Updated
Where AI helps when you’re building a business
How founders can use AI for research and execution while checking assumptions and keeping responsibility for decisions
AI can help a founder put together a business plan, product requirements, code, and marketing copy. A small team can get through work that would otherwise take more time or require help it can’t yet afford. That makes it easier to start building, and it also makes it easier to get a long way into an idea before checking the assumptions behind it.
You can end up with a roadmap, landing page, customer personas, and competitive analysis that all came from the same description you put into a prompt. That gives you a more detailed plan without necessarily teaching you anything further about the customers you expect to buy it.
Calling the system a co-founder can make that distinction harder to keep track of. It can help you produce work and examine information, but you still need to establish where its conclusions came from and whether they give you a sound reason to proceed. The amount it produces tells you very little about whether customers will pay for what you’re building.
Give the analysis something outside your own assumptions
When you describe an idea to an AI system, you choose how to frame it. You explain the problem, the customer, and why your solution should help. If those things are mostly assumptions, the answer is working from those assumptions too.
The model can organize them into a convincing explanation. Because the response comes back in someone else’s words, it can feel like another party has evaluated the business. In practice, you may be reading a more detailed version of your own proposal.
A generated customer persona is a good example. It can include buying habits, frustrations, goals, and objections before you’ve spoken with anyone who might buy the product. Those details can help you plan what to investigate. Treating them as things you already know gives the rest of the plan a weak foundation.
A competitor analysis needs similar care. It can help you organize a market and find areas to explore. You still have to check whether the information is current, whether the sources support the claims, and whether it explains why customers choose one product over another. Language models can produce confident factual errors, a limitation OpenAI describes in its research on hallucinations.
Customer interviews, sales conversations, support tickets, and usage data give the analysis material that didn’t originate with the founder. A conversation might show that the customer describes the problem differently. Sales research might reveal that the person who wants the product has no authority to buy it. Usage data can show that a feature the team spent time on gets very little attention.
AI can help work through that evidence. It can organize interview notes, compare responses, and suggest questions for the next conversation. The person reviewing the analysis needs to be able to trace a conclusion back to its source and distinguish what a customer said from what the model inferred. Otherwise, the summary can introduce another set of assumptions while making the research look more settled than it is.
The product decisions still need attention
While building an AI-powered market-research tool with LangChain, Redis, Python, Svelte, PostgreSQL, and S3, I had to make decisions that went beyond the technical stack. The product also needed a clear answer about what a user should receive at the end of a session and what information would be necessary to produce it.
Those decisions affect the interface, the research process, and the way the result is presented. A user needs to understand weak evidence, inspect useful sources, and recover from a misunderstood request. If an answer sounds more convincing than its sources justify, the product needs to make the uncertainty visible enough for the user to notice.
AI can help write code, explore edge cases, and compare implementation options. Decisions about the user experience also need information from people using the product. Where they get confused and which results they find useful give the team a basis for deciding what to change.
An agent completing a research task leaves plenty for the product team to work out. Someone has to decide how much freedom it gets, which information it can use, when it should stop, and what a person needs to review. A technically successful run may still give the user a result they can’t check or act on.
Put AI where the existing work is expensive
The order in which I’d consider AI uses depends on the work people are already struggling with. A new capability can be interesting without being a good place to spend the team’s time.
For a market-research product, gathering information and making sense of it belong high on the list. A user may spend hours moving between sources and organizing what they find before they can make a decision. There’s value in shortening that process when the result keeps enough of the evidence available for them to check it.
Repetitive work with predictable inputs is another useful place to look. Classifying requests, preparing routine summaries, and drafting messages from structured information give the reviewer a clearer idea of what a correct result should look like. That makes it easier to judge whether the tool is saving effort.
I’d put open-ended idea generation lower when the team already has more ideas than it can evaluate. Another list of possible features means someone has to compare more suggestions with customer needs, technical limits, and the company’s priorities. It can increase the amount of work waiting for a decision.
A support team handling many similar requests could reasonably put classification and drafting first. A company entering an unfamiliar market might get more from research. The ranking should change with the bottleneck, the frequency of the task, and the consequences of getting it wrong.
The cost of checking the output belongs in that decision. A task that takes seconds to generate can still be expensive if reviewing it requires substantial expertise or sends someone back through all the original material.
Borrow workflows carefully
Long prompts, agent loops, skills, and multi-agent systems are ways of organizing work. Their usefulness depends on the job you’re trying to do.
A coding workflow may need repository access, tests, version control, and retries. A workflow for reviewing customer interviews needs careful source tracking and a clear distinction between statements and interpretations. Copying the setup from one task to the other can add complexity without helping with the parts that matter.
Frameworks can also make a system harder to inspect when something goes wrong because they add layers between the developer and the model’s inputs and outputs. Anthropic’s engineering guidance describes this tradeoff and recommends adding complexity when the task calls for it.
There’s ongoing work after the setup. Prompts need changes, connections fail, and a model update can affect behavior you were relying on. A process used every day may justify that maintenance. For something you do once a month, completing it directly could take less effort.
Before adopting someone else’s workflow, I’d want to understand what problem it solves, how often that problem occurs, and how the person checks the result. Those assumptions help explain why it works for them and whether it fits another business.
Keep responsibility with the people making the decisions
A human co-founder brings experience, relationships, and information you may not have. They also share responsibility for what happens to the business. That gives disagreement a different context from asking a model to generate an opposing argument.
AI can help you investigate an unfamiliar area, prepare for customer conversations, draft material, write software, and test the reasoning behind a decision. Its usefulness grows when it has relevant evidence and a way to check its work. The founder still has to decide which evidence is strong enough and which risks the business can afford.
I’d use review requirements that reflect the consequences of the task. A routine draft and a recommendation affecting customers, spending, or hiring need different levels of attention. The team needs enough visibility into the work to understand what it’s approving and correct mistakes before they become commitments.
That leaves plenty of room for AI to help a small team. It can reduce the work involved in research and execution while giving the founders more information to consider. They still need to spend time with customers, choose what deserves their effort, and take responsibility for the business they build.