Most AI copilot business cases I see assume one thing: the experts whose work feeds the agent will say yes when you ask them.
I used to think the only question was whether the workflow was clean enough to capture. Then I started asking the people whose names were on the pitch decks and the performance reviews. The answer is not yes.
The sales director who said no to the agent
A sales director at a B2B SaaS firm in Munich, on a coaching call last week. Her company had bought a sales enablement agent that trains itself on the top closers' decks, their call recordings, their email cadences. Three quarters' worth of her best work. She was reviewing the consent flow before her team was onboarded.
She told me she would have signed the same release if a new VP had asked for her playbook. She was not going to sign it for the agent.
I asked why.
She said the agent did not need her approval the second time it used the work. The VP would still be a person she could call about it next quarter.
That is the whole paper, on a Zoom call, before I had read it.
What the research says
Song, Zhu, Cao and Lee at NUS and Emory ran a between-subjects experiment with 61 full-time professionals in management, sales, HR and marketing. Each person did a simulated workplace task. Each one then decided whether to let the output be reused, either by a human or by an AI.
Permission to reuse was granted by 96.8% in the Human condition. 76.7% in the AI condition. Fisher's exact OR=0.11, p=.026.
The workers who did say yes to AI reuse were not giving blanket consent. They attached specific demands. Credit, measured on a 1-7 scale, came in at Hedges' g=1.30, the largest signal in the paper. Control: g=0.60. In open-ended responses, 16.7% of the AI condition named professional replacement, versus 3.2% in the Human condition.
AI and human reuse differ primarily in whether workers choose to share their knowledge.
The authors' line, not mine.
Most internal AI programs I see priced out this year treat voluntary expertise capture as a technology problem to solve. The paper measures the consent problem underneath. There is a 20-point gap baked into the pipeline before a single prompt has shipped. The workers who do consent are naming the terms: credit on the output, control over the reuse, honesty about what the agent is for.
The replacement-anxiety number is a disclosure. The sales rep, the HR partner, the marketer, the manager can all read what "my expertise becomes agent skill" means for their position.
Confront the career architecture question before you build the copilot. The workers are already past that part.