Google Research and the Network for Teaching Entrepreneurship (NFTE) reported a randomized trial in which students who practised with AI teammates built to challenge them averaged 67.6% on their final business plans, against 60.5% for control students who had standard AI tools (arXiv 2611.02016, posted November 1).
The paper puts the gap at 7.1 percentage points (Cohen’s d=0.62, p=.047) across 52 students in four summer programs. By our arithmetic, that is about 12% higher. The authors call the findings “preliminary”.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how a harness around models shapes what they deliver. It teaches the general idea only and does not discuss Vantage or this study.
The key insight:Google Research and NFTE reported on November 1 that students who practised with Gemini-run AI teammates built to challenge them averaged 67.6% on their business plans, against 60.5% for control students with standard AI tools, in a 52-student randomized trial.
The platform is Vantage, which the paper describes as “an experimental tool developed by Google Research”. Students work on business tasks in chat with several AI teammates. A single “Executive LLM”, which the paper says is Gemini 3.1 Pro, writes every teammate’s turn.
The teammates are prompted to test the student. The paper gives examples: a teammate “may suggest a flawed product idea that the student must evaluate, introduce a competing idea the student must reconcile, or state a false claim to test whether the student verifies it”.
They also hold back. The paper says the AI peers “never complete, solve, or submit the task for the student, even when the task is framed as a team deliverable”. It adds that the Executive LLM eases off when a student shows signs of frustration.
Average Capstone Business Plan score for control students with standard AI tools (grey) and students using Google’s Vantage AI peers, with two non-randomized subgroups of the Vantage group by practice tasks completed. Data: Globerson et al., Google Research and NFTE, arXiv 2611.02016. The 12% relative gap is our arithmetic.
What Was Posted on X
New Google paper finds that students who practiced with AI teammates that push back wrote better business plans than students using normal AI tools.
Skills like critical thinking and collaboration matter for starting a business, but they are hard to teach and hard to practice at scale.
They extended Vantage, a Google Research platform where students solve business tasks in chat with AI teammates run by Gemini 3.1 Pro. The teammates pitch flawed ideas, disagree, and state false claims, but never do the task for the student.
In a trial with 52 students, the Vantage group averaged 67.6% on the final businessplan, versus 60.5% for controls using standard AI tools. Students who did 5 to 6 practice tasks averaged 79.7%, while those who did none scored 59.3%.
If you build AI for learning, make it a teammate that challenges people, not an assistant that finishes their work.
A post by Rohan Paul at 21:00 UTC on 7 November summarising the paper: it says the Vantage group averaged 67.6% on the final businessplan against 60.5% for controls in a trial with 52 students, and that students who did five to six practice tasks averaged 79.7%. The closing advice to build AI teammates that challenge people is the poster’s view, not a finding of the paper.
How the Trial Worked
The trial ran inside NFTE’s BizCamp summer programs at middle and high school sites in the Southeast, Midwest and West of the United States. Students were randomized within classrooms at a ratio of about 65:35 in favour of treatment, the paper says.
Both groups got the same six tasks, written by Google Research and NFTE. Treatment students did them in Vantage; control students got the same descriptions on a worksheet and had access to standard generative AI tools. The outcome was the BizCamp Capstone Business Plan, scored by teachers or volunteer judges on NFTE rubrics that the paper says were developed before and independently of the study.
Nine sites took part, but the impact sample is four randomized sites: 36 treatment students and 16 controls. The paper says four sites had no control group and one recorded no business-plan scores. The gap held in every one of the four sites, ranging from +2.0 to +9.9 points.
More Practice, Higher Scores
Inside the treatment group, scores rose with the number of Vantage tasks completed. The 10 students who did none averaged 59.3%, close to the controls’ 60.5%. The 5 who completed five or six tasks averaged 79.7%.
The paper estimates a gain of 2.3 to 2.9 points on the businessplan for each additional task. That split is not randomized: students chose how many tasks to do, so it shows an association rather than a measured effect of each task.
The Business Lever
Most AI assistants are built to finish the user’s work. Vantage is built to argue, to plant errors and to refuse to do the task, and in this trial that design beat standard AI tools on a graded outcome. Our earlier coverage tracked how labs measure the opposite trait, sycophancy, as a product axis.
As we read the paper, the tested variable is productdesign rather than model quality: the AI peers were prompted to plant errors and to refuse to do the student’s task.
As we read the dosage numbers, students assigned to Vantage who did no tasks scored like the controls, so the gap appears tied to using the tool rather than to being assigned it.
As we read it, this is one small trial in summer programs for school-age students, and the paper itself calls the findings preliminary.
Globerson et al., arXiv 2611.02016
“These preliminary findings suggest that orchestrated AI peers can meaningfully support short-term performance on tasks requiring complex, durable skills.”
Three Implications
FOR AI PRODUCT TEAMS The paper says the Vantage peers never complete, solve, or submit the task for the student, even when the task is framed as a team deliverable.
FOR EDUCATION BUYERS The paper says business plans were scored on NFTE rubrics developed before and independently of the study, with treatment ahead in all four randomized sites.
FOR RESEARCHERS The paper says a planned before-and-after Vantage skills test was not analyzed because only 24 students had valid scores on both.
The essay describes “an operating system that wraps your AI”, “a harness that compresses your thinking, your team’s feedback, and the model’s raw capability into something replicable.”
As we read it, Vantage is such a harness around Gemini 3.1 Pro: the same model, wrapped in rules to challenge and to hold back, produced the graded gap in this trial. The essay does not discuss Vantage or this study.
What Is Not Established
The sample is small: 52 students in the impact analysis, with 16 controls, and the p-value of .047 sits just under the usual 5% threshold. Rubrics differed by site, though scores were standardized as a share of possible points. A planned before-and-after Vantage skills test was not analyzed, the paper says, because only 24 students had valid scores on both. By our count, 20 of the 24 authors list Google Research as their affiliation and 4 list NFTE. The study does not show lasting skill gains, results for adults, or effects outside entrepreneurship. This piece makes no forecast.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
In a 52-student randomized trial, students who practised with Gemini-run AI teammates that disagreed, planted errors and refused to do the work averaged 67.6% on their business plans, against 60.5% for peers using standard AI tools. The paper calls the result preliminary.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Figures and quotations are from the paper by Google Research and NFTE posted to arXiv on 1 November 2026 (2611.02016). We did not rerun the analysis. Percentages comparing scores and the author count are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
