My favorite AI paper of 2026 is about a map.

Hyunjin Kim at INSEAD, with Dahyeon Kim and Rembrand Koning at Harvard Business School, took 515 high-growth startups going through a three-month accelerator and gave every one of them the same package of API credits, AI tools and technical training. That is the package most companies buy when they say they are “adopting AI.”

Then they flipped a coin.

One group got a curriculum, the other got a map

The control half followed a standard entrepreneurship curriculum, customer profiles and market validation. The treated half studied cases of AI-native companies that had reorganized production around AI. Shorter product cycles. Prototypes built in parallel. Receivables handled by automation. Service businesses that started to run like software.

Nobody wrote code for them. Nobody gave them an extra tool. They were shown other companies’ maps.

The results, from the paper Mapping AI into Production: A Field Experiment on Firm Performance and the authors’ HKU abstract, compared with the control group, were these.

• 44% more AI use cases (2.7 more per firm), mostly in product development and strategy
• 12% more tasks completed
• 18% more likely to land paying customers
• 1.9x the revenue
• 39.5% less outside capital requested, roughly $220,000
• no change in demand for technical staff

Ethan Mollick called it a “big deal paper” when it came out in April. Chamath Palihapitiya brought the 1.9x back on September 25 in his deep dive on why AI is booming but productivity is not. Both are right to flag it. It is a randomized experiment, which almost nothing else in the AI ROI debate is.

The part of the result that should keep you honest

Almost all of the revenue gain came from the top tenth of firms. For the median startup, revenue barely moved in either group.

I like that the paper says so. A good map raises the ceiling. It does not guarantee anything to the firm that follows it.

Why a 200-person company should care about a study of startups

The control startups had AI and knew how to use it. They put it where everyone puts it first, drafting emails, summarizing documents, writing some code. Each person looked for the spot where AI saved them time. One good write-up of the paper puts it in six words I keep repeating: AI search is local, value is systemic.

A mid-sized company has the same problem with more moving parts. More functions, more handoffs between departments, more habits nobody questions. The number of places you could reorganize is larger, and they are harder to see from inside.

Yesterday I wrote about MIT’s finding that half-done AI is associated with lower margins in the S&P 500. That was a correlation. This is a controlled experiment showing the other side: teach a company where to redesign and it earns more with the same tools. Different methods, same direction.

What we do with it at IQ Source

Stage 1 of AI Maestro looks a lot like the treatment arm. We map how the company actually works, function by function. We look for use cases outside the obvious function, because that is where the experiment found its extra 44%. And we compare against how other companies restructured the same work.

Two differences. The accelerator handed out generic cases, and we draw the map on your operation. And every candidate ends in a Go or No-Go call, because the top-decile result carries an uncomfortable lesson too. Some maps lead nowhere, and the right answer is to not build.

If you already bought the licenses and trained the team, you did what the control group did. The next step is the one the treated group took. Send me a note and we can start looking for your 44%.

Map my operation