Pressure-test a big decision with an AI red team

A structured adversarial review of a big decision, run by an AI you explicitly task with attacking it: the premortem, the strongest opposing case, and the assumptions that, if wrong, sink the whole thing.

What you'll have when you're done

A repeatable way to stress-test a major decision before you commit: you write a short memo, and an AI runs a premortem, argues the strongest case against you, and surfaces the three assumptions the decision secretly rests on. You get a checklist of risks to address, generated by something with no stake in the outcome. The decision stays yours; the blind spots get found first.

The people around you are not incentivized to tell you it's a bad idea

Here is the uncomfortable truth about big decisions. By the time you are seriously considering one, the people around you mostly want it to happen, the deal team wants the deal, your reports do not want to be the one who killed the boss's idea. So the hard questions go unasked, not out of malice, but because everyone has a stake. You end up pressure-testing your own thinking with a room full of people rooting for it.

An AI has no such stake. As one private-equity firm put it about running an AI devil's advocate on every investment, it has no career ambition and no ego. It will ask the question your team is too polite or too invested to ask. I have made the mistake this guards against: I walked into a decision I was excited about, surrounded by people who were excited with me, and we all quietly agreed not to poke the thing too hard because poking it felt like betraying the momentum. The objection that would have saved us money existed the whole time. Everyone half-knew it. Nobody wanted to be the one to say it out loud and slow the room down. The one catch is real and worth naming: AI models tend to agree with you by default (a tendency called sycophancy), so you have to explicitly assign it the adversary's role, or you just get a smarter cheerleader.

What you need first

• A decision worth the effort: a hire, an acquisition, a pricing change, a big spend. Not every call needs this; the consequential ones do.
• A one-page decision memo: what you are deciding, why, the key assumptions, the expected outcome.
• A capable model. Use a strong reasoning model (a frontier model) for this; depth matters here.

Step-by-step

Step 1Write the one-page memo

Get the decision out of your head and onto a page: what you are deciding, the reasoning, the assumptions it rests on, and what you expect to happen if you do it. The act of writing it is half the value, and it gives the AI something concrete to attack.

Step 2Explicitly assign the adversary role

This is the step that makes it work. Do not ask "what do you think?", you will get agreement. Assign the role:

You are my red team. Your job is to find why this decision fails, not to support it.
Be direct and specific, not hedged. From my memo:
1. Run a premortem: it's 12 months later and this decision failed badly. Tell the
   most likely story of how.
2. Argue the strongest possible case AGAINST doing this.
3. List the 3 assumptions this decision rests on that, if wrong, sink it.
Do not reassure me. Do not soften. Find the holes.

Here is the shape of what comes back, illustrative, on a decision to acquire a smaller competitor for $3M:

Premortem (12 months on, it failed): You closed the deal, but their lead engineer, who held most of the institutional knowledge, left inside 90 days, and the codebase turned out far less reusable than the demo suggested. You spent the year integrating instead of growing, your own roadmap slipped two quarters, and the acquired customers churned once the product stopped getting attention. The $3M bought a distraction, not a moat.

Strongest case against: You are buying revenue you could likely have won organically for less than the integration will cost, and you are spending your team's focus on a merge during the exact window you needed it on your core product.

Three load-bearing assumptions:

1. Their key people stay (your memo names no retention mechanism).
2. The tech integrates in roughly a quarter (based on a demo, not a code review).
3. Their customers are additive, not already overlapping with yours.

None of that tells you not to do the deal. It tells you the three things to nail down before you do, which is exactly what the people rooting for the deal were never going to raise.

Step 3Run it twice, with different framings

Run the red team, then run it again with a slightly different frame ("now argue it from the perspective of our biggest competitor" or "from the perspective of the customer"). Different framings surface different holes. You are not looking for one answer; you are stress-testing from multiple angles.

Step 4Treat the output as a checklist, not a verdict

The AI's job is to find risks, not to make the call. Take its premortem and its three assumptions and turn them into a checklist: for each, can you address it, de-risk it, or accept it? A confident-sounding objection might be wrong, your judgment decides which risks are real. The red team widens your view; it does not replace it.

Step 5Make the call, with eyes open

Now decide, having seen the case against. Often you proceed anyway, but with a mitigation for the top risk in place, or a tripwire ("if X happens, we pull out"). Sometimes the premortem shows you something that changes the decision. Either way, you decided with the holes visible instead of discovering them after the wire cleared.

Walk the acquisition example through to a real decision and you can see the value is rarely "stop." Assumption 1 (key people leave) becomes a retention package and a 90-day earnout written into the deal. Assumption 2 (the tech is messy) becomes "we are not closing until our CTO has done a real code review, not a demo." Assumption 3 (overlapping customers) becomes a quick analysis of their customer list against yours before you sign. You still do the deal, but you do a de-risked version of it, and you set a tripwire ("if the lead engineer signals they are leaving in the first 60 days, we pause integration and reassess"). That is the whole move: the red team does not make the call, it converts three vague worries into three concrete things you handle before they become the post-mortem.

How you'll know it's working

You catch the flawed assumption before it costs you, not after. Decisions get a little slower and a lot sounder. And you stop having the post-mortem realization, "everyone kind of knew this was risky and nobody said it", because the red team said it, out loud, before you committed.

When it breaks

• It just agrees with you. You did not assign the adversary role firmly enough. Restate it: "your job is to find why this fails, do not reassure me." Sycophancy is the default; override it explicitly.
• The objections feel generic. Your memo was thin. Give it the real specifics, the actual numbers and assumptions, and the critique sharpens.
• It raised a confident objection that's wrong. Expected and fine, it is a checklist, not an oracle. Your judgment filters real risks from noise.
• You treat its output as the decision. Do not. It finds holes; you decide. Outsourcing the call to the AI is the opposite of the point.
• It catastrophizes everything into a reason not to act. Over-corrected from sycophancy into pure pessimism, which is just as useless. Ask it to also rate each risk by likelihood and severity, so you can tell the fatal flaw from the survivable one and do not talk yourself out of every good move.
• You only run it on decisions you already doubt. Then it just confirms cold feet. Run it on the decisions you feel most confident about, that is where your blind spots are biggest and the red team earns its keep.

Make it yours. Match the rigor to the stakes. A reversible, cheap decision does not need a red team; a one-way door (an acquisition, a layoff, a pivot, a bet-the-quarter spend) does. For the biggest calls, run the red team, then bring its sharpest objections to a trusted human advisor, the AI surfaces the questions cheaply and tirelessly, and a person who knows your business helps you weigh which ones are real. Keep a short list of your recurring blind spots (you over-trust demos, you under-weight integration cost) and feed it in, so the red team targets your specific failure patterns, not generic ones.

Where this fits in your harness

This is a judgment-amplifier, not an automation. It pairs with the signals from your other workflows: your Monday market scan and competitor monitoring surface the moves worth pressure-testing, and your customer-call roadmap signal grounds product decisions in evidence before you red-team them. (See red team for the origin of the practice.)