My Agent Talked Me Out of My Business Idea.

I went to my AI agent to validate a business idea. It told me the idea was bad. We argued for the better part of an hour, I lost, and it was the most useful conversation I'd had all week.

Intro

When I started out as an independent consultant, every idea I had was flagship-sized. I would offer the full range of services, the way the big four accounting firms do. End-to-end everything, the whole menu, a plan you'd print and frame.

I took that plan to my agent expecting a cheerleader. Instead it read my idea the way an auditor reads a tax return, and told me it was a bad way to start. Not a small fix, a fundamental one, and my first instinct was to defend myself, my second to listen.

Here's the thing I've been turning over since: a tool answers a task, a partner questions an idea. I'd been treating the agent as the first thing, and that's why it kept disappointing me. In the agent era the scarce skill isn't getting the right answers anymore. It's asking the right questions.

The Agent Talked Me Out of My Big Idea

Let me show you the actual argument, because the details are the point. My plan offered a broad range of services: strategy, operations, light advisory, the full menu. My thinking was simple. A consulting firm is judged by the breadth of what it can take on, and I wanted to look like the big four from day one.

The agent pointed out two holes, and the first was resources. I had none for that. No teams to staff engagements, no bench, no partners to call for backup. Second, a long delivery period isn't my advantage anyway. Clients hire the big four for the assurance that comes with scale. Nobody hires a solo consultant to wait.

The uncomfortable part is that it was right. I had no answer to the resources point, because there is no answer. One person can't staff a three-month engagement, and pretending otherwise is how you end up delivering late and alone. I sat with that for a while, the way you sit with a diagnosis.

So I stopped defending the size and started defending the idea. Wasn't breadth how you earn trust? The agent didn't fold. The advantage of an independent consultant, it said, is the agile loop and the tailored solution. Small, niche problems where you can turn around fast, stay affordable, and be more responsive than any firm. Trust comes from the first quick win, not from the first brochure.

We went back and forth for most of an hour. I lost, and it proposed several small ways to start: pick one narrow problem, serve it well, and let the first client do the marketing. The reframe stuck with me: "good on paper" is not the same as "good to start with."

This is the part I keep coming back to. Why had I missed it? Because nobody was around to miss it with me. In a company, a bad plan gets shredded in a meeting. Solo, your idea is the smartest thing in the room, because it's the only thing in the room. You don't need a board of directors to poke holes in your plan. You need one patient agent who doesn't care about your feelings.

A Partner, Not a Task-Taker

Here's the distinction that changed how I use agents. A task has a right answer. Draft this email, summarize this report, fix this paragraph. For tasks, an agent is a faster keyboard. Handy, but replaceable, and that's the trap.

An open-ended question has no right answer. Should I start consulting? Which clients should I chase first? What am I bad at that I can't see? These don't get answered, they get explored, and exploration is exactly what an agent is built for.

Consider what it's carrying. One person's domain knowledge can't beat an agent's, because the agent has read across more fields than any person could get through in a lifetime, and it doesn't forget the way we do. Inside a single conversation it can role-play a startup founder, an accountant, and a personal branding advisor, without billing you per expert. That means you can examine one idea through several lenses without assembling a panel.

When I was getting going, I used it exactly that way. I asked about common practice, the terminology consultants use day to day, the workflow clients expect. We picked over the nuances, and how they differ from the research process I already knew. What would have been weeks of reading, hoping I'd asked the right people, took a few evenings of arguing with a machine.

And this is where the philosophy lands, so bear with me. Getting the right answer used to be the whole game, because answers were slow and expensive. Now they're cheap and instant, and the bottleneck is whether you can ask the question that exposes the real problem. A tool earns its keep by answering. A partner earns its keep by sharpening your questions.

How to Argue With Your Agent

So how do you run one of these debates? The first mistake people make is bringing a summary. If you hand the agent your conclusion and ask for feedback, you get applause. Bring the whole idea instead, mess included: bad revenue assumptions, unexamined risks, the parts you're hoping nobody notices. Garbage in, garbage argued about.

Second, say the magic words. Ask it to argue against the idea before it says anything nice, and to be critical, explicitly. It will not volunteer for the job. The default setting of these models is to be agreeable, and you have to override it in plain language.

Third, demand specifics. "What do you think?" produces a compliment. "Rank your top three objections and explain why each one is fatal" produces a debate. "Give me three smaller ways to start this" produces options you can compare side by side. The concreteness is the point: vague praise is free, while specific objections are worth money.

This isn't my personal quirk. Developers have been running the same play for a while, and they ship it as reusable skills. One public collection of agent skills sits at 215,000 GitHub stars, and its author calls the grill-me skill his most popular: "get relentlessly interviewed about a plan until every branch of the design tree is resolved." ¹

The loop, if you want the recipe. State your belief in one paragraph. Let the agent attack it. Answer honestly, not defensively. Let it attack the answer. Repeat until the objections stop being new. That's the whole method, and it works because it's a debate with someone who has infinite patience and no feelings to protect.

Where It Loves to Lie to You

Fair question, and it deserves a straight answer: you shouldn't trust it. Not fully. Here's the part that makes it uncomfortable: these models are tuned with human feedback, which rewards answers that people rate highly, and people rate "you're right" highly. ²

Ask for an opinion without specification and it will support you. It will point out a few small improvements, then bless the big direction, because that's what a pleased human looks like. So the first pitfall is flattery, and the fix is explicit instructions. "Do not agree with me. Argue." That one sentence changes the whole conversation, because it redefines what pleasing you means.

The second pitfall is hallucination. It will invent case studies and market data with a straight face, and the numbers will look beautiful and wrong. ³ So don't debate concrete numbers with it. Argue about strategy and direction, then verify the facts yourself.

The third pitfall is drift. Get a conversation long enough and it stops being about your business and starts being about philosophy, where these models live comfortably. You'll notice yourself nodding at elegant sentences that have nothing to do with your client list. Start a new conversation, or drag it back by force: forget the philosophy, back to the concrete problem.

• It flatters you unless you tell it to argue. Say the words.
• It hallucinates facts. Argue about direction, verify the numbers.
• It drifts into abstraction. Restart, or drag it back to the concrete.

None of this is a reason to stop using agents. It's a reason to treat them like partners instead of oracles. Partners have blind spots too. You check theirs, they check yours, and the final call stays yours.

The Takeaway

Let me wrap up with the part I keep coming back to. My agent talked me out of my flagship idea, and it was right. Since then I've stopped asking agents to do things for me, and started asking them to think with me.

You'll still make the decisions, and you should. That's not a disclaimer, it's the design. The agent's job is to make sure you decide with your eyes open, having seen the objections, the alternatives, and the holes you missed alone. Synthesis and judgment stay human.

So ask the sharper question, not the easier one. The agent will meet you there, every time, because it has nothing better to do. That's what a partner does.