How Much of Their Own Work AI Does at Tech Companies

Google says 75% of its new code is AI-generated and approved by engineers. Airbnb says nearly 60% of its code is coauthored with AI. Uber says more than 10% is written autonomously by agents. Anthropic says Claude leads 26% of its AI research and development work. Each of those numbers is on the record, dated, and true as stated. None of them measures the same thing, and any article that lines them up as a league table is comparing four different questions.

This page collects every figure of this kind we could trace to a first-party, on-record source, read on September 18, 2026: earnings calls, shareholder letters, official posts and papers, and one recorded interview. Twenty-eight figures from thirteen companies, each with its exact definition in the same row, because the definition is what a reader citing the number needs and what most coverage drops. Figures that exist only in secondary reports, social posts we could not fetch, or forecasts are listed in the methodology as excluded. The one company that has published how it measures, Anthropic, gets its method set out in full, since it is the only one a team could reproduce.

There are two honest ways to answer the question and several dishonest ones. The first honest way counts artefacts: what share of the lines of code merged this quarter did a model write, and under what definition of "write": generated and then accepted by a person, coauthored with a person, or produced by an agent with nobody typing. The second counts tasks: of the work the company does, weighted by how much time it takes, what share does a model do most of. Only Anthropic has published the second kind. Everyone else has published the first, in wordings that differ from quarter to quarter, or has published something adjacent, such as how many engineers use a tool or how many pull requests each one merges, which says nothing about share.

Two labels recur in the tables. "Vendor-run" marks a figure about a product the company also sells, because a company measuring its own tool has a reason to measure generously. "Third-party transcript" marks a figure spoken on an earnings call whose text we read from a transcription service rather than the company's own investor-relations document; the call is on the record and the words are the speaker's, but the transcript is not first-party. Everything else is from a document the company published or a recording of the speaker.

These are the percentages that get quoted. Read the last column before the second one. The rows are in no order of merit; they are grouped by company and then by date.

Three rows deserve a note. Google's four figures, the latest in its Cloud Next post, are the only series from one company across time, and the wording moves from "generated by AI" to "written by coding agents" and back to "AI-generated", so the series is a trend in Google's vocabulary as much as in its code. OpenAI's 99.8% is included because it is quoted as a share of work, and it is not one: it measures which of two OpenAI tools its staff generate tokens through, which is a fact about tool choice. And Anthropic's 26% is the only percentage in the table with a denominator you can inspect, which is why section 05 is about it; the same post's oversight figures are covered in our post on Anthropic's three oversight numbers, and what an agent leading a task looks like in practice is the subject of our post on Anthropic's model-optimisation report.

These figures are also on the record and also quoted, usually next to the ones above as if they were the same kind of number. They are not shares of anything. They measure how much more gets produced, how many people use a tool, or how much support work a product absorbs.

The most useful rows here are the ones with a stated unit of work: Coinbase's pull requests per engineer and Robinhood's commits per engineer are numbers a company could compute for itself tomorrow and compare with its own last year. Meta's output per engineer would be too, if Meta said what output meant. Our collection of agent productivity statistics holds the survey and study figures that sit alongside these company disclosures.

• The denominators differ. New code at Google; code in repositories at Microsoft; code engineers produce at Airbnb; code committed at Uber; person-time-weighted research tasks at Anthropic; output tokens by tool at OpenAI; pull requests or commits per engineer at Coinbase and Robinhood. Seven denominators for one question.
• What counts as AI differs. "Generated, then reviewed and accepted" is a different bar from "coauthored with", which is different again from "written autonomously by agents". A line an engineer accepted from an autocomplete counts under the first and not the third.
• Only one company shows its method. Anthropic publishes sampling, weighting and a judge-agreement rate. Every other figure is a single sentence in a call or a letter, with no statement of how it was counted or over what period.
• All of it is self-reported. No figure on this page has been audited by a third party, and six of the thirteen companies sell the tool their figure describes.

Anthropic's figure comes from a procedure described in the appendix of its September 17 post, and it is the only one on this page that a company could reproduce at its own scale.

1. Sample the people, weekly. For each week in July 2026, a random 20% of staff from every department in the model R&D loop.
2. List the tasks. A Claude agent reviewed each sampled person's week in messages and documents and listed what they did, giving about 15,000 granular tasks.
3. Build a tree and freeze it. The tasks were organised into 542 nodes with 378 leaves, then held fixed so later months are rated on the same map.
4. Rate each leaf on a six-level scale. An independent Claude judge, seeing only evidence from that month or earlier, assigned one of six autonomy levels from Epoch AI, running from no AI involvement, through assists and collaborates, to leads and autonomous.
5. Weight by time. Each person is one unit per week, split evenly across their tasks; a category's weight is the sum, so a task that takes many people many hours counts for more.
6. Check the judge. Staff who own each area rated it blind. The model agreed exactly with humans 59% of the time, against 35% agreement between humans, and was within one level 97% of the time.

A fifty-person company can run the same procedure with a spreadsheet and any capable model. Sample ten people a week, have each list last week's tasks with hours, rate each task on the six-level scale with a written definition of each level, and weight by hours. The result is a percentage with a denominator you can defend, which is more than twelve of the thirteen companies on this page have published. It is also the same measurement discipline that turns an AI adoption claim into something a board or a buyer can check, which is where our AI transformation service begins with clients.