I pay for my AI subscriptions in my own name, as a consumer, but very little of what they do is personal. I have a couple of volunteer and nonprofit roles, I handle a fair amount of family and household admin, and I have side projects, including this site. Most of the AI work for all of it is building and maintaining websites, keeping email and admin moving, tracking money, and putting together small tools.
A few years ago that work would have meant a website builder, an email or security service, a finance tracker, or a few hours a month from a part-time admin. Now an agent does a lot of it, and the products I might have bought for each job don't get bought.
Websites are the clearest case. I no longer pay for a website builder or hosting. Cloudflare Workers and Pages host my sites for free, and Claude Code designs, builds and deploys them. When I started my career, putting a website on the Internet was hard enough that companies like Rackspace, where I later worked, were built around doing it for people. Now it couldn't be simpler, and for sites like mine, I think web design and hosting are solved.
I also build most of my own small utilities now, like apps that find duplicate files and images, and one that downloads videos, transcribes them and summarizes them. And I finally have an assistant again. Grok Bot doesn't press the button to book a trip or schedule a meeting, but it does most of the legwork ahead of time, so booking or approving is easy for me. That's real work, but I probably wouldn't have hired anyone to do it, since I'm not sure I needed someone full time.
All of it runs through a handful of subscriptions in my name. Right now I pay for Claude Max at $100 a month, ChatGPT Pro at $100, Cursor Pro+ at $48 a month on an annual plan, SuperGrok on an annual plan that works out to about $25 a month, and Google AI Pro at $19.99. That comes to roughly $300 a month, or about $3,500 a year, and it doesn't count the separate bills I pay for API usage, like the scripts that send my drafts to OpenAI, Google and xAI models for review.
The purchase nobody records
Most of the SaaSpocalypse talk is about companies cutting seats. There's another effect that's harder to see, which is the purchase that never happens. When an agent can build a small tool that does what you need, a subscription to a product that does it more generally stops looking necessary, and the decision never shows up anywhere as a cancellation or a line item.
Companies are doing the same thing on purpose. In McKinsey's State of AI survey this year, 32% of respondents said their organizations had decided against buying at least one software product or feature because they could build it internally with agentic coding tools, decisions that only show up because McKinsey asked.
Some of this isn't a lost sale, like the assistant I wouldn't have hired. A small organization might never have paid for a better website, and plenty of people would have kept using spreadsheets. In those cases the AI subscription is new spending rather than a replacement, but it still goes to an AI vendor and onto a personal card.
I'm not the only one billing business work to a consumer plan. Chris Lausten wrote on X that he pays $300 a month for SuperGrok Heavy and spent another $1,500 in overages in one week, "just running simple bot stuff for a small business."
This looks a lot like the shadow IT that arrived with smartphones and cloud apps, except for what the tool does. When people brought Dropbox to work, they were using the kind of product IT had declined to buy. An agent can do the job of a product, so the product doesn't get bought at all, and nobody has to approve or turn down anything.
I see a small version of this in my own habits. Last week I posted that Grok Bot is a better search engine for my Gmail than Google is. Gmail is free, so nobody loses a sale there, but it's the same habit that makes a paid tool's renewal start to look optional: I didn't cancel anything, I just open the app less.
A flat rate at home and a meter at work
There's a plainer reason the work ends up on personal plans: they're a much better deal. A consumer subscription is a flat rate, and a heavy user gets far more out of it than the same work would cost at API prices. AgentsView, which estimates what my coding sessions would have cost at API prices, puts my Claude Code sessions on one laptop over the past month at more than ten times what I pay for Claude Max. SemiAnalysis estimates that subscriptions are only about 10% of Anthropic's revenue but can use over 40% of its inference compute, so I'm clearly not the only one getting more than I pay for.
A company doesn't get that deal. On Anthropic's Enterprise plan, the seat fee "doesn't include any usage," and "every token your team consumes is billed separately at standard API rates," according to Anthropic's billing page. OpenAI now meters eligible ChatGPT Enterprise plans by the token, on top of seat fees. Aaron Rubin joked about the gap back in May: "Big news. I migrated from my personal Claude Max to our company account so I can pay $1,600 instead of $200 for the same monthly usage." Someone paying for a plan out of pocket can get far more done than their company would get for the same money, and none of it shows up on the company's books.
Reuters reported, citing people at its investor day, that Anthropic's third-quarter revenue was $16.8 billion. If its IPO filing reports individual Pro and Max plans separately from Team and Enterprise, we'll at least see how big the personal side is, though not which of those accounts are doing company work.
Inside companies, one employee at a time
Deloitte UK published its first GenAI Workforce Survey on September 16. Based on 25,000 UK workers, it found that 17% of people who use GenAI for work pay for at least one tool themselves, which Deloitte puts at about £958 million a year, and that 31% use GenAI without their employer knowing. When Deloitte asked self-funders and unapproved-tool users why, 21% said their tools beat what the company provides, and 14% said the tools are essential to the job but their employer doesn't pay for them.
In its shadow AI write-up, Deloitte says all this makes it hard for IT teams to measure usage, "meaning internal tools and expensive licenses often go unused." A company can keep paying for seats while its people do the work in subscriptions they pay for themselves, and the only signal is that the seats get quieter.
When Wes Bos asked on X how much employers cover for AI each month, the 208 replies were all over the map. One reply said his employer covers about $100 a month for work only, while he pays for Claude Code and OpenCode himself for side projects. Luca Capone wrote, "Day job covers zero for me. I pay my own Claude sub out of pocket," and called it the best money he spends.
The data goes with it
When work moves into personal accounts, company data goes with it. My own accounts hold email and documents from organizations I volunteer for, and the questions I raised in Superintelligence Is a Permissions Problem about what an agent can touch get harder to answer when the account belongs to an individual. An organization can't revoke access it never granted or audit an account it doesn't know about.
How much of your "personal" AI spend is really for work, and what have you stopped buying because an agent does the job? Tell me on X, LinkedIn, or at contact@exagentica.ai.