyping tools, a product manager, a designer, or even someone in operations can cobble together a working demo or complete a preliminary experiment in hours or days.
For example, if you want to know how well an AI model handles a certain kind of user request, don’t guess; use an AI tool to write a simple script and run it. If you want to know how users react to a new interaction method, use AI to quickly generate an interactive prototype and have a few people try it. What these quick experiments produce are the facts we urgently need. A “fact” here means an observable, reproducible experimental result under specific conditions, not some absolute truth. Suppose you run an experiment with model X and a particular System Prompt and find that “after 5 rounds of interaction, the probability that it accurately understands this kind of complex instruction is 70%.” That is a fact. It’s objective, it doesn’t involve “I think,” and the discussion can start from there.
Experiments at the Center
With this capability, you can build a team culture centered on experiments. The principle is simple: encourage (or even require) people to turn ideas into small, low-cost experiments as quickly as possible, and let the facts those experiments produce drive discussions and decisions. In practice it goes roughly like this. First, based on your understanding of the problem, state a clear hypothesis. Then design a small experiment: what you want to verify, what prerequisites you need (tools, data, models, etc.), and how you’ll measure the results, keeping it small and fast enough. Next, use AI tools to run it quickly, building a POC or a demo, or running a simulation. When it’s done, record the process, prerequisites, and results clearly and share them. Transparency is key.
The discussion that follows is built on these shared facts. First find the consensus (“We all saw that under condition A, the result is B”), which consolidates what everyone understands. Then identify the disagreements and unknowns (“This experiment didn’t cover situation D,” “Why is the result different under condition C?”), which naturally become the directions for the next experiments. That’s how you keep accumulating facts and get closer, step by step, to the truth of the problem. To cultivate this culture, you need to encourage curiosity and initiative, accept “failed” experiments (they also provide valuable facts and help rule out wrong paths), emphasize transparent sharing, and keep discussions focused on the experimental process and data, without targeting individuals.
An experimental culture doesn’t rule out predictions about the future, or opinions; innovation itself involves imagining the future. The difference is that a prediction has to be explicitly linked to facts already obtained through experiments, rather than being a castle in the air. Discussions start to sound like this: “Based on the experimental results A and B that we’ve seen, I predict that if we try direction C, we might get result D.” That way, the basis of an opinion is transparent and traceable.
Going further, a prediction can itself be treated as a hypothesis to be tested. When someone makes a prediction, the team shouldn’t rush to refute it, or to accept it. First ask two questions: “What facts and assumptions is this prediction based on?” and “What small experiment could we design to test it quickly (or the key assumptions behind it)?” Once those are answered, the discussion about the prediction becomes the starting point for the next round of experiments. It helps us find the most important uncertainties and explore them at the lowest cost, and it saves us from arguing endlessly about a future no one can verify.
The shift from “I think” to “I tried” brings real benefits. When discussion is built on facts, it stops being a contest of eloquence and volume. Decisions get faster and smarter, far less time is wasted on unproductive arguments, and trial and error, learning, and iteration all speed up, which is exactly the pace innovation should have. Consensus built on facts everyone has seen is far more stable and meaningful than a “pseudo-consensus” reached through compromise or power, so collaboration and psychological safety improve a lot, and the negative emotions and personal attacks that come from clashing opinions lose their breeding ground.
Also, with the barrier to “trying it out” lowered, members who usually “haven’t done it” can join in more easily and contribute their observations and thoughts, while those who “have done the experiment” are less likely to develop unnecessary superiority, because the facts are shared. Everyone who runs experiments firsthand gains direct experience, and the whole team understands the problem more deeply. This way of working also fits the principles of efficient communication: establish consensus first, then focus on the differences, and meetings and discussions become more productive.
Of course, the change won’t happen overnight. Putting it into practice will run into some practical difficulties, but they can all be overcome. Team members need to know how to use the relevant AI tools, which may take some training or guidance. The team needs to learn to design meaningful, lightweight experiments, and avoid experimenting for experimentation’s sake. And you have to watch out for over-interpreting or misinterpreting experimental data. The biggest challenge probably comes from cultural inertia: the habit of “meetings are for debating” took a long time to form, and breaking it takes time, as well as leaders who keep pushing and setting an example.
Besides raising productivity, AI is also changing how teams work together. In the past, disagreements in meetings were often settled by whoever was loudest, most senior, or most stubborn; now many of them can be settled by spending a little time trying things out with AI and then discussing the results. In innovation work, which is uncertain by nature, I think this is one of the most practical changes AI brings to a team.