How much autonomy is too much autonomy? While some companies take a rigid approach to assigning tasks, it’s become increasingly popular to allow employees a greater degree of freedom in determining what they work on, and with whom. Companies such as Spotify, GitHub, and Google have publicized their policies allowing employees to self-select the projects and teams they work with, arguing that by spurring higher levels of ownership and creativity, this strategy leads to better, more innovative ideas.

This may seem intuitive, but our recent research suggests that it’s possible to take autonomy too far. We conducted a field experiment with more than 900 students in an 11-week undergraduate course on lean startup entrepreneurship, in which students were grouped into three-person teams and tasked with developing a startup idea and pitch deck. Unbeknownst to both the students and teachers, we randomly assigned students to one of four conditions, varying whether they had autonomy over the ideas they’d work on and the composition of their teams. In the first group, students were both assigned to teams and told what idea to pursue. In the second group, students could choose their teammates, but they were assigned an idea to work on. In the third group, students were assigned to teams, but were given the autonomy to choose their own idea. And in the final group, students were allowed to choose both their teammates and their ideas.

The teams produced pitch decks which were evaluated by a panel of practitioners, including entrepreneurs, angel investors, and venture capitalists. To ensure fair evaluation, names and photos were redacted, judges were not allowed to communicate with each other before submitting their responses, and each pitch deck received three independent evaluations. The judges were first asked to assess the pitch decks’ performance in five areas: novelty, feasibility, market potential, likelihood of success, and likelihood of inviting the team to a follow-up meeting. Next, once all the pitch decks had been individually scored, the judges were asked to allocate a fictional investment budget of $1 million across all the projects they evaluated (they could spend some, all, or none of their budget).

So, what did we find? Interestingly, our analysis suggested that some — but not total — autonomy yielded the best results. The teams that weren’t allowed to choose either their ideas or their teammates performed the worst, but those with full autonomy performed only marginally better: They were rated as less than 1% more likely to succeed than the teams with no autonomy.

Conversely, the teams that were given some autonomy significantly outperformed both those with full autonomy and those with no autonomy: Teams that could choose either their ideas or their teammates (but not both) were rated as 50% more likely to succeed than those with no autonomy, and 49% more likely to succeed than those with full autonomy. They also received 82% more of the fictional investment budget than those with no autonomy, and 23% more of the budget than those with full autonomy. Overall, the top performing teams were those who were assigned their teammates but allowed autonomy over their ideas, closely followed by those who were assigned ideas and allowed to choose their teams.

Why might this be? We found that the main factor driving these effects was that the teams with some autonomy were best able to match ideas to team members. Teams with assigned teammates and autonomy over ideas could choose ideas based on members’ interests, while those with assigned ideas and autonomy over membership could strategically choose teammates who were best suited to work on the idea. Teams without autonomy over their membership or their ideas had no way to do this kind of matching, and it was similarly challenging to match teams and ideas effectively when both were up for discussion.

In addition, a secondary driving factor was that getting to choose both their teams and ideas fueled overconfidence, which is known to negatively impact performance. We asked the teams to describe how their own performance and capabilities compared to those of a reference group, and we found that the teams with total autonomy were most likely to overestimate their own abilities.