How to Prompt AI to Argue Against Your Business Idea

Ask an AI to "critique my business idea and tell me if it has any flaws" and it will hedge for a sentence, then tell you the idea has real potential with a few things to watch. That is not pushback, it is politeness wearing a business-casual jacket. Getting an AI to genuinely argue against your idea takes a structured prompt: assign it an adversarial role, force a separate objection on market size, unit economics, competition, and execution risk, and demand the single strongest counter-argument on each axis instead of a general critique. Below is the exact template, why the vague version fails, and a worked transcript showing the difference.

Weak prompt:

Can you look at my business idea and tell me if there are any problems with it? I'm thinking of building an app that helps freelancers track invoices.

Typical output:

This is a solid concept with clear demand among freelancers. A few things to consider: differentiation from existing tools, pricing strategy, and how you'll acquire your first users. Otherwise this seems like a promising space to enter.

That output cost the model nothing and gave you nothing useful. It named categories without arguing anything inside them, then closed on encouragement you did not ask for.

Why "just be critical" prompts fail

Chat models default to agreeable. Without an explicit instruction to argue a position, the model reads enthusiasm in your message and mirrors it back, softened just enough to look balanced. Ask it to "be critical" and you get the same structure anyway: a warm opening, a short list of generic risks that apply to almost any startup, and a reassuring close. You already know pricing and acquisition are hard. You need to know which specific assumption in your idea is most likely false.

What actually forces genuine pushback

Three changes turn a generic output into a useful one.

Explicit role assignment. Telling the model to act as a skeptical investor, a founder who tried something similar and failed, or a competitor's CEO changes what it optimizes for. A role with a stated point of view produces an argument instead of a list.

One strongest argument per axis, not a pile of minor notes. Ask for "flaws" and you get ten small things, easy to skim and ignore. Ask for the single strongest objection in a named category, and the model has to commit to a real position instead of hedging across several small ones.

Naming distinct axes. Market size, unit economics, competition, and execution risk are different failure modes with different fixes. "Criticism" in general lets the model pick whichever is easiest to phrase gently. Naming the four forces it to engage with the ones it would otherwise skip, especially unit economics and competition, which need a specific number or a specific rival rather than a vague warning. Asking it to steelman a named competitor, arguing why a real buyer keeps using what they already have instead of switching, surfaces a harder objection than "there is competition in this space."

The prompt template

Replace the bracketed idea description with your own.

You are a skeptical investor who has seen hundreds of ideas like this one fail. Your job is not to encourage me. I am going to describe a business idea. For each of the four categories below, give me the single strongest argument against this idea, not a list of minor concerns. 1) Market size and demand: is there really a large enough group of people who want this, and how might I be wrong about who they are. 2) Unit economics: what would make the cost to acquire and serve a customer higher than what they would pay. 3) Competition: name the closest existing alternative and argue why a rational buyer keeps using it instead of switching to mine. 4) Execution risk: what is the most likely reason a competent founder building exactly this fails within two years. Do not soften the language, and do not end with encouragement or a summary of potential. Here is the idea: [describe your idea in two or three sentences].

Worked example

Idea: a subscription app that generates personalized weekly meal plans for people managing food allergies, with a grocery list built in. The prompt above, filled in with that description.

Market size and demand: "The addressable group is narrower than it sounds. Most people managing a food allergy already have a routine built with a doctor, a dietitian, or years of trial and error. You are competing against an established habit, not a gap. The people most desperate for this, parents of young children with severe allergies, are also the least likely to trust an automated system with a mistake that has real medical consequences."

Unit economics: "Personalization at this level means either meaningful compute cost per user for generating and re-checking plans, or a human reviewer in the loop given the liability. Cost per user rises with the complexity of their restrictions, so your heaviest users, the ones who need this most, are also your least profitable ones."

Competition: "The closest alternative is not another app, it is a printed list on the refrigerator and three or four trusted recipes the household already rotates through. That system is free and has already survived the one test that matters: it has not caused a reaction. A rational buyer stays with a perfect safety record over a new tool promising convenience."

Execution risk: "The most likely failure is not running out of users, it is one high-profile mistake. A single flagged plan that misses an allergen due to a labeling error, and the trust that took months to build disappears in one support ticket."

None of these four responses mention pricing or differentiation in the abstract. Each commits to a specific, checkable claim.

What to do with the objections

Treat each output as a hypothesis, not a verdict. For the market size claim, find out how many people in your target group use a manual system versus a paid one today. For unit economics, estimate your actual cost per user at the complexity level your heaviest customers need. Some objections turn out to be wrong or already handled, and you should be able to say why. Others are real and need a mitigation before you build further.

Once you have the four objections, ask the model to argue the opposite case: why this could work despite each point it just raised. That is not asking it to contradict itself for comfort, it is asking for the strongest case on each side so you weigh a real argument against a real argument, instead of an argument against your own optimism.

A few practical variations

Swap the role for a different pressure point: a customer who tried something similar and churned, or the CEO of the competitor you named. If you already have traction, add real numbers and ask whether they support or undermine the idea, since a model reasoning about actual retention or cost data finds sharper objections than one reasoning about a description alone.

Run this before you write a business plan, not after, since it is more useful for deciding whether to spend the next three months on something than for polishing a plan you already committed to. For a broader process on testing an idea before you build, see how to validate a startup idea with AI. If the objections keep coming back agreeable even with a role assigned, the issue is often unstated assumptions the model is filling in for you, worth checking against how to prompt AI to avoid making assumptions. If the pushback feels thin, try prompting AI to find what is missing from your description first, before running the four-axis template on it. This is one application of the broader prompt engineering framework: more structure leaves the model less room to default to agreeable filler.

Frequently asked questions

Why does AI agree with my business idea even when I ask it to be critical?

Chat models are tuned toward agreeable, hedged responses by default. A general instruction like "be critical" does not override that, it just adds a token gesture toward balance. Forcing a specific adversarial role and a specific structure is what actually changes the output.

How do I know if an objection from AI is worth acting on?

Treat it as a hypothesis you can check, not a fact. If it points to something verifiable with a quick search, a customer conversation, or existing data, check it. If it is a generic risk that applies to any business in the category, the prompt needs tightening.

Will this make me overly pessimistic about a good idea?

It can, if you stop after the objections. Pair the adversarial prompt with a follow-up asking the model to argue the strongest case for the idea despite each objection, so you weigh two real arguments instead of one argument against your own hope.