Scaffolding Human-AI Collaboration: A Field Experiment on Behavioral Protocols and Cognitive Reframing
Paper URL: arXiv:2604.08678
arXiv subjects: econ.GN (General Economics), cs.HC (Human-Computer Interaction)
Giving employees access to generative AI is not the same as helping them use it well. This field experiment asks a more useful question: once people have the same AI tool, what kinds of support improve their work?
Farach and colleagues studied 388 employees at Gap Inc. All participants had access to Microsoft Copilot. The researchers tested two interventions. The first was a behavioural scaffold: a structured pair protocol, called Create-Out-Loud, that required colleagues to use AI together. The second was a cognitive scaffold: partnership-oriented training that encouraged employees to treat AI as a thought partner rather than simply a tool for producing an answer.
The clearest result concerns the pair protocol. Employees assigned to it completed fewer documents and received lower LLM-graded quality scores than employees who used AI more naturally. This matters because the intervention was designed to improve collaboration, yet its structure appears to have added coordination and compliance costs. A rigid protocol can turn a flexible tool into another meeting, especially when the task does not genuinely require close coordination.
The training result is more tentative. Partnership training did not significantly improve average document quality in the pre-specified analysis. It was associated with a greater chance of producing a top-scoring document in an exploratory analysis. That is an interesting signal, but not strong evidence that cognitive reframing reliably improves performance.
The paper is valuable because it studies AI use inside a real organization rather than in a short laboratory task. It also distinguishes two ideas that are often treated as one: structuring behaviour and shaping people’s mental models. The results suggest that the two can have very different effects.
Its limitations matter just as much. The behavioural intervention is entangled with an AM/PM session difference, the pair task, varying completion rates, and the demands of following the protocol. The quality measure may also reward longer documents because it relies partly on an LLM grader. These issues make the narrow conclusion more defensible than the broad one: this particular synchronous protocol performed poorly in this particular setting. The study does not show that all behavioural scaffolding harms AI-supported work.
For organizations, the practical lesson is not to avoid guidance. It is to design guidance around the work. Support should reduce uncertainty, encourage good judgement, and leave room for people to adapt their AI use. Before standardizing a collaboration protocol, organizations should test whether it solves a real coordination problem or merely creates a new one.
The next studies should separate timing and task effects cleanly, measure quality without rewarding length, and test which coordination costs matter most. They should also examine whether partnership-oriented training has durable effects on judgement and work quality rather than a short-lived effect on one task.