I picture the managing partner at 6:40 p.m., laptop still open, staring at three different things at once. One partner is asking whether the firm is falling behind. Another is worried the team is getting lazy and over-relying on AI. A third is asking whether any of this will actually show up in margins, recruiting, or client value. This overwhelm is real.
The mistake is thinking the choice is between speed and caution.
The firms that pull ahead will build a flywheel. Not a tool stack or a governance policy. A flywheel, where every use of AI leaves the firm a little smarter than it was the day before.
That is why Simon Willison’s story matters more than it looks. In “Don’t classify. Hallucinate!” he points to Doug Turnbull’s move: don’t force the model to pick from an existing list of 1,856 tags. Let it generate possibilities first, then map those imagined tags back to the firm’s real vocabulary using embeddings. That is a practical reminder that the first answer is rarely the system. The system is the loop around the answer, generation, comparison, grounding, and reuse.
Ben Thompson gets to the same place from the economics side. He writes that the real opportunity is in building “a learning loop on top of models” where human capital and token capital compound together. That matters because your advantage has never been raw information alone. It is how your people recognize patterns, frame judgment, and improve the next engagement from the last one. If AI does more of the first-pass work, your moat shifts to whether you capture and compound what your people learn.
That is also why the labor stories should not be read as a simple automation forecast. Azeem Azhar notes that AI was cited as the top reason for nearly 40% of U.S. job cuts in May, and in a separate piece he reports that AI-native startups are 25% smaller at similar funding and growth, with denser expertise and fewer managers. Casey Newton’s interview with Town’s CEO on the self-organizing company points at the same thing. The shape of the firm is changing. But for mid-sized professional-services firms, the practical question is not whether headcount goes up or down. It is whether your people are being turned into better managers of intelligence, or replaced by a tidal wave of unowned output.
This is where I use my Autopilot Paradox framework. The more capable AI becomes, the easier it is for smart professionals to stop thinking too early. The staffing podcast puts it bluntly through Michael J.’s warning that “the most dangerous thing AI can steal isn’t your job. It’s your next thought.” The education piece, “YOUR CHILD IS NOT READY,” is really about the same fear in a different environment. If people outsource too much cognition before they have built judgment, they lose the muscle they most need to supervise the machine.
This is not abstract. A second-year associate who accepts an AI summary without testing it. A senior accountant who stops tracing edge cases because the cash flow looked clean. A broker who uses AI market language that sounds polished but misses the local signal that wins the deal. That is not efficiency. That is cognitive atrophy.
Noah Smith’s 23 low-regret policy recommendations and the Center for Humane Technology’s “messy middle” argument both tell me the same thing: society is still figuring out the rules while the machines and incentives keep moving. You do not get to wait for perfect clarity.
So I come back to a framework I trust, the 10-20-70 Rule. In my experience, about 10% of this is the model or tool, 20% is workflow design, and 70% is people, habits, judgment, incentives, and management. The best move for a mid-market firm is to stop measuring AI by usage and start measuring it by learning velocity. Are your teams documenting what worked, what failed, what had to be escalated, what prompt pattern improved quality, what client objection keeps repeating, what exception really mattered? Are you creating what Thompson calls compounding learning, or are you just generating more text? And most importantly are you changing your AI workflows to take in this new information.
This is also why I would connect this compounding learning to my AI Whisperers framework. Every firm needs a small network of key employees who do more than cheerlead tools. I want them capturing good prompts, review patterns, red flags, and judgment calls from real work. They are the human bearings inside the flywheel. Without them, AI stays personal. With them, it becomes institutional.
The concrete decision for this quarter is simple. Pick one high-frequency, client-facing workflow, proposal drafting, diligence summaries, tax memo first drafts, candidate brief prep, whatever matters in your firm, and require a closed learning loop around it. Every output gets human review. Every correction gets logged. Every recurring miss gets turned into a better instruction, checklist, or escalation rule. Do it for 90 days. Then review whether the firm got smarter, not just faster.
The rest of this brief examines how the conversation should be opened — what specifically to say to clients who are silent, how to structure the disclosure so it lands as competence rather than panic, and three ways the Friction Audit reveals exactly which client conversations to have first…