← All exercises

Chapter 6 · Demonstration · 6 minutes

Let them fight

From Actually Using AI

What this shows

Chapter 6’s claim is that asking one AI to critique another’s work is the most powerful single move you can make with AI. Not because AI is unreliable. Because the AI that helped you build something has spent hours getting invested in its conclusions, while a fresh AI seeing the same document for the first time has none of that history. The disagreements are where the real problems are.

You’re about to watch this on something with money on the line: Alex’s meal-kit business plan. He’s spent two weeks building it with an AI that has helped shape every section. He’s about to take it to investors. Before he does, he runs a single experiment — paste the whole plan into a fresh AI conversation and ask it to find every reason to pass on the deal. What comes back is not a polishing suggestion. It’s three things he’d have walked into the investor meeting not knowing.

The demonstration

Alex has been working on this for two weeks. The plan is for a subscription meal-kit service — fresh ingredients, weekly box, urban professional market. He’s worked through it section by section with his AI: who the customer is, what to charge, how to find them, how the money works. The document is now thirty-two pages. The financial projections balance. The competitor analysis names three direct rivals. The pricing has a defensible margin. The executive summary opens with “The UK meal-kit market is forecast to reach £1.2 billion by 2028…” — the market-size lead that every other plan in the pile will also have. He’s ready to send it to the first investor on his list.

Before he does, he wants a sanity check.

Take 1 — the collaborator AI’s review

Alex asks the AI he’s been working with for a final review.

Here’s what comes back:

AI: The plan is in good shape overall. We’ve covered the major components a sceptical investor will look for — market sizing, customer acquisition pathway, unit economics, financial projections, and the competitive landscape. The financial model is consistent: your 3.2% conversion assumption flows correctly through to the year-one revenue figure, and the gross margin calculation supports the pricing strategy we landed on.

A few things to consider before you send it:

1. The executive summary opens with the market size, which is standard for the format but doesn’t differentiate you. Consider opening with the specific gap in the market you’ve identified, then the size.

2. The “team” section is currently quite brief. Even a single sentence on why you’re the right person to build this would help an investor frame the rest of the document.

3. The customer-acquisition section is strong on channels but lighter on the timeline — when do you expect to hit the projected subscriber count? An investor will want to see the cadence, not just the destination.

The financial projections look defensible. Your competitor analysis covers the three main direct rivals. The risk section is honest about the main risks without overstating them.

Overall I think this is ready to send. The investors you’re approaching will be reading it on the strength of the numbers and the proposition, both of which hold up. A few openings to tighten in the next pass, but nothing that would stop you sending it today.

Alex reads it. The feedback is exactly what he expected. Some polish on the executive summary, a sentence about himself in the team section, a note about acquisition timeline. He decides to tighten the opening tomorrow morning, add the team line, work in the cadence point, and send by Friday. He’s about to close the tab.

Then he remembers something from the book. Before acting on AI’s first response, paste the whole document into a fresh AI conversation and ask it to find every reason to pass on the deal. He’s not sure what a fresh AI would catch that the one he’s worked with for two weeks wouldn’t. But ten minutes is cheap insurance before sending to investors.

Take 2 — the fresh AI as sceptical investor

Alex copies the entire business plan into a different AI tool. He hasn’t used this one on the project at all. The conversation starts blank.

Here’s what comes back:

AI: I’ve read the plan. Here are the three issues I’d raise in the meeting, in roughly the order I’d raise them.

1. The 3.2% conversion rate assumption.

Your financial model rests on the assumption that 3.2% of people who see your paid acquisition will convert to paying subscribers. This drives your customer acquisition cost, your year-one revenue forecast, and your runway calculation. I’d want to understand where this number comes from.

If it’s based on general e-commerce or online retail benchmarks, that’s likely too high for this category. Meal-kit subscription conversion from paid acquisition tends to be substantially lower than general e-commerce conversion — often in the range of 1.0% to 1.5%, sometimes lower depending on the channel mix. If you’ve applied a general benchmark to a meal-kit-specific business, your CAC could be underestimated by a factor of two or more, which would materially change the runway picture.

I would want to see the source of this number traced to a meal-kit-specific benchmark before relying on anything downstream of it.

2. The seasonality of the subscriber base.

Your monthly subscriber and revenue projections show relatively smooth growth across the year. Meal-kit services don’t typically grow smoothly. Subscriber retention tends to follow a seasonal pattern, with cancellations and pauses concentrated in summer months (people are on holiday, eat out more, the weekly box becomes inconvenient) and again over the Christmas period.

Industry retention figures often look reasonable on an annualised basis but mask a summer trough that can reduce active subscriber counts by 20-30% across June, July and August. If your model doesn’t account for this, your summer cash position will be worse than the spreadsheet shows.

I would want to see a stress test of the financial projections assuming a 25% pause/cancellation rate across the summer quarter, with reactivation patterns built in for September.

3. The competitor analysis.

Your competitor section names three direct rivals (FreshFare Direct, Plate & Pantry, Weekly Greens). It does not mention two companies that have launched in the same space in the last six months: UrbanPlates (London-focused, similar price point, professional-target positioning launched in November) and FreshPrep UK (the UK arm of an established US brand, launched February, has secured Series A funding).

Both target urban professionals at a similar price point to your plan. Both have been visible to investors in this space. If I’ve seen either of their pitches this quarter, your plan looks out of date the moment I open it. You don’t necessarily need to argue you’re better than them, but you do need to demonstrate you know they exist and have thought about how you compete with them.

One thing I should flag: I’m reading the plan as a standalone document, without access to any analysis or research that may have informed it. There may be answers to all three of these in conversations or workings that didn’t make it into the document. But the plan as written doesn’t address them, and an investor will read it as written.

Alex reads it twice. Then he stops planning to send the plan on Friday.

The 3.2% number wasn’t researched against meal-kit benchmarks — he’d taken it from a general e-commerce conversion guide his collaborator AI had referenced. The financial model assumes year-round retention because that’s how they built it; neither of them flagged the summer pattern. He’d never heard of UrbanPlates. He’d vaguely heard of FreshPrep but assumed it was a US-only brand.

None of these kill the plan. But walking into an investor meeting with a conversion rate he hadn’t checked, a financial model that ignores how the business actually behaves through the year, and a competitive landscape that’s six months out of date would have killed his credibility.

He spends the next week revising. He finds meal-kit-specific conversion data (the number he ends up with is 1.4%, not 3.2%) and reworks the financial model around it. He adds a summer stress test. He researches UrbanPlates and FreshPrep, confirms FreshPrep launched in the UK in February, and adds both to the competitor section with honest analysis of how he differentiates from each. The plan he sends on the following Wednesday is less exciting on paper than the version he was going to send on Friday — more cautious numbers, honest gaps flagged, competition acknowledged.

It’s also the plan that gets the second meeting.

The point

Both AIs were the same underlying technology. The first one had spent two weeks helping Alex build the plan. The second one had no history with the project at all. That single difference produced three findings the first AI did not catch.

The first AI was not wrong. Its review was thorough, the polish suggestions were sensible, and the plan as it stood was internally consistent. But internal consistency is not the same as external defensibility. The first AI had inherited every assumption Alex had made over two weeks, because it had helped him make them. The 3.2% conversion rate flowed correctly through the model because the AI had been there when the number was chosen. The year-round retention was modelled smoothly because that’s how they’d built it. The competitor section covered the right rivals because the AI had helped research them — six months ago.

The fresh AI didn’t inherit any of that. It read the plan as an investor would — cold, fast, looking for reasons to pass. That framing did the work.

This is the move. Not “use AI to check AI” as a generic technique. Specifically: the AI you’ve been working with for hours or days is invested in the work it helped build. A fresh conversation, given an explicit adversarial framing — act as a sceptical investor, find reasons to pass, be specific — sees what the collaborator can’t. The disagreements are where the real problems are.

Try it yourself

The next time you’re about to send something that’s been a long collaboration with AI — a business plan, a proposal, a strategy document, a major piece of writing — try this:

Get your collaborator AI’s view first. Take its feedback, make the polish moves it suggests. Then open a fresh AI conversation, paste in the whole document, and ask it to find every reason to reject it. Use specific framing — act as a sceptical [investor / reviewer / hiring manager / board member], find every reason to pass, be specific.

A few things worth knowing before you do this. Adversarial framing pulls hard in both directions: a fresh AI given a “find reasons to pass” prompt will usually surface real problems, but it will also generate plenty of pedantry, hallucinated concerns, and objections that don’t apply to your situation. Expect more noise than signal. The catches that matter will usually be the ones that point to something concrete — a specific number you used, a specific section that’s missing, a specific competitor or risk you didn’t think about. The pedantry — “the formatting could be more consistent”, “the executive summary should be tighter” — is noise. Ignore it. The other warning: fresh AI doesn’t always get the named details right. Competitor names it cites might be real, partly real, or made up. Treat the category of finding as the signal — the names need verification.

That said, ten minutes with a fresh AI told Alex what to fix before anyone else saw it. The signal is worth the noise.