We are using AI to make yesterday cheaper

Ask a room of executives what AI is for, and the answers tend to sound familiar: faster reports, fewer support tickets, less time spent on routine work. Those are real gains. If a task takes an hour and AI helps finish it in ten minutes, take the win.

But there is a floor. You can only remove so much time and cost from a task. Once you have made yesterday’s work cheaper, the larger question is what you can do that you could not do before.

That question is already changing science. AlphaFold’s public database now gives researchers access to more than 260 million predicted protein structures.^[1] In 2023, Google DeepMind’s GNoME project reported 2.2 million candidate crystal structures, including 381,000 predicted to be stable against the expanded set of competing structures.^[2] In 2020, an MIT-led team used a model to identify halicin as an antibiotic candidate and demonstrated antibacterial activity in laboratory tests and mice.^[3] Predictions still need experiments; promising preclinical results do not establish safety or effectiveness in humans. The progress lies in how much more scientists can investigate.

This is a different use of AI from asking it to summarize the meeting about an existing plan. It expands the range of things we can examine, test and build.

That was the useful tension at Moonshots LIVE in Los Angeles on September 25, 2026.^[4] XPRIZE put two kinds of creation on the same stage: Future Vision XPRIZE, a competition for hopeful visions of the future, and Build with Gemini XPRIZE, a challenge to launch an AI-powered business in 90 days.^[5][6] One asks people to imagine a future worth making. The other asks them to make something that works. We need both.

Most organizations can pursue both, too. Use AI to give people time back, then decide what that time is for. Let it handle part of the repetitive analysis, then investigate a customer problem the team has never had capacity to study. Speed up a research workflow, then run more experiments. Reduce the cost of a prototype, then test ideas that once seemed too uncertain to justify building.

Doing both creates a management problem. Experiments need room to change direction. Reliable delivery needs shared methods, clear ownership and work people can repeat. A company that standardizes every promising idea too early may stop learning. One that treats every project as a fresh experiment may struggle to deliver what customers already depend on.

We have been managing that tension at ThinkForge Labs ourselves. It raises questions that a decision to “invest in innovation” does not answer: Which work should become a repeatable system? Who has permission to change it? How do we protect time for uncertain ideas while keeping promises already made?

That deserves its own article. We’ll look at how companies can configure their operating models to run experimentation and standardization in parallel, and why the answer may look different for a large organization with established systems and a small one still discovering what works.

The risk today is that we use a remarkable new capability to preserve every old assumption. We keep the same products, the same questions and the same ambitions, only with a smaller budget attached.

Cost savings can fund the next move. They cannot be the whole destination. The organizations that matter most in the coming years may be the ones that ask, alongside “How do we do this for less?”, “What is now possible that we have not yet tried?”