A different role for each task
A founder asking about an unfamiliar market keeps the lead in the conversation. Building a prototype together gives the assistant more room to shape the result. Routine editing often means requesting a bounded change and checking the output. Our study distinguishes these arrangements from autonomous coordination across tools, which founders mostly described as something they wanted in the future.
AI takes different roles in a venture
The task changes how much initiative founders give an assistant.
- Getting oriented
- The founder directs the inquiry. AI retrieves information, explains unfamiliar subjects, or offers feedback.
- Creating together
- The founder and AI shape an idea or prototype through repeated suggestions and revisions.
- Routine work
- The founder specifies a bounded task, such as proofreading. AI carries it out with little discretion.
- Coordinating across tools
- Mostly envisioned. The founder would set goals and limits; AI would initiate actions across tools and channels.
In the interviews, founders remained cautious about allowing assistants to contact customers or partners directly.
Delegation stays conditional
Founders usually kept reviewing the assistant’s work, particularly when a mistake could affect customers or the venture’s reputation. They wanted more proactive help, such as reminders and warnings about unresolved issues. Some imagined handing over bounded tasks after repeated checks had built confidence. Their accounts suggest that delegation can expand or be taken back; they do not establish a fixed path toward full autonomy.
Preparing the next customer conversation
Illustrative application: a founder could ask an assistant to summarize interview notes and challenge an assumption about customer demand. The founder would check the notes and decide what to test next. The assistant could then draft follow-up messages and suggest meeting times, while the founder keeps approval over what goes to customers. These choices give AI a useful role without treating plausible suggestions as evidence of demand.
Watch what becomes part of the workflow
Founders described storing pitch decks, code, and other venture knowledge in their assistants. This context made the tools more useful, but also made switching harder. Team rules often lagged behind everyday use. A practical question is therefore what the team needs to retain outside any one tool: the knowledge behind decisions, who can act on it, and which actions still need approval.
About the study
Who we spoke with
We conducted semi-structured interviews with 19 founders of active German startups between December 2024 and August 2025. Participants already had experience using AI in innovation work. Recruitment combined accessible networks with deliberate selection for variation, including ventures receiving university startup support and ventures featured in startup magazines. The sample covered software and hardware businesses at different stages of development.
What the interviews captured
Founders described the tools and tasks involved, what they delegated, and how they judged the results. We also asked about team practices and the support they wanted from an ideal assistant. Video interviews were audio-recorded with consent and transcribed verbatim. The findings therefore include both accounts of current use and desired capabilities; autonomous coordination was mostly an expectation rather than established practice.
How we analyzed the accounts
Two researchers divided the transcripts and coded them line by line, staying close to founders’ own descriptions. They then jointly compared codes and grouped them into broader themes. The author team discussed the resulting categories and their interpretation. This Gioia-inspired analysis connected motives for adoption and tool requirements with task types, patterns of oversight, and emerging team rules. Data collection and analysis proceeded iteratively until further interviews added little to the developing categories.
What the findings support
The study explains how these founders described using and controlling AI. It does not measure productivity, output quality, or venture success, and it did not observe their work over time. All participants were active users in Germany, so the findings do not represent founders who avoid or abandon AI. The accounts also reflect the tools available during the interview period. The example above applies the findings hypothetically; it was not an evaluated workflow.