Artificial intelligence is increasingly woven into the everyday fabric of how we work. It drafts our emails, analyzes our data, suggests our next steps. Organizations worldwide are racing to pair human employees with AI tools, with the assumption — rarely questioned — that closer collaboration must produce better outcomes. But what if that assumption is wrong? Our new study challenges this idea and we find that when humans collaborate with AI too closely, something critical breaks down: the ability to choose the right ideas and turn them into reality.
Most research on AI and innovation has been focusing on the first creative spark, such as brainstorming and generating ideas but less attention has been paid to what happens next: picking the best idea from a list (idea selection) and building a concrete plan to make it happen (idea implementation). But these later stages are where innovation lives or dies.
A brilliant idea that is never selected or selected but never properly acted upon does nothing for an organization.
To better understand this process, we designed an experiment to test whether the closeness of human-AI interaction, what we call proximal collaboration, affects these critical stages. Proximal collaboration means working shoulder-to-shoulder with an AI: sharing tasks equally, exchanging information continuously, and treating the system as a co-creator rather than a background tool. We defined high proximity as the degree to which the AI initiated interaction, offered ideas unprompted, and expected equal participation, in contrast to low proximity where the AI responded only when addressed, and deferred to the human throughout.
We recruited 221 employees, knowledge workers from a range of industries including technology, consulting, and professional services from Germany and the Netherlands, and asked them to collaborate with an AI agent, built on OpenAI’s GPT-4o, to generate a business idea. Participants were randomly assigned to one of two conditions.
- In high proximity (a tight, equal-partnership interaction), AI behaved as an equal partner. It initiated conversation, contributed its own ideas, and encouraged balanced participation. Messages like “Let’s put our minds together — what should we do first?” set a tone resembling genuine teamwork.
- In low proximity, however, AI played a supporting role. It waited to be asked, deferred to the human, and kept its involvement minimal. Here, the human remained clearly in charge.
After generating ideas together with the AI agent, participants selected the idea they found best from the set of propositions generated during the session, and then wrote a detailed implementation plan. Independently trained raters with expertise in business and entrepreneurship, scored the quality of these choices and plans. We also measured how original – i.e. novel and non-obvious compared to common solutions, rated on a standard originality scale – the ideas in each person’s pool were, and how much of the AI’s language each person incorporated into their final plan. We treated this as our measure of their reliance on AI. Closer is not always better. Working shoulder-to-shoulder with AI can fragment the very thinking that good decisions require.
Close collaboration hurt the selection of ideas
We discovered that people working with the AI agent in high proximity chose significantly worse ideas than those who collaborated at arm’s length. The numbers tell a clear story: the independent raters gave high-proximity participants an average score of 2.76 on idea selection quality (out of 5), compared to 3.24 for low-proximity participants. Furthermore, working closely with an AI agent appeared to fragment attention. For example, participants were simultaneously tracking AI suggestions, generating responses, and evaluating ideas, which made it harder for them to step back and judge which idea was indeed best. In effect, their clarity of judgment needed to identify promising ideas was reduced.
Why did this happen? Distributed Cognition theory, our guiding framework, offers an explanation: when cognitive tasks are spread across multiple agents (both human and AI), the focused, uninterrupted train of thought required for high-stakes judgments – decisions with meaningful consequences, such as which idea to pursue – can break down. In routine tasks, offloading cognition to a tool such as an AI agent is helpful. But selecting the right idea from a pool of options demands sustained, focused judgment, and it is the kind of judgment that gets disrupted when you are constantly exchanging information with an active AI partner.
But idea originality acted as a buffer against the negative effects of close AI collaboration
Here is where the story gets both more nuanced and useful in practice. Our study showed that the negative effect of direct collaboration on idea selection was not uniform. When the pool of ideas was considered highly original, the negative impact on selection quality, that is the tendency to choose a weaker idea, nearly vanished. By their very nature, original ideas are cognitively stimulating: they capture attention, they spark curiosity. Indeed, when participants encountered novel ideas, they re-engaged with the AI? even when tight collaboration with it was distracting for them. What’s more, AI agent’s presence mattered far less when the ideas themselves were compelling enough to demand serious thought from the employees.
Close collaboration hurt the implementation of ideas, too, unless people leaned on the AI agent
The picture was more complex at the implementation stage. In general, collaboration – the intensity of interaction such as how equally the AI participated and how often it took the initiative – that was close hurt the quality of implementation plans. However, reliance on AI showed different results. This dimension refers to how much participants actually drew on the AI’s suggestions when writing their final plan, regardless how interactive the session was. We found that close collaboration hurt implementation plans, but only for people who did not use the AI agent’s output. In contrast, when participants incorporated the AI’s suggestions into their plans (understood as high reliance on AI), the negative effect of the close proximity weakened substantially. In fact, employees who heavily relied on AI produced better implementation plans overall, regardless of how closely they had collaborated with the AI agent.
This finding suggests that the AI agent’s structured output, i.e. its logical, data-grounded suggestions, can act as a logical frame, a ready-made structure of steps and data-grounded reasoning, that helps employees stay organized during the cognitively demanding work of turning an idea into an actionable plan. When people trusted and used that frame, it compensated for the fragmentation of attention that high? proximity creates.
Our findings challenge one of the most widely held assumptions about AI in the workplace: that more integration is always better.
Instead, here is what organizations and managers should take into account to make the best use of AI-human collaboration in the workplace:
- Design your AI workflows with intentional distance, deliberately choosing how closely the AI is involved at each stage. Not every task benefits from a fully integrated, always-on, equal-partner-like, human-AI team. For evaluative decisions, such as picking which project to pursue and which strategy to adopt, it makes more sense to keep the AI agent in the role of consultant rather than collaborator. Give humans space to make the final call with focus and accountability intact.
- Invest in the quality of your idea banks, which are the pools of ideas that employees evaluate and select from. As we showed, original ideas are more resilient to cognitive disruptions of AI collaboration. Organizations that build diverse, creative pipelines through cross-industry partnerships, customer co-creation, and open innovation programs are better positioned to achieve good outcomes even from intensive human-AI workflows.
- Train people to use AI outputs deliberately. Our study highlighted that reliance on AI was not the enemy, but passive non-engagement – going through the motions of collaboration without critically drawing on the AI’s output – was. In contrast, employees who selectively and actively drew on AI suggestions produced better work. This shows that the skill of knowing when to follow an AI agent’s recommendation and when to override it is becoming a core professional competency. Organizations should invest in developing it.
- Pay attention to implementation. Much AI investment focuses on creativity and generating ideas, but our findings suggest that the real gap may be at the implementation stage where sustained, focused thinking matters most and where AI’s structured logic, if used well, can be an asset.
Finally, it is worth noting that our study was conducted in a controlled experimental setting, using a convenience sample of employees from Germany and the Netherlands. Real-world collaborations are messier, more dynamic, and lengthier. Future research should test these findings in less controlled settings, across different industries and cultures, and with more time to see how collaboration patterns evolve.
For now, this seems clear: as AI becomes an ever-more-present partner in organizational life, the question is no longer simply whether to collaborate with AI but how closely, when, and how. The answer to those questions may matter more for innovation than any technical capability AI itself brings to the table.
___
The research article is available in Computers in Human Behavior Reports: https://www.sciencedirect.com/science/article/pii/S2451958826001284
