Building a working prototype used to take a team and a quarter. Now it takes one person and a weekend. Most companies have noticed the first half of that change. Very few have noticed what it does to their portfolio.
When starting is cheap, everything gets started. And when everything gets started, the scarcest skill in the company is no longer building. It is stopping.
The cost of starting collapsed, the cost of keeping did not
Every product, feature or internal tool has two costs. The cost to build it, and the cost to keep it alive: maintenance, support, documentation, security reviews, integration with everything else, and the attention of the people responsible for it.
AI cut the first cost dramatically. It barely touched the second. A prototype built in a weekend still needs someone to own it on Monday. A customer-facing feature still needs support, still breaks when a model is updated, and still carries risk when it gives a wrong answer.
So the arithmetic changed. A team can now launch ten experiments in the time it used to take to launch one. If it keeps all ten, it has multiplied its maintenance burden by ten, while each experiment gets a tenth of the attention.
That is how companies end up with dozens of half-alive AI pilots and no clear winner. I wrote about the result in most AI pilots die in the middle, and the cause is usually not failure. It is the absence of a decision.
Why nobody kills anything
Killing a project is emotionally expensive. Someone built it. Someone championed it in a meeting.
Someone's quarterly goals are attached to it. And there is almost always a metric that looks promising if you squint.
Organizations solve this the wrong way. They wait for the project to fail clearly, which rarely happens. Most bad ideas do not crash. They drift.
Usage is flat but not zero, a few customers like it, the team believes one more iteration will turn it around. Six months later it is still running, still consuming people, still not working.
From my own experience building products at GEOflux and Madison AI, the hardest calls were never the obvious failures. They were the things that worked a little. A little traction is the most dangerous signal in product development, because it is just enough to justify continuing and never enough to justify the investment.
How labs actually work
Good research labs handle this with a simple discipline. Every experiment has a hypothesis, a time box and a kill criterion, agreed before the work starts.
If the criterion is not met by the deadline, the experiment ends. No debate about whether the metric was fair. The debate happened at the start, when nobody was attached.
The second discipline is separating exploration from scaling. Exploration is cheap, fast and expected to fail most of the time. Scaling is expensive, slow and expected to succeed. The mistake most companies make is running exploration with scaling rules, demanding business cases for experiments, or running scaling with exploration rules, letting a promising prototype grow into a product without anyone deciding it should.
The third discipline is celebrating clean kills. A team that stopped three experiments on schedule and moved resources to the one that worked has done better work than a team that kept all four running. If your culture only rewards launches, people will hide failures instead of ending them.
Setting up a lab inside a normal company
You do not need a research department to do this. You need a few rules, written down.
Give every experiment a one-page brief: what we believe, how we will know, by when. Make the success metric something that matters commercially, not activity. "Customers use it weekly and would pay for it" is a criterion. "People clicked on it" is not.
Time box everything, and keep the boxes short. Four to six weeks is enough to learn whether an AI-assisted product idea has a pulse. If it needs a quarter to show any signal, the idea is probably too big for an experiment and should be broken down.
Hold a regular review where the only possible outcomes are kill, continue with a new deadline, or graduate to scaling. Graduation should be rare and should come with real resources, a real owner and real expectations. Everything else either earns another short box or ends.
The rare skill is judgment, not building
There is a quieter consequence here for people's careers. If anyone can build, the valuable people are the ones who can tell which things deserve to exist.
Product judgment, market sense, a feel for what customers will actually pay for. Those were always valuable. Now they are the bottleneck.
The same applies to leaders. A CEO who can approve ten experiments is common. A CEO who can end eight of them without drama, and move the people to the two that matter, is rare. In a world of cheap prototypes, that is the job.
Anyone can start something now. The companies that win will be the ones that know when to stop.