Apparently, I Read S-1s Now

Until last week, I couldn't have told you what an S-1 was. I'd never read one. I'd never wanted to.

Then Anthropic filed theirs on October 26, and my feeds filled up with people arguing about the numbers. Revenue growth, compute commitments in the hundreds of billions, what it all means for the valuation. I have no opinion on any of it. Big numbers, crazy numbers, no context. I'm not qualified to tell you whether they're good.

But the threads on X and Reddit kept saying the same thing: this is a very different S-1. And Anthropic and OpenAI are so ingrained in how my life and work has changed, that I was just curious and very interested.

So I did the obvious thing for someone in my line of work. I asked Claude, ChatGPT and, just to see what an open-source model would do, GLM 5.3. I told them I didn't really care about the numbers. What I cared about was what the filing actually says about the nature of work, and why people thought it was so unusual. Explain it like I'm not a finance person, because I'm not.

There's an obvious irony in using Claude to understand the company that makes Claude.

It became a long back-and-forth. And every few turns I'd pause, because something in there made me think about what I'm doing at OneDigital, or a question we've been arguing about for months internally. Sometimes it was something I'm running into at SmartCue, which I've been trying to run almost entirely with AI agents.

I had nothing to compare the S1 to understand why people said it was so different, so obviously, I also asked Claude to check it against other IPO filings going back to 2001 and tell me what the big deal was. Long, alarming lists of risks are standard. Lawyers write them defensively, so a scary risk factor isn't a confession. But a few things stood out to me. The filing values the company's own use of its product against its payroll i.e. AI spend versus human spend. And the company says it may spend money on the problem it's trying to solve with the potential for no return, or that its own product could be an existential crisis for humanity!

Then there was the number that was surprising to a layman like me. All that revenue people are arguing about, comes from companies that have barely started using AI. For the average Fortune 500 company paying Anthropic more than $100,000 a year, that bill is still less than a tenth of one percent of its own total expenses. Even for the ten biggest spenders, the average is under 1%.

Inside Anthropic, it's a very different story.

Their AI bill passed their payroll

In April, Anthropic's own use of Claude was worth more than its entire monthly payroll. Between January and September, it grew more than 250 times.

I've spent a good part of this fall in conversations about AI costs at OneDigital. Some of that is about getting the bill down. More of it is about something harder: connecting what we spend on tokens to what actually changed in the business. Not how much people used AI. What it did. Did it change or improve the outcome for the individual using AI? Or for the practice? Or for the company as a whole?

Anthropic doesn't seem to have that problem with its own bill, which makes sense. It is one of only a handful of companies like that, because each version of Claude helps build the next one. And it can point at what the token spend resulted in, because most of its code is now written by Claude to build a better Claude!

They're in a different position, to be fair. Their AI spend is their product getting better. But for most businesses, including OneDigital, AI is a tool inside a business that helps consultants advise their clients on benefits, insurance and retirement.

When Anthropic looks at its AI bill, it can point at the latest flavor of Claude. When I look at ours, mostly I can point at a usage chart. What I want to be able to point at however, is a client who signed up with us because the consultant used AI to get that deal over the line. Or a  disgruntled client stayed with us because AI allowed a consultant to do something that seems almost magical because s/he wouldn't have been able to do it better. Or perhaps someone on our team who got to go home on time or early. 

Although - there are probably a handful of people who do that already! I don't fault them, but I don't know if their managers or HR would feel the same!

Their engineers stopped writing code. Now they only check it.

Over 90% of the code contributed at Anthropic in September was written by Claude. Each engineer is putting out more than 45 times as much code a day as in 2024. The filing says the engineers' job is increasingly to "direct and review."

Then, the same company warns that people reviewing AI's work can get "alert fatigue," and may not have enough context to judge what the AI actually did.

Put those two together. The AI produces more and more. A person still has to check it. There's far more to check than anyone could have written themselves. And the people checking are getting tired.

At OneDigital we talk a lot about AI giving people time back. I believe it does. I'm just not sure how much of that time goes back into checking and validating the AI's output. If AI can produce in an afternoon what used to take a team weeks, somebody still has to read it, catch what's wrong and decide what's safe to use. That somebody might be the new bottleneck!

So did productivity actually go up? Or did the work move from doing to checking, and did it take the same hours, or more?

I've felt a version of this at SmartCue. Last Christmas I spent an inordinate amount of time cleaning up its code so an AI could understand it. I wasn't adding anything. My goal was very clear: I can't be doing day-to-day work and building/selling SmartCue. If every change needed me to read it line by line, I'd have jeopardized my day job. So for me, it was imperative to set the AI up in a way that I could trust it to do the job in my absence. No checking. Or rare checking every other weekend!

I spent my time up front, so I wouldn't have to spend it checking forever. Not every business has that luxury or is as binary.

Anthropic's paid when you use more. I'm trying to make our people do more with less.

Anthropic says it believes enterprise customers get a "compelling return on investment." The evidence it offers is that customers keep spending more. Then, in the risk factors, it says customers who can't "demonstrate measurable return on investment" may cut back.

That second sentence is my whole job now!

This week I was on a prep call with a few colleagues before a session with our leaders on the usage dashboard I built. The dashboard shows who's using our AI coworkers and how much. It doesn't show whether anyone does their job better. Someone pointed out that you can get a lot of output from an AI and have no idea whether any of it is good. Someone else said you might not have used it at all, just given up and done the work another way.

Where we landed was almost the opposite of Anthropic's revenue model. The goal is for someone who barely cares about AI to type two lines and get what our best power user got after hundreds of rounds of prompting. If we pull that off, the people who use AI the most won't look very different from the people who use it the least. Essentially - less token usage AND more/better output is the goal.

A long conversation with an AI doesn't mean someone's good at it. A short one might mean the tool finally works. Neither is black and white though.

The agents come with access. The people don't come with authority.

Anthropic calls its agents Virtual Collaborators. We call them coworkers. They "carry the access rights, system connections, and context that their organizations grant them," and the company is building toward work that runs for "hours, days, or weeks." So are we at OneDigital.

Picture an account manager with a few of those coworkers. In an afternoon, they can pull together ten different analyses of a client's renewal, the kind of work that would have kept a few people busy for days.

They now have the output of a team. What they don't automatically have is the team lead's job. Deciding which of the ten actually matters. Spotting the one that's wrong. Choosing what to tell the client, and what they're allowed to promise on the company's behalf.

AI is giving people the output of a team before anyone has worked out the job that comes with it.

That's close to how we've been rethinking jobs at OneDigital. We split work into reducible and irreducible parts and try to help people spend more of their time on the irreducible kind. I'd mostly thought of that as client relationships and expertise. The filing made me think it also includes deciding what's worth doing, what to trust, and what to put your name on.

I don't want to get too comfortable with that either. AI is getting better at judging and prioritizing too. Today's irreducible work could be next year's reducible work, which is why we keep redrawing the line and think of more and more irreducible work that only a human can do. For now.

Anthropic can grow into the time it saves. Not every company is Anthropic!

From the end of 2024 to September 2026, Anthropic's headcount grew more than five times. Over the same stretch, its quarterly revenue grew 86 times. In the last year alone, revenue per employee went from $1.4 million to more than $6.6 million. I promise this is the only financial piece I understood because I thought the correlation was weird and unique! I'll explain why -

That's one way a company can absorb a huge jump in capacity without anyone losing a job. It grows faster than the capacity does.

Most companies don't grow 86 times in two years. So when a team's work gets faster, the time has to go somewhere else. More clients. Better service for the clients you have. Less overtime. Fewer future hires. More experiments. Higher margins.

Every one of those is a management choice. None of them happens by default. I don't think most companies, have sat leadership, HR and talent teams down to decide which one they're choosing, and for which roles. I don't know of a playbook for it. I know we are trying to write that playbook as we speak.

There's also the employee's side. Going home earlier is a real benefit. Treating every recovered minute as capacity the company has to take back is a choice someone makes.

And I ask this question to whoever will listen, probably because I work at a benefits company - If AI turns a 40-hour job into a 25-hour job, is it still a full-time job? Could someone reasonably hold two? Under the Affordable Care Act, full-time means 30 or more hours a week. Benefits eligibility, non-competes, NDAs and our whole idea of loyalty to one employer rest on the assumption that one person does one job for about 40 hours. I'm not sure that assumption survives this.

Anthropic seems to know the jobs question is sitting there. Its risk factors say that if job displacement rises quickly, governments may step in to slow AI adoption. They also say the company's own public statements about job loss "may be used against us." Also - this was crazy for me to see in a filing! Who would say that about their own company?! Or at least in those stark terms?

What happens to kids entering the workforce for the first time?

I asked the models what the filing says about entry-level work. The answer was nothing. "Entry-level," "early career," "apprentice," "reskill" and "upskill" don't appear anywhere in it. "Junior" appears once.

What the filing does have is a $100 million Frontier Academy, aimed at what Anthropic calls a key constraint on enterprise AI adoption: "the shortage of engineers who can take AI from an initial use case into production." The plan is to train 10,000 of them by the end of 2027.

So a company building toward AI with "the judgment and initiative of a brilliant colleague" has decided its bottleneck is trained people, and is paying to train them. That's an apprenticeship program, whatever they call it.

At OneDigital, our AI coworkers start as interns. They get promoted to apprentice, then full-time. I hadn't thought about it until this week, but those are the job titles junior people used to hold.

Which leaves another interesting and I think important question - If the junior work goes to agents, how does anyone become the senior person who can tell when an agent got it wrong? Do companies stop hiring freshers? If they don't, what does a fresher do on day one? What should a degree even teach?

There's a hopeful version at OneDigital. For our Medicare phone agents, I built an AI trainer that lets new hires practice calls with an AI before they talk to a real customer. Done well, AI training might give a new person more feedback and more reps than sitting next to someone experienced for six months ever did. But this is just one example. There are still so many, unanswered and almost uneasy questions that most of us don't have an answer to. And the jury is out whether or how soon can the AI just take those calls instead of a human.

Am I an analyst now? LOL

I went in just curious to see what the hype is about.

What I have is a list of questions, from a document I'd never have opened in my adult life:

• What is our AI bill buying, and who decides whether that's enough? How?
• When people get time back, is checking answers the new bottleneck? How do you overcome it?
• What does good AI use look like, if it isn't simply more AI use?
• Who decides which of an agent's ten analyses gets acted on? Who's the throat to choke if AI makes a mistake?
• What's a full-time job when the job realistically only takes 25 hours?
• How do juniors even break into a workforce?

Obviously don't buy the stock or not based on my analysis! PS - GLM 5.3's analysis was meh - kind of like a college grad's report. Academic and underwhelming. Make what you will of it.