The short answer
A 2026 randomized field experiment handed 515 high-growth startups the same AI tools, credits and training. The only extra was information: half of them saw how peers had reorganized production around AI. That half found 44% more use cases, completed 12% more tasks and finished the quarter at 1.9x the revenue of the control group. The variable was never the model. It was the map.
Give a few hundred companies an identical AI stack, then tell half of them one extra thing: here is how firms like yours moved real work around AI. Kim, Kim and Koning ran that experiment in a global startup accelerator for three months (INSEAD Working Paper 2026/20/STR; SSRN 6513481). Revenue in the treated group climbed to roughly twice the control group's. Demand for external capital fell 39.5%.
TL;DR
- One randomized field experiment, three months, every firm on the same AI tools, credits and support.
- Only the treated firms saw how peers had reorganized production around AI. They found 44% more use cases, most of them in product development and strategy.
- The gains landed almost entirely at the 90th percentile and above. The median startup barely moved in either group.
What did the experiment actually test?
The study randomized firms inside a startup accelerator. Both arms received the same AI tooling, credits and support. The treatment was a single channel of information: founders saw how other companies had reorganized production around AI.
That nudge changed where founders looked. Treated firms searched across a broader set of company functions, while control firms stayed with the obvious ones. The trial is preregistered as AEARCTR-0016746, so the outcome measures were fixed before anyone saw results.
The results, without decoration
| What was measured | Result | Where it showed up |
|---|---|---|
| AI use cases discovered | +44% versus control | Product development and strategy |
| Tasks completed | +12% | Same headcount, more shipped work |
| Likelihood of acquiring paying customers | +18% | Commercial proof, not dashboards |
| Revenue | 1.9x the control group | Across the experiment window |
| Demand for external capital | −39.5% | Growth without a bigger check |
| Demand for labor | Unchanged | No replacement story in the data |
Note the last row. Labor demand unchanged means AI did not cut the team. It changed what the team spent its day on, which is a workflow debate rather than a headcount one.
Why the median startup barely moved
Most coverage of this paper skips the distribution. For the typical startup, revenue hardly budged in either arm. Almost all of the gain sat with the firms at the 90th percentile and above.
Chamath Palihapitiya read that curve the same way in his September 2026 post on AI ROI: AI multiplies output rather than adding to it, so the strongest firms pull further ahead. That framing is his, not the paper's.
Identical tools do not produce identical outcomes. They widen the distance between companies that reorganized and companies that subscribed.
The subscription is not the edge
Ramp's AI Index tracks card and bill-pay data from more than 70,000 US firms. Median AI spend stands at $11.95 per employee per month on July 2026 data, published 12 August 2026. The June letter put the same median at $11.38.
Compare that with the average monthly labor cost of about $8,500 per employee, the baseline Social Capital cites in its own ROI deep dive. AI spend only needs to return roughly 0.15% of that cost to pay for itself.
So the money is not where the decision is. Two companies pay the same $11.95 and end up an order of magnitude apart, because one knows which processes the tools should touch and the other is still guessing. That is how AI tool sprawl starts.
What MIT found, and what Chamath added
MIT Sloan summarized the workflow literature in April 2026, and its conclusion points the same way. AI's largest effect comes from how an entire workflow is sequenced, grouped and handed off between people and machines.
Chamath's gloss goes one step further: group the AI steps so a person checks the work once, not after every step. Treat that as his reading of the research, not a quote from it. The finding is well supported; the one-check rule is a design choice built on top.
The line every consultant quotes here carries the same caveat. Automation applied to an efficient operation magnifies the efficiency, and applied to an inefficient one it magnifies the inefficiency. It is commonly attributed to Bill Gates (1995).
5ac sells the map, not the license
Our reading of the 515-startup result is blunt. Vietnamese SMBs do not fail at AI because they bought the wrong model. They fail because nobody wrote down where an agent belongs in their process, and where it does not.
That map has three parts: which processes an agent can own end to end, which it supports while a person approves, and which stay untouched because the rule is unclear or the downside is too big.
Identical tools, identical credits, identical training. The only variable left is whether someone drew the map.
Multi-agent orchestration, meaning several agents with separate roles, memory and permissions, only becomes useful after that map exists. Our orchestration work starts from the process.
Do this before you pay for another seat
Four steps, all cheap, all reversible.
- Record one week of work: the tasks, who did them, how long they took.
- Sort by function, not by tool. Product, strategy, sales, operations. The treated firms won in product development and strategy, where most small teams never look.
- Mark the approval point for each process, once per batch instead of after every step.
- Run one process for four weeks and log hours, errors and approvers. A Kanban board turns "AI feels useful" into a number you can argue with.
If the numbers do not move in four weeks, kill the project. That costs less than a year of subscriptions nobody switched on.
Takeaway
The 515 startups bought what you can buy tomorrow. The winners were shown how peers reorganized around those tools, then did the reorganizing themselves. Pick one process this week, write down what it costs you today, and put the human approval where it belongs.
Then read our note on AI for Vietnamese SMBs before you add another seat. If budget is the open question, G-Company OS pricing is the next stop.
Frequently asked questions
What did the 515-startup AI field experiment find?
Kim, Kim and Koning (2026) randomized high-growth startups inside one accelerator into two groups. Both groups received the same AI tools, credits and training; only the treated group saw how other firms had reorganized production around AI. That group found more use cases, shipped more work and finished the quarter with roughly double the control group's revenue, without adding headcount.
Why did the treated startups earn more when they had the same AI tools?
The treatment was information, not technology. Seeing how peers reorganized work around AI pushed founders to search across a broader set of company functions. The differentiator was knowledge about where AI creates value, not access to a model.
What does this mean for a Vietnamese SMB with a small AI budget?
The tool bill is not the deciding factor. Measured against the average monthly labor cost that Social Capital cites, AI spend only needs to return a fraction of a percent to pay for itself. The scarce input is a written map of which processes an agent should own, which it should support, and where a person approves.
Where does the "group the steps so a person checks once" rule come from?
MIT Sloan (April 2026) summarizes research showing AI's biggest impact comes from how an entire workflow is sequenced, grouped and handed off between people and machines. The narrower rule about grouping AI steps so a human checks the work once is Chamath Palihapitiya's interpretation of that research, not a direct quotation from the MIT work.
Sources
- Kim, Kim & Koning (2026), "Mapping AI into Production: A Field Experiment on Firm Performance" — INSEAD Working Paper 2026/20/STR; 515 startups, +44% use cases, +12% tasks, +18% paying customers, 1.9x revenue, −39.5% external capital, labor demand unchanged. Faculty listing with the full abstract: Harvard Business School.
- Chamath Palihapitiya, "AI ROI" post (25 September 2026) and the Social Capital deep dive — the 90th-percentile reading, the 8,500 USD labor baseline, and the one-check interpretation of the MIT workflow research.
- Ramp AI Index, August 2026 letter (12 August 2026, Ara Kharazian) — median 11.95 USD per employee per month on July 2026 data. Earlier median 11.38 USD in the June 2026 letter.
- MIT Sloan, "How AI is reshaping workflows and redefining jobs" (22 April 2026) — sequencing, grouping and handoff as the source of AI value.
- AEA RCT Registry, trial AEARCTR-0016746 — preregistration of the field experiment, fixing outcome measures before results.
The tool is the cheap part. The map is the work. At 5ac.vn we map which processes an AI agent should own, which it should support, and where a person approves, before anyone buys another seat.