What the Anthropic Economic Index Tells Enterprise AI Teams About Where Agents Are Used

The Anthropic Economic Index is built mainly from usage logs. Anthropic samples real traffic to its Claude models and has a classifier label each record with the work task it resembles. It then publishes the totals, as the report of 15 January 2026 describes. One part of that data covers traffic that companies send through the programming interface (the API), which is how software products and automated workflows call a model.

No new edition has appeared this month. The Anthropic economics team page, checked on 11 November 2026, lists the report of 26 June 2026 as the most recent full Index report, followed by a country brief on Canada dated 14 July. The public dataset agrees, with its newest folder labelled release_2026_06_26. Related reports followed on 16 October and 5 November. This article reads the business figures in them for one question: what can a head of AI learn about which agent use case to pick?

What the business data covers

Anthropic began reporting on business use in its report of 15 September 2025 and has carried it forward since. The business sample is traffic sent directly to Anthropic's own API, which the reports call first-party API traffic. The report of 15 January 2026 drew on one million such records from 13 to 20 November 2025. It calls them "enterprise data" because the sample "mostly represents enterprise use". Each record is a single prompt and its response, with nothing linking it to the calls before or after.

The report of 24 March 2026 states that first-party traffic is separate from "third-party platforms such as Amazon Bedrock or Google Cloud Vertex". Calls that a company routes through its cloud provider are therefore absent from the sample. The unit is also a record, so the shares below are shares of calls, and they say nothing about spending or business value.

Which tasks companies send through the API

Software work comes first. In the January 2026 report, computer and mathematical tasks were the largest category in the November 2025 API sample, at around half of records. The report gives 46% in one chapter and 52% in another without explaining the difference, so we use the rounded figure. The single most common API task was changing software to fix errors, at about one record in ten. The ten most common tasks accounted for 32% of API records, up from 28% three months earlier.

The second group is routine office work. Office and administrative support tasks rose by 3 percentage points over the same three months to more than one API record in ten, and the report reads this as businesses automating "routine back-office workflows". A footnote in the January report lists examples with their share of all API records: personalised sales emails to business buyers at 0.47%, replies to business email at 0.28%, invoice processing systems at 0.24% and email classification at 0.23%.

The March report added two workflows whose share at least doubled between November 2025 and February 2026. One was sales and outreach automation, such as lead qualification research and cold email drafts. The other was automated trading and market operations. The same report of 24 March 2026 notes that customer service tasks, including automated support for billing problems, were already common in API data. It put the ten most common tasks at 33% of API traffic.

How automated the use is

Business use is delegated far more often than consumer use. The Index groups interaction patterns into automation, where the user hands over a task, and augmentation, where a person learns from the model or works back and forth with it. Some records fit neither group, and the automation label says nothing about whether the task succeeded. In the September 2025 report, 77% of API records showed automation patterns and 12% showed augmentation. The January 2026 report put automation at about three quarters of API interactions, against less than half in the consumer app.

Delegated and independent are separate measures in the Index. The January report scores autonomy, meaning how much decision-making is left to the model, and gives a translation request as an example of a task that is fully delegated with almost no decisions to make. API traffic showed "uniformly lower autonomy at all levels of complexity" than the consumer app, according to the same report. The work in an average API record would take a person an estimated 1.7 hours alone, against 3.1 hours for a consumer conversation. Our reading is that a business call in that sample was often a narrow, well specified step inside a larger program.

Success is lower on the API than in the app. Claude, acting as the classifier, judged the task complete in 49% of API records and 67% of consumer conversations. API success ran at around 60% for tasks under an hour and roughly 45% for tasks that would take a person five hours or more, crossing 50% at 3.5 hours. The January 2026 report cautions that customers choose which tasks to send, so these rates may overstate what the model could do on tasks nobody tried. The sample was taken before the release of Claude Opus 4.5, so it says little about newer models.

The automation share has not moved in one direction. The March 2026 report says augmentation rose slightly in both the app and the API, and that automation "decreased sharply" in the API data. The exact shares sit in an appendix figure and the page gives no cause. It does say that Claude Code, Anthropic's coding agent, had grown to "a large share of sampled traffic" on the API and that it splits work into many smaller calls. A change in what the sample contains may explain part of the shift, though that is our inference.

What changed in June and after

The report of 26 June 2026 changed the method. It moved to continuous daily sampling, and a footnote says the first-party API figures do not include Claude Code. Its chapters deal mostly with the chat, Cowork and Claude Code products. The breakdown of API output types is left to an appendix, and the page gives no API automation share. The November 2025 business figures therefore have no like-for-like update for the middle of 2026.

That report does compare delegation in Claude Code with delegation in the chat and Cowork products. On a five-point autonomy scale, Claude Code sessions averaged 0.37 points higher than chat and Cowork conversations, according to the June 2026 report. About two thirds of the gap came from the same tasks being handed over more fully.

One of the reports dated 16 October looks inside a single company, Anthropic itself. By August 2026, running time for interactive agents outnumbered employee hours six to one, with employee hours estimated from headcount at 40 hours a week. Interactive sessions were idle, waiting for human input, for around 80% of their elapsed time. What the report calls meta-work, such as writing instructions for agents and monitoring pipelines, rose from 10% of the average employee's interactive sessions in fall 2025 to 22% in spring 2026. The figures are in How employment is changing at Anthropic. An AI developer is an unusual employer, so read this as one case. The 5 November report on scientific research uses 207,049 sessions on consumer plans and adds no business data.

What a head of AI can conclude

What follows is our own reading of the data, which the reports do not offer themselves, set out as four working rules for picking a first or next agent use case. The tasks that show up in volume in the API samples tend to have a clear input and an output that can be checked: code fixes, document and invoice processing, email triage, support replies and sales research. A candidate that looks like these has precedent, at least with this vendor. If it depends on judgment across many open choices, the business data offers little precedent.

Task length is the second rule. With API success falling from about 60% under an hour to about 45% past five hours in the January 2026 data, we suggest testing a long job as a chain of short steps with a check between them. Context is the third. The September 2025 report found that API customers using the model for complex tasks tended to supply lengthy inputs, and it warned that company knowledge which is scattered or not yet digitised could hold back deployment. Check that the documents and records a task needs are in one place before you commit to it.

The fourth rule concerns people. In Anthropic's own account, human attention became the limit on agent output, and a growing share of employees' sessions went to instructions and review. A plan that counts hours saved should also count the hours needed to supervise. The October report covers one firm, so make your own estimate.

What the Index cannot tell you

The business data is one vendor's traffic. It covers Claude only, and only calls made directly to Anthropic, so it leaves out traffic through the cloud platforms named in the March 2026 report and all use of other models. It reports shares of records. We found no breakdown by industry or company size on the pages cited here, and no measure of whether a deployment paid off. In our view a task being common says nothing about whether it earns its cost.

The labels come from a model. Anthropic describes its measures as "directionally accurate" in the January 2026 report and says individual classifications are imperfect. An API record is one call, sometimes from the middle of a workflow, so the API figures cannot show a whole agent run or its final result. The definitions have also shifted between editions, as the Claude Code change shows.

How to use it this quarter

1. List your candidate use cases and mark the ones that resemble the high-volume API tasks in the Index.
2. For each candidate, write down what a correct output looks like and how a person or a test would check it.
3. Estimate how long the task takes a person, and split anything longer than a few hours into checked steps.
4. Count the review time the agent will need each week and name the team that will provide it.
5. Run the shortlisted task on a sample of your own cases and record the success rate before any wider rollout.

Where QueryNow fits

QueryNow builds the foundation enterprise AI runs on, and the agents that run on it. Each engagement starts with one scoped workflow and a written definition of done. If you are choosing a first or next agent use case, you can talk to our team or read our guide to AI agents.