Founders and AI

Early-stage founders make decisions under what the literature calls Knightian uncertainty: outcomes cannot be reliably predicted, and relevant information may not yet exist. Sensemaking research looks at how people form a workable interpretation so they can act. Cognitive frames are the structures of knowledge they use to organize and interpret what they encounter.

Generative AI introduces advice into that process with very little practical or interpersonal friction. Research on advice-taking suggests that reliance rises with task difficulty and falls with domain expertise. Human–AI research also shows that the order of a person’s judgment and AI advice affects reliance. Less is known about what happens when AI enters while someone is still trying to understand the decision.

I wanted to understand how that starting point affected their use of AI: did they already have a way of thinking about the decision, or were they relying on the model to help them form one?

In the research, I called this “frame availability”: whether someone had a working interpretation of the problem and a basis for judging possible responses before asking AI.

I conducted 29 semi-structured interviews with pre-seed technology founders in the US and UK. Each lasted 35–45 minutes and focused on particular decisions they had made with AI. I used criterion-based purposive sampling, recruiting through founder networks, LinkedIn, and referrals. Participants described past decisions in detail, following the logic of the critical incident technique.

I analyzed the interviews using the Gioia methodology, tracing participants’ accounts from initial codes into broader themes. The unit of analysis was the decision episode rather than the founder. I coded a frame as available when the connection between the problem and an appropriate response was evident before substantive AI input, and as not yet available when it wasn’t.

I built the literature review through Google Scholar, covering sensemaking, cognitive framing, advice-taking, and human–AI reliance. The Gioia analysis provided an audit trail from participants’ own terms to the theoretical dimensions.

Figure 1. Project plan and delivery, June 15–August 31, 2026. The evaluation moved nine days because of participant availability; completing the qualitative work early kept the project on schedule.

3.1The same founder could use AI in very different ways

For nine founders, the interviews showed different patterns across decision domains. One founder had a clear understanding of her market but no frame for an unfamiliar codebase. Another had a frame for code architecture but not for pricing.

Frame availability therefore varied within founders. It made more sense to examine their understanding of a particular decision than to treat it as a stable personal characteristic.

3.2Having a view first changed how founders assessed the answer

When founders already understood the problem, they described checking AI’s output against their own judgment. When they didn’t, the model could provide the first explanation they worked from. Sycophancy and surface-level validation could reinforce anchoring on that first explanation. The concern was that AI could provide the first convincing account of the decision while the founder was still making sense of it.

Figure 2. The process model follows two pathways, depending on whether the founder had a working understanding of the decision before consulting AI.

3.3Checks often came after something went wrong

Regulation was the largest dimension in the codebook. Founders checked more carefully when the stakes were high, returned to human feedback, set boundaries around particular domains, and overrode the model where they already had a frame.

Those habits often followed a reality check, such as a lost deal or investor pushback. The correction came from external feedback rather than the founder noticing the problem independently. When feedback is slow or a decision is irreversible, that correction may arrive too late.

4.1Making room for an independent first thought

COM-B identified physical opportunity as the main barrier: consultation was immediate, with no built-in pause. Automatic motivation was a secondary factor. PRIME helped explain why an intention to think critically did not necessarily translate into action when the founder opened the chat.

The Behaviour Change Wheel pointed to environmental restructuring as the main intervention function: change the interaction so there is space to think before advice appears. I kept enablement for founders who could not proceed on their own. I rejected restriction because the aim was to support better use of AI, rather than prevent its use.

4.2Choosing where the pause belonged

I compared six concepts against six design criteria and combined elements of several into the final concept. Two criteria were decisive. Triggering at the moment of consultation ruled out concepts that required earlier action. Keeping the founder’s answers out of the model’s context ruled out approaches that asked them to explain their position in the chat.

Figure 3. Comparing six concepts against the design criteria.

4.3Two questions before the answer

I built the intervention as custom instructions in Claude, installed once and included in the model’s context on subsequent turns. Founders did not need to remember to invoke it for each decision. The two questions drew on the diagnostic and prognostic aspects of cognitive-frame research: defining the decision and what a successful response would achieve. For a consequential or hard-to-reverse judgment request, the instructions ask the model to pause and have the founder write down:

1. What decision am I actually making?
2. What would have to happen for me to call it the right decision?

The founder writes their answers somewhere outside the chat and confirms when they’re done. The model then gives its advice, and the founder can compare it with their own starting point.

Keeping the notes outside the chat was intended to reduce the chance that the model would simply echo the founder’s view. The pause was designed to last one turn and trigger once per decision episode. It avoided alert styling and never suggested what the founder should write. Because they were written instructions rather than a hard-coded rule, I tested whether the model followed them.

Five founders from the interview sample installed the instructions and used them for one live, unresolved decision. All five completed installation; two needed to switch to a desktop browser and a personal account.

The pause triggered on the first qualifying question in all five sessions. I observed no false triggers, but the study deliberately focused on decisions that should trigger it, so it offered limited evidence about everyday use. No participant used the “I can’t say” route.

The evaluation did not suggest weakening the core mechanism. Both trigger conditions, the two questions, and both refusal routes stayed unchanged. Following the MRC framework, the next step was to repeat the evaluation with these refinements.

I presented the research and intervention to a panel of two pre-seed founders and design engineering experts in a 15-minute presentation followed by 30 minutes of questions.

The interviews suggest that AI can play different roles depending on whether someone has already formed a view of the decision. That gives designers a specific moment to investigate: what happens just before the advice appears.

The study contributes to sensemaking research by showing how AI can encounter an existing frame or become involved while one is still forming. It also suggests that prior knowledge should be considered at the decision-domain level.

The evaluation provided evidence about feasibility, usability, and acceptability. It did not establish whether the pause reduced anchoring or improved decision quality.

The study also had limits. I coded the interviews myself, participants described past events, and everyone in the sample already used AI. The evaluation involved one model, one use per person, no comparison group, and no ground-truth measure of decision quality. All five participants had also taken part in the interviews, creating a risk of demand characteristics: their understanding of the research could have influenced their responses.