You Can't Fire a Large Language Model

In financial markets, the big question on everyone's mind these days is whether or not AI will produce enough profits down the road to justify sky-high stock valuations and aggressive corporate bond issuance. I think the answer is no, for two reasons:

1. Most of the benefits of AI will go to individual consumers, not corporations
2. Corporations pay for accountability, not output, and LLMs can't provide accountability

The Benefit to Consumers

I have been a paid subscriber to model providers almost from the beginning--first OpenAI, and now Anthropic. The value I've gotten vastly exceeds the $20 I spend every month. I will give three examples:

1. I noticed pooled water in one of my bathrooms a few days ago. Once upon a time, I would have been paralyzed by indecision and anxiety about that problem for weeks. It's not a big problem, but I don't know anything about plumbing so I wouldn't even have known where to start in trying to fix it, and fear of screwing something up, or searching for a long time and not finding a solution, would have prevented me from even starting. But with LLMs? I just described the problem to Claude and said I wanted it fixed. It did not solve the problem for me. But it did give me enough basic information about the problem that my "fear of the unknown" turned into annoyance at a routine breakdown, and suggested two fixing options: DIY or Taskrabbit. I picked Taskrabbit, but the basic point is that it transformed a problem that felt paralyzing into something that felt actionable. It held my hand long enough for me to shift my emotional state into one capable of solving the problem.
2. I started using Claude Code as an agent on top of a file vault I've built through Obsidian. This is completely wild and it is difficult to describe how powerful this is until you start using it yourself. It's like having an executive assistant that is constantly tracking problems, new contacts, and my mishmash of data dump ideas and sorting and organizing them for action. My former co-founder, David Strawn, thinks this is the end of software, where we no longer use apps or programs but just have data and we tell AI agents to do stuff with it. I'm not that forward of a thinker but I can tell you that in four days my default working system has shifted from random files scattered all over my computer to this Obsidian file vault with Claude Code open on a terminal window next to it. It isn't reliable enough yet to give me true peace of mind, but it's vastly better than me trying to remember all this stuff on my own.
3. It is an extremely powerful search engine. I do not use chatbots to write my articles, but these days, I usually start with a chatbot for research. I have a separate piece cooking about the Treasury's General Account. When I started researching this piece, I told Claude what I wanted to write about and asked it to find five primary sources (from places like the NY Fed, BIS, and IMF) about the TGA and the flow of tax receipts. It found me three excellent pieces and two okay ones. I never would have found the excellent pieces with traditional search (instead I would have gotten a bunch of SEO-optimized garbage.) The value of the learning I got for those three good pieces easily exceeds $20, and that was just one morning of usage.

(Anthropic isn't going to lose me as a sub. Their revenue is safe. But I don't think having Claude+Obsidian manage every Walmart employee's calendar so that they don't ever forget about a meeting isn't going to 3x Walmart's profits.)

Corporations Do Not Care About Output, They Care About Accountability

The problem is that the consumer benefits I described above won't translate to increased profits for corporations, because these things aren't where corporate employees spend their time and energy. Here's a story:

At McKinsey, I once worked on a media study with an analyst named Andrew. It was a typical consulting workload--travel M-Th, 14 hour days, mostly at the client site.

In a lucid moment, I asked Andrew, "In a typical 14-hour workday, how much real work do you think we get done?" His response: two or three hours.

When I said real work, I meant deep work, the kind that Cal Newport talks about. In 2009, Cal Newport wasn't a thing and there was no term for this. But I meant the kind of high-focus, high-effort cognitive work that leads to a new insight or a meaningful change for the client.

The rest of the work, 11 or 12 hours, was basically slop. Pointless analysis, meetings, check-ins with distracted clients, filler PowerPoint pages, and the like. This is where the majority of corporate employees spend the majority of their time. This is because corporations are risk-averse environments, where people aren't rewarded for taking risks that pay off, but are punished for taking risks that go wrong. That's why everything has to be approved by a committee, discussed at a meeting, and signed off by twelve executives. It's safer for everyone if there's nobody to blame when something goes wrong.

This is a form of accountability--diffused accountability, with blame spread throughout multiple layers across dozens of people. This blame-absorbing accountability structure is expensive to maintain, and doesn't rest on analytical output.

AI isn't going to short-circuit this process. 5x faster spreadsheets from the finance department is not going to speed up Walmart's decision-making process or change where its employees put their attention. The real attention goes into consensus-building, risk mitigation, policies and procedures, and the like.

Individual Accountability Requires Knowledge, Not a Report

Moreover, to the extent that true value lies in understanding the analytical artifacts--being a true subject-matter expert in your area--AI isn't going to reduce human headcount either. The New York Fed has to decide whether it wants an in-house municipal bond expert or if it doesn't. If it does want an expert, having generative AI produce a bunch of municipal bond reports isn't going to obviate the need of having a person whose opinion you can draw on when something goes wrong. If it doesn't want an expert, maybe there was no value in being a muni bond subject-matter expert at the Fed in the first place.

I am a finance professor who teaches a discussion-based class. At the beginning of the summer, I asked Claude if there was any part of my teaching workflow that I could outsource to an agent. An obvious candidate was my weekly market discussion notes:

It's immediately obvious that an LLM can generate a vastly better summary in a much shorter time than the 5-6 hours it takes me to do this every Sunday night. But having a good summary isn't the point--it doesn't produce accountability to my students. For me, as a professor leading a classroom discussion, I have to invest the time reading these stories, looking up price action, thinking of potential discussion questions, to internalize the knowledge necessary to have a rich conversation with the students. I can't lead a discussion if I don't know what I'm talking about. There's no shortcut--the time spent working is the entire point.

So what part of my job can be outsourced to an LLM? I can't outsource the learning I have to do myself. Students can't outsource the engagement with a real human in a physical classroom that drives the attention, focus, entertainment and unexpected twists and turns needed to produce real learning anyway.

So all this comes back to the value for the buyers. An individual like me will pay because I got results--a fixed leak, learning TGA, or whatever. But what results is a corporation getting? If they don't get the value they need (accountability), why would they pay? And if they won't pay, does the revenue supporting these stock valuations and this bond issuance exist?