Red Team Thinking: How Adversarial AI Improves Strategy

In 1963, Pope John XXIII formally established the role of the Promotor Fidei — the "Devil's Advocate" — in the Catholic Church's canonization process. The role's explicit purpose was to argue against sainthood, to find every reason why the candidate was unworthy, so that the final decision would be robust against the strongest possible objections.

The concept wasn't new. Military organizations had practiced red teaming — assigning a team to argue the enemy's position — for centuries. What was notable was the recognition that important decisions require structural opposition. Not because opposition is always right, but because the process of surviving genuine challenge produces better outcomes than the process of building consensus.

Today, the most consequential application of red team thinking is moving from military exercises and cybersecurity audits into strategic business decision making. And multi-agent AI is making it practical at a scale that was previously impossible.

Why Every Strategy Needs a Devil's Advocate

The psychological research is unambiguous. Groups that include a designated dissenter make better decisions than groups that seek consensus, even when the dissenter's specific arguments are wrong. Charlan Nemeth's research at UC Berkeley found that authentic dissent — not role-played devil's advocacy, but genuine disagreement — stimulates divergent thinking that improves the quality of the majority position.

The mechanism is straightforward: when you know your strategy will face serious challenge, you build a better strategy. You think through the failure modes. You identify the assumptions you're relying on. You develop contingency plans. You do the work that consensus-seeking organizations skip.

The problem isn't that leaders don't understand this. It's that creating genuine structural opposition within a real organization is incredibly difficult. The person assigned to play devil's advocate knows it's a role. Their career isn't served by genuinely undermining the CEO's preferred strategy. The challenge is performative, and everyone in the room knows it.

The Problem with Groupthink in AI

The AI era has, counterintuitively, made this problem worse. When organizations use AI as a strategic thinking tool, they typically use a single model in a single conversation. And single-model AI is architecturally incapable of genuine red teaming for the same reason a single person can't play chess against themselves: you can't simultaneously hold a position and genuinely try to destroy it.

When you ask ChatGPT to "red team this strategy," you get a polite list of risks that reads more like a CYA section in a consulting deck than a genuine adversarial attack. The model identifies surface-level concerns without committing to any of them. It hedges every criticism with "however, the overall strategy is sound." This isn't red teaming. It's the AI equivalent of a subordinate who raises just enough concerns to seem thoughtful without actually challenging the boss.

The structural fix requires separate experts with separate objectives — one genuinely motivated to build the strongest case for the strategy, another genuinely motivated to destroy it.

Red Teaming in Cybersecurity vs. Strategic Decision Making

The cybersecurity industry offers the best existing model for what effective red teaming looks like. In penetration testing, the red team's job is to break in — not to provide a balanced assessment of the security posture, but to find the single vulnerability that makes everything else irrelevant.

The principles transfer directly to strategic red teaming:

Asymmetric objectives. The red team's only goal is to find fatal flaws. They don't need to present a balanced view. They don't need to acknowledge the strategy's strengths. They need to find the thing that kills it.

Realistic adversarial behavior. In cybersecurity, red teams simulate actual attackers — not theoretical ones. In strategic red teaming, the adversarial expert should model what real competitors, regulators, or market forces would actually do, not what the strategist hopes they'll do.

No mercy rule. An effective red team doesn't pull punches because the finding is politically uncomfortable. If the strategy depends on an assumption that's demonstrably false, that's the finding — regardless of how much money has already been spent on the current course.

Structured debrief. The value isn't in the red team's attack itself — it's in the organization's response. A good red team exercise surfaces the specific vulnerabilities that need to be addressed, and the strategy is refined accordingly.

How Multi-Expert AI Creates Structural Red Teams

Multi-agent AI systems can create genuine red team dynamics because different experts have genuinely different objective functions. In SynthBoard's AI Boardroom, the adversarial experts aren't performing a role — they think for themselves and are configured to find flaws.

The Devil's Advocate doesn't receive instructions to "consider some risks." It receives instructions to construct the most compelling argument for why the proposed strategy will fail. It's measured on the quality of its objections, not on how balanced its assessment is. This is a fundamentally different cognitive task than what a single model does when asked to "red team" a strategy.

A multi-agent red team session might include:

• The Advocate — presents the strongest version of the strategy, steelmanning rather than strawmanning
• The Adversary — constructs the most compelling case for failure, with specific mechanisms and evidence
• The Competitor — models how intelligent competitors would exploit the strategy's weaknesses
• The Historian — identifies analogous strategies that have been tried before and analyzes why they succeeded or failed
• The Synthesizer — reconciles the debate into an honest assessment of the strategy's robustness

The tension between The Advocate and The Adversary is the core engine. Neither expert defers to the other. Neither hedges. The Advocate builds the strongest possible bull case. The Adversary builds the strongest possible bear case. The resulting synthesis captures both — and more importantly, it identifies the specific assumptions on which the strategy's success depends.

Five Red Team Questions Every Strategy Should Survive

Before committing resources to any major strategic initiative, subject it to these five adversarial tests:

1. "What has to be true for this to work?" List every assumption the strategy depends on. Market size, competitive response, execution timeline, customer behavior, regulatory environment. Then ask: which of these assumptions am I most uncertain about, and what happens if that assumption is wrong?

2. "How would an intelligent competitor respond?" Model the competitive response — not the one you hope for, but the one a smart, well-resourced competitor would choose. If your strategy depends on competitors being slow or stupid, it's not a strategy.

3. "What does the realistic failure mode look like?" Not the catastrophic black swan failure, but the mundane, predictable failure. The one where everything takes 50% longer and costs 40% more than projected. Does the strategy survive that scenario?

4. "What would make us abandon this in 12 months?" Pre-commit to the kill criteria. If you can't articulate the conditions under which you'd abandon the strategy, you'll never abandon it — you'll just keep investing past the point of rational commitment. This is the sunk cost fallacy in organizational form.

5. "Who disagrees, and why?" If nobody disagrees, you haven't asked the right people — or the culture has eliminated dissent. Identify the strongest objection to your strategy and engage with it seriously. If you can't articulate the strongest objection, you don't understand your strategy well enough to execute it.

Building a Red Team Culture with AI Assistance

The ultimate goal isn't to run occasional red team exercises. It's to build an organizational culture where adversarial analysis is a standard part of every significant decision.

AI makes this practical in several ways:

Lowering the social cost of dissent. When the red team analysis comes from an AI expert rather than a junior employee, nobody's career is at risk. The analysis can be genuinely adversarial without the political consequences that suppress honest feedback in most organizations.

Making it routine. When red teaming requires assembling a team, scheduling a war game, and dedicating multiple days, it only happens for the largest decisions. When it requires describing a strategy and running an AI boardroom session, it can happen for every meaningful decision.

Preserving institutional memory. AI red team sessions produce structured records — the strategy, the arguments for and against, the specific vulnerabilities identified, and the refinements made. Over time, this creates an institutional knowledge base of adversarial analysis that compounds in value.

Calibrating confidence. After adversarial analysis, decision makers don't just know what they've decided — they know how robust that decision is. A strategy that survived aggressive red teaming warrants more confidence and faster execution than one that hasn't been tested.

The organizations that outperform over decades aren't the ones that avoid mistakes. They're the ones that catch mistakes before they're committed to. Red team thinking, powered by multi-agent AI, makes that capability available to every organization that's willing to hear what it doesn't want to hear.

Run your first red team session with SynthBoard's AI Boardroom — assemble adversarial experts against your most important strategy and see what survives.