AI's Best Customer Right Now Is AI Itself
Three stories from one week show an AI economy increasingly built to spend on itself, while the rest of us get the demo and the bill.
Anthropic Just Showed Us What 80% AI-Written Code Looks Like, And Why They Built A Reviewer Around It
On June 4, Anthropic published a research post called "When AI builds itself." The headline number is the one that traveled: as of May 2026, more than 80% of the code merged into Anthropic's production codebase was authored by Claude, up from low single digits before Claude Code launched in February 2025. Engineers now ship roughly eight times more code per day than they did in the 2021 to 2025 baseline. One Anthropic engineer reportedly has not written a line of code in five months.
But the details that did not travel are the ones I keep returning to.
Buried in the same paper, Anthropic described what now sits on top of every line of AI-written code: a mandatory automated reviewer. Every proposed change is checked before it can merge. A retrospective analysis found that automated reviewer would have caught roughly a third of the bugs behind past claude.ai incidents before they reached production. In April, Claude shipped more than 800 individual fixes that reduced a class of API errors by a factor of 1,000. The engineer overseeing it said a human would have taken four years.
What that means in plain terms: even at the lab building the model, raw output is not trusted to ship. The bottleneck did not disappear. It moved. Writing used to be the constraint. Reviewing is now the constraint. Anthropic's report says so explicitly.
Anthropic published this report one week after confidentially filing for an IPO. The paper closes with a call for a verifiable, multi-country mechanism to slow frontier AI development if recursive self-improvement gets closer. That is a strange thing for a company to say while it is pricing its public offering. Treat the financial and the technical claims separately when you read it.
90% Of Big Companies Cannot Get AI Agents Past Pilot. Here's What "Agent-Washing" Actually Means.
Also on June 2, a French AI and data firm called ChapsVision released "The State of Enterprise Agentic AI in 2026: Agentic Reality Check." They surveyed 740 senior executives at companies generating $1B to $20B in annual revenue. The headline finding: only 10% have moved autonomous AI agents from pilot phases into full-scale production.
The blockers are not what you would expect.
71% said they are investing "a lot" in agentic AI. Money is not the constraint. The constraint is trust. 43% cited reliability and hallucinations as a top blocker. 42% cited security and privacy. 40% cited accuracy. 86% of executives cited at least one of those three as a primary reason their pilots are stuck.
This is where the term "agent-washing" enters the story. ChapsVision defines it as the vendor practice of slapping "agent" onto every product feature whether or not it does anything autonomous. 88% of the surveyed executives said agent-washing has actively eroded their trust in AI as a category. Not their trust in one vendor. Their trust in the whole space.
That number should reset any pressure you feel about being behind. The companies you assume are running circles around you, the ones with $5 billion in annual revenue and dedicated AI teams, are looking at the same blank space between pilot and production that you are. The ones that crossed that gap, ChapsVision found, did so by investing in what they called "trust infrastructure": review frameworks, real-time kill switches, governance baked into the agent orchestration itself. The unglamorous layer.
GitHub Copilot Quietly Ended Flat-Rate AI Last Monday
Starting June 1, every GitHub Copilot plan moved from flat-rate subscriptions to a token-based model the company calls AI Credits. Unlimited agentic use is no longer part of any tier. Each plan now includes a fixed allotment of credits, and overages are pay-as-you-go.
The community took it badly. The official announcement thread drew over 400 comments and nearly 900 downvotes in the first weeks. That is unusual for a product release of any kind from GitHub.
The product news is one thing. The pattern it sits inside is bigger.
For two years, the consumer-facing AI economy ran on a flat-rate model. $20 a month, $30 a month, unlimited use. That model was always a loss leader subsidized by venture capital and infrastructure spend that the vendors planned to recover later. "Later" is now. As the labs file for IPOs and start showing investors the path to gross margins, the flat-rate cap is moving. Token-based pricing has been quietly arriving across the stack: API pricing first, then enterprise tiers, now consumer-adjacent products like Copilot.
If you have AI tools in your business, two practical implications. First, predictable monthly cost is a smaller and smaller part of how AI will be priced going forward. Second, the cost of an AI workflow now depends on how often you run it. Workflows that ran cheap when AI was a fixed-cost subscription can quietly turn expensive when AI becomes a variable cost.
This is not a problem so much as a new line item on the operating side of the business. But it is one most owners have not budgeted for, because there was nothing to budget for a year ago.
What This Means
Three stories in one week, and they line up.
The vendors are using their own products on themselves to ship faster. Anthropic's engineers are eight times more productive than they were two years ago because they delegated writing to Claude and built review on top.
The customers in the middle are watching their flat-rate access turn into a meter. GitHub is the visible example. It will not be the last.
And the enterprises that were supposed to be the proof of concept, the trillion-dollar tier of companies buying agents to transform their operations, mostly cannot get them past pilot. 90% of them, on the most generous read of ChapsVision's data.
So who is the AI working for, right now, fully and productively? The companies building it.
AI's best customer right now is AI itself. The vendors are using it on themselves. The pricing is restructuring around how often you reach for it. And the customer at the end of the chain, the design firm or law practice or wellness studio, gets the bill, gets the demo, and 90% of the time gets a pilot that never ships.
The 10% who made it past pilot did not get there by picking better tools. They built the unglamorous part around the tools. That is the gap, and that is the work.
What Business Owners Should Actually Do
Audit your AI bills this month. If you have not looked at usage per tool in the last 90 days, you are pricing yourself on yesterday's model.
For any tool labeled an "AI agent," ask the vendor exactly what it does without a human in the loop. Watch the answer get smaller. That smaller version is what you actually bought.
Build a review step between AI output and customer. The 10% have one. The other 90% are stuck because they do not.
Stop measuring AI adoption by how many tools you are paying for. Start measuring by how many of those tools made it from trial into something your business actually depends on.
If you feel behind, you are not. The data this week says 90% of much larger companies are stuck in the same gap you are. The pressure to be further along is being manufactured by the people selling you the next tool.
Chantal Emmanuel is co-founder of BAMPT and CTO of LimeLoop. This Week in AI publishes every Monday.