The False Safety of Agreement
There's a pattern showing up in ops teams right now: a question goes into three AI models, all three return the same answer, and the team ships it. The reasoning is intuitive — if independent systems converge, the answer must be solid.
It isn't. Models trained on overlapping data with similar RLHF tuning will confidently produce the same wrong answer. Consensus is a correlation artifact, not a verification signal. Nate Jones published a practical guide to this exact failure mode this week, and it's the most operationally useful framing we've seen for non-tech operators deploying AI across sales, finance, and ops without in-house ML expertise.
What Friction Maxxing Actually Means
Jones's framework introduces deliberate resistance steps between AI output and execution. The core techniques:
Counter-prompting: After getting your answer, explicitly instruct the model to argue against it. Ask Claude to steelman the opposite position. Ask Grok to identify what conditions would make the original answer wrong. Models are sycophantic by default — you have to design around that.
Source-forcing: Require the model to cite specific, verifiable sources for any factual claim. Then check two of them. This single step eliminates most hallucination risk in business-decision contexts because it forces the model to surface its actual evidence base rather than synthesizing from training data.
Domain expert spot-checks: For any AI output that will affect pricing, hiring, legal exposure, or customer commitments, route it through one human with direct domain expertise before it moves. This isn't a bottleneck — it's the last gate. Build it into your workflow as a required step, not an optional review.
The decision tree Jones structures this around is worth mapping to your own team's AI touchpoints: which outputs go straight to execution, which get counter-prompted, which require source-forcing, and which need expert review before they ship.
The Actionable Step This Week
Audit the three AI-assisted workflows your team runs most frequently. For each one, answer: what's the blast radius if the output is confidently wrong? For anything touching customer data, pricing, or external communications, add at minimum one friction step before execution.
This isn't about slowing down. It's about not automating errors at scale. The operators who build verification into their AI workflows now will have a structural advantage over those who discover the failure mode after it surfaces in a customer call or a bad quarterly number.
If you're running AI across GTM or ops without in-house ML review capacity, friction maxxing is your quality control layer. Build it deliberately or it won't exist.