How people use AI matters more than whether they have it.
A field experiment with 388 Fortune 500 employees. Microsoft designed it. Their treatment condition — the one that won — was our AI Mindset curriculum. Here's what happened when it went head-to-head with standard Copilot training.
Same tool.
Different scaffolding.
194 pairs, randomized. Everyone got Microsoft Copilot. Everyone tackled the same two real-world tasks. The only variable: the structure surrounding their AI use — one behavioral, one cognitive.
The Strategy Task
Each pair produced a one-page "AI Adoption Action Plan" tailored to their organizational function.
The Communications Task
Each participant drafted a strategic response addressing three distinct stakeholder concerns:
- Data governance and transparency
- Workforce transition and displacement
- Sustainability of AI infrastructure
Two interventions.
Two opposite outcomes.
Structure that was meant to help collaboration actively broke it. A brief mental-model shift measurably lifted individual output quality. The gap between the two is the story.
Forcing a protocol made pairs worse, not better.
Pairs assigned a rigid "Create-Out-Loud" protocol scored nearly 5 points lower on quality — and were eight times more likely to produce no document at all.
The protocol required synchronous meetings, verbal discussion, then AI drafting from the transcript. It imposed real coordination costs without adding information the AI could actually use. Less than a quarter of treatment pairs even managed to follow it.
The pattern was uniform across every rubric dimension: opportunities, risks, action plan, insight.
Document Quality Score (out of 22)
A single session of reframing lifted 15 more people to top quality.
77% of Mindset-trained participants hit a perfect score — compared to 62% of controls. Odds ratio of 2.07, significant at p = 0.022.
Partnership training didn't teach anyone a new Copilot feature. It shifted the mental model — from "search engine" to "thought partner." That reframing alone shifted the probability of producing top-tier work.
The effect held at the ≥18 threshold too (OR = 1.87, p = 0.049). A ceiling effect in the AI grader — 68% of documents scored perfectly — made the continuous model noisy, but the binary signal was clear.
Hit-Rate · Perfect Score (20/20)
Microsoft's researchers, on what they found.
A short from Microsoft Research on what the study uncovered — including why they concluded the mindset intervention mattered more than the tools themselves.
The full experiment is live on Microsoft's research microsite, with interactive data and supplementary appendices.
Visit the micrositeACCESS the one-page explainer
Not all "teamwork"
is the same.
The research categorized every treatment pair by what they actually did. Three of the four modes performed no better than working alone. Only one actually worked. The rest were a social placebo.
Natural Use
Pairs left to work however they preferred — the reference point for what "normal" AI adoption looks like.
Reference · 15.6 quality scoreTrue Joint
Partners actually shared a conversation and the prompt. They thought together, then asked together. The highest-reported experience scores.
Best outcome · 12.5 qualityParallel Play
Partners met and talked — then prompted the AI individually. Looks like collaboration. Produces outputs indistinguishable from working alone.
9.8 quality · no gainStranded
Tech failures. No-shows. Scheduling breakdowns. The collaboration the protocol mandated never happened. 37% of treatment pairs.
12.3 quality · forced frictionScaffolding, layered.
The study mapped two approaches to a three-layer architecture for AI adoption. Each layer depends on the one below — and the layers only work in order. Skip a step and the layer above collapses.
Behavioral Scaffolding
Mandated interaction protocols. The most ambitious layer — and the hardest to execute. Without the two layers below, it actively backfires. (This is what Task A tested.)
Cognitive Scaffolding
Reframing AI as a thought partner. Low-cost, portable, measurable. Works when the foundation is solid. (This is what Task B tested — and it lifted quality.)
Mechanical Fluency
People can actually operate the tool. Without this, nothing above it functions. The prerequisite no one talks about because it feels too obvious — until it isn't.
Practitioners saw it
coming.
The pattern the research surfaced echoes what the practitioners inside the study have been saying in the field for years. Mindset first. Mechanics later. Mandates last, if at all.
If you had to choose — an employee with an AI-first mindset navigating an analog process, or an employee with an analog mindset navigating an AI-first process — which would you pick?
When we stop treating AI like a tool and start treating it like a teammate, we unlock its real potential.
Enable, optimize, reinvent. You enable the people, you optimize the work, and only then do you scale transformation.
With AI Mindset training it was 77 [out of 100 clearing the bar]. That's 15 more people reaching top-quality work — from a single session.
"The sole differentiator was behavioral."
Microsoft's paper is explicit: treatment and control had identical Copilot access. No new features. No technical capabilities. The only thing that differed between the two groups was the training itself — ours versus theirs.
Their finding: participants trained on the AI Mindset curriculum were more than twice as likely to produce top-quality work, with an odds ratio of 2.07 (p = 0.022). The paper concludes that performance gains in AI use stem from how people engage with the system, not from what they know about its features.
In plainer language: the same tool, in the hands of differently-trained people, produces materially different results. That's the lift Microsoft measured. That's the curriculum.
What we actually
teach.
The treatment condition was our curriculum — the same AI Mindset training Conor Grennan has delivered to enterprises for years. No new Copilot features. No technical wizardry. Three behavioral components, all designed to shift how people work with AI.
Cognitive Reframing
Challenge the default assumption that AI is a search box. Reposition it as a collaborator that deserves the same engagement a human thought partner would — context, follow-ups, real-time correction.
Mental Model Replacement
Put the old model in direct contrast with the new: the "search engine" frame (one-shot, extractive) versus the "thought partner" frame (multi-turn, generative). The "smart intern" metaphor makes it actionable.
Guided Practice
Structured exercises in iterative, conversational prompting. The goal isn't to memorize syntax — it's to build fluency with dialogue itself. The rest follows naturally.
The people behind the study.
Microsoft Research
- Alex FarachSenior Data Scientist
Corresponding author - Alexia CambonDirector, Applied Research
- Lev TankelevitchSenior Researcher
- Connie HsuehSenior Researcher
- Rebecca JanssenSenior Applied Scientist
Gap Inc. Leadership
- Sven GerjetsChief Technology Officer
- Mario DiazSenior Manager
Future Skills
Training Curriculum
- Conor GrennanFounder
AI Mindset
This is the curriculum
Microsoft tested.
You just read what happened when it ran alongside standard Copilot training. The AI Mindset enterprise program is what we deliver every week — to technology, finance, retail, and professional services teams rolling out AI at scale. The same framework, the same behavioral shift, the same measurable lift.