← All episodes

November 2, 2026 · 11 min

The Governance Gap

0:00 / 11:14

In this episode

Transcript

This is The Judgment Layer, from Advisers Give Back — a nonprofit working to increase access to pro bono financial planning, pairing households who can't afford a financial planner with CFP professionals who volunteer their time. This is your weekly briefing on AI: what happened, what it means, and what it changes if you're the one running the firm. Written and voiced with AI; directed by Matt Iverson-Comelo, executive director of Advisers Give Back. Nothing here is investment, legal, or compliance advice.

This week, two things happened that make the AI question harder to defer — and one of them opens your books for the first time.

Anthropic filed its S-1 this week, and the number that caught my attention wasn't the revenue figure. That story was already in play, and we've covered the trajectory. What stopped me was a detail that Timothy B. Lee — the technology journalist who writes the newsletter Understanding AI — pulled out after reading all four hundred and twenty-one pages. The gross margin. Anthropic is now running gross margins in the mid-sixties, percentage-wise. For a company that, eighteen months ago, was losing money on every API call, that is a structural turn. Not a quarterly blip. Lee's read is that Anthropic has solved the unit economics problem faster than almost anyone expected, and the S-1 makes that legible to the public for the first time.

Here is what that means at a level deeper than the headline. When a frontier AI lab has healthy gross margins, it stops being a company that needs to choose between serving enterprise customers and funding research. It becomes a company that can do both simultaneously. And that changes what it builds. Anthropic has been the most deliberate of the major labs about enterprise trust — its Acceptable Use Policy, its Constitutional AI approach, the way it talks about reliability over raw capability. That posture has sometimes read as cautious, maybe even commercially timid. Profitable gross margins remove the pressure to abandon it. Azeem Azhar — the founder of Exponential View, a research and analysis publication focused on technology's long-term effects — put it plainly in his breakdown of the S-1 this week. This filing answers the central question that skeptics have been asking since the beginning of the AI investment boom, which is whether there is a real business under all the hype. The answer, in Anthropic's case, is yes. And the business is more durable than most people modeled.

Now here is the non-obvious part, and this is what I'd want you to sit with. The Hard Fork podcast from The New York Times — hosted by Kevin Roose and Casey Newton, the paper's technology columnists — noted this week that Anthropic is the first pure-play large language model company to go public and open its books. That means it is the first time anyone outside a handful of investors has seen what the economics of this technology actually look like at scale. And the answer is. They look like software. High gross margins, heavy upfront investment in compute, operating leverage as the customer base scales. That is the business model the skeptics said couldn't exist. They argued that AI would always be compute-bound, that margins would stay thin, that it was more like a utility than a platform. The S-1 refutes that. And if Anthropic's economics look like software, then every assumption your firm has made about the long-term pricing of AI access needs to be updated. About whether frontier capability will stay expensive. About whether smaller firms will eventually get access to the same tools the largest institutions use. Software economics mean the price of capability keeps falling while the capability keeps rising. That dynamic does not stop.

For someone running an advisory firm, here's where this gets operational. One of the subtler things in the S-1, as Lee analyzed it, is how Anthropic describes its enterprise customer base. Large institutions, with long contract cycles, who are embedding Claude deeply into workflows rather than using it as a supplemental tool. That description should sound familiar, because it describes the trajectory the most AI-forward advisory firms are already on. The ones who aren't there yet tend to have the same gap. They have experimented with AI in isolated pockets, but they haven't written down what their actual policy is for how AI-generated work gets reviewed, verified, and signed off on before it reaches a client.

This is the moment to write that policy down. Not because Anthropic went public — that's not the cause — but because the S-1 marks the point where the ambient uncertainty about whether this technology is real and durable has been replaced by public, audited evidence. The firms that treated AI as an experiment pending further proof now have their proof. The question that should be in your next leadership meeting is this. What would it look like to treat AI assistance as a permanent part of how work gets done here? And what oversight structures do we need if that's true? That's different from the question most firms have been asking, which is how do we run a pilot. A policy question is. Who checks what, how often, and who is accountable when an AI-assisted output is wrong. That's not a technology question. That's a governance question. And it's overdue.

One practical version of this. Pick one workflow where your advisers are already using AI — meeting preparation, draft communications, research summaries, whatever it is — and have someone document exactly how that works today. Not how it's supposed to work. How it actually works. Who prompts the model, what they do with the output, whether anyone else reviews it before it lands in front of a client. That documentation exercise will surface the governance gaps faster than any framework document will, and it will give you something concrete to build a policy around. Schwab, per a report in InvestmentNews this week, is now offering credits to help registered investment advisers cover enterprise AI seat costs. Which tells you something about how seriously the custodial layer of this industry is treating AI adoption as a competitive factor. That's useful context. But the policy question is yours to answer, not theirs.

The takeaway from the Anthropic S-1 for a firm leader is this. The economics of frontier AI now look like software, which means the cost of access will keep falling. And that makes the governance question more urgent, not less, because more of your people will be using these tools more of the time.

The other thing I'd want you to know about this week is something that sounds like a gaming story and isn't. Jacob Sansbury, the founder of a gaming company called Spawn, published a detailed account this week of how his company now operates. Simon Willison — the developer and writer who tracks AI at the technical frontier more carefully than almost anyone — flagged it. Spawn is a massively multiplayer game where players can prompt changes to the game world in real time. A castle here, a dragon there. What Sansbury describes, though, is less about the game and more about how it's built. A swarm of autonomous AI agents, running on a model tier Sansbury calls Fable-class, is now writing and deploying live code to production without a human in the loop. The company is spending, by Sansbury's account, more than three million dollars a month on AI tokens. Thirty times what it spends on human payroll.

I'm not telling you this because you're building a game. I'm telling you because the ratio is the story. Thirty times payroll on AI. That is not a company that uses AI to help humans work faster. That is a company where AI is the primary productive force and humans are the supervisory layer. Most firms are nowhere near this, and most firms shouldn't be. The risk tolerance and the output domain are completely different. But the reason to pay attention is that Sansbury's architecture is going to become cheaper and more accessible. And the pattern of a small human team supervising a much larger AI workforce is going to appear in more conventional professional services contexts over the next few years. The question it raises for an advisory firm isn't whether to do this. It's whether you have thought about what the supervisory role actually requires — what judgment, what domain knowledge, what accountability — so that when this pattern arrives in your corner of the industry, you're ready to staff for it rather than scrambling to define it.

A quick note on something that appeared this week in a different register. Azeem Azhar's Exponential View covered Microsoft's decision to cut a significant portion of its workforce — a quarter of staff, by Azhar's account — in a period when the company is simultaneously reporting its highest margins ever. Azhar's framing is worth holding. This is what the transition looks like when it arrives at a large, mature organization. Revenue goes up. Headcount goes down. The ratio of output per person changes structurally, not cyclically. For a firm that's thinking about where AI shows up in hiring plans and team sizing over the next few years, this is a leading indicator worth watching. Not because advisory firms are Microsoft, but because the dynamic plays out at smaller scale, in slower motion, in any professional services organization that takes AI seriously.

Before the close, one item I'd steer you away from spending time on. There was noise this week from Ben Hylak, a product thinker who writes about AI interfaces, to the effect that every software product should eventually dissolve into a headless service feeding an AI agent, with no user interface at all. Except, as Hylak dryly noted, for your own product, which customers love and would never want to lose. It's a sharp observation about how every software company simultaneously believes AI will commoditize everyone else's interface while theirs is special. Worth a smile. Not worth a strategy meeting.

Here's what I'd be watching in the week ahead. The Anthropic IPO will generate a wave of coverage, most of it focused on the share price and the investor story. Try to stay focused on the customer composition data and anything Anthropic says publicly about how enterprise clients are embedding Claude into workflows. That's the operational signal. Everything else is financial noise for your purposes.

And the one thing worth bringing into your leadership meeting this week. Ask your team to describe, concretely, one workflow where AI is already being used inside the firm. Not hypothetically — actually being used. Then ask. Who reviews that output before it reaches a client or a colleague who depends on it? If no one can answer that question cleanly, you have found your governance gap. Writing down the answer — just for that one workflow, just to start — is the most useful thing you could do with the next thirty days.

One thing to try this month

Pick one workflow where AI is already being used inside your firm. Ask: who reviews that output before it reaches a client? If no one can answer cleanly, you've found your governance gap — and writing down the answer for that one workflow is the most useful thing you could do in the next thirty days.

Questions for your leadership team

  • If frontier AI costs keep falling and capability keeps rising — which Anthropic's S-1 now makes a documented probability rather than a prediction — which parts of our current workflow design will look obviously wrong in eighteen months?
  • For every AI-assisted output that reaches a client or a compliance-sensitive process right now, can we name the person accountable for its accuracy? If not, what would it take to be able to name them?
  • As the ratio of AI output to human payroll rises across the industry, what does the supervisory role at this firm actually require — and are we hiring and developing for it?

Sources