If you are running a small or mid-sized firm right now, I know what this feels like. One partner forwards you a note about Kimi K3. Another sends a screenshot about Claude Fable 5 pricing. Your IT lead is worried about agents with shared credentials. Your associates are quietly wondering whether junior work is disappearing. And somewhere in the middle of all that, you are supposed to make a calm, durable decision.
The temptation is to keep shopping for a better engine.
But that is not what matters, what matters now is the gearbox.
The biggest signal across recent news is that AI value is moving away from the raw model and toward the learning loop around the model. Ben Thompson put it cleanly when he wrote that “the real opportunity is not choosing the best model but building a learning loop on top of models in which human capital and token capital compound together.” That matters a lot for a law firm, CPA firm, brokerage, search firm, or deal team. You can offload a task. You cannot offload your firm’s learning.
That is why the latest model shock matters less than it appears. Yes, prices are moving. Yes, capabilities are jumping. Yes, product plans are getting scrambled. But if your team’s prompts, review habits, escalation rules, precedent libraries, client-context patterns, and correction loops all live in scattered chats and private experiments, then every new model release resets you back to zero. You are renting intelligence by the drink instead of building a firm that gets smarter every month.
This is also why the agent stories should make you alert, not frozen. VentureBeat reported that 54% of enterprises have already had an AI agent incident or near-miss, and that only about a third give every agent its own scoped identity. That is not just a security story. It is an operating-model story. If an agent acts without clean identity, bounded permissions, and a review path, you do not have a learning loop. You have a loose machine in the hallway.
The same pattern shows up in context. VentureBeat’s separate piece says the issue is “a trust problem, not a retrieval problem,” and that most enterprises are building context infrastructure faster than they can trust it. I agree with that framing. In my Four Cultural Dimensions, trust and data fluency sit together for a reason. If your AI can retrieve five thousand pages of policy, contracts, comps, or workpapers but your people still cannot tell when the answer is confidently wrong, then your system is not getting smarter. It is getting louder.
Simon Willison’s note on AI mania sharpens the people side of this. He surfaced the story of an executive who had “never even used ChatGPT” after producing a strategy centered on AI for a $2B+ organization. I do not read that as a funny anecdote. I read it as a warning. Firms are making directional calls about AI without a lived feedback loop between leadership assumptions and frontline reality. In my experience, that is how you get theater instead of adoption.
Then layer in the labor signal. Azeem Azhar cited data from more than 21,000 US firms showing that heavy AI adopters grew employment by about 10% over two years after adopting AI, while entry-level roles grew 12%. But the same piece says some firms fired too quickly, mistaking task automation for human obsolescence. That is the heart of the issue for professional-services leaders. Junior work is changing. Supervision work is growing. Review, verification, and quality control become part of the value stack. If you do not intentionally redesign the apprenticeship loop, AI will hollow out your bench. If you do redesign it, AI can strengthen it.
This is where I use two of my frameworks.
First, the 10-20-70 Rule (10% tools, 20% process, 70% people). Only a small part of the win will come from the tool itself. More comes from workflow design. Most comes from people, habits, governance, and reinforcement. If your partners are still debating vendors while ignoring review discipline, knowledge capture, and training, they are overinvesting in the tools and starving the people.
Second, the AI Whisperers idea. The firms that pull ahead will not be the ones with the flashiest subscriptions. They will be the ones that teach more of their people how to brief AI clearly, challenge outputs, spot edge cases, and feed corrections back into the system. Better human sophistication in AI collaboration beats endless tool chasing.
I also think the gig-work warning matters here. The AI Now Institute piece points to “gigification” spreading into skilled work. For a mid-market professional-services firm, that is not just a labor-policy debate somewhere else. It is a strategic fork in the road. You can use AI to deskill, atomize, and thin out your human core. Or you can use AI to build a Centaur Firm, where humans and AI together outperform either alone and where judgment gets more valuable, not less.
The concrete decision this quarter is simple. Pick one workflow where junior staff produce high-volume first drafts and senior staff already review them anyway, think diligence summaries, lease abstracts, tax memo drafts, candidate research briefs, or client update memos. Turn that into an AI Impact Project with a visible human review layer, structured feedback capture, and a shared repository of what good looks like. Do not ask which model seems smartest this week. Ask whether the work leaves your firm smarter after every cycle.
That is the gearbox.
The firms that build it will keep moving, even when the engine changes again next month.
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…