The AI agent most companies use doesn't code — it consolidates, drafts, and organizes
Anthropic analyzed 1.2 million sessions of its autonomous agent tool and found that only 8.7% involve software development. The largest block — 33.4% — is pure operational work: reports, checklists, consolidating scattered data. The myth of agents as a tech team tool breaks apart with data from the lab itself.
There's a persistent myth about AI agents: that they're a tool for technical teams, that you need to code or have an engineering department to build them. Data Anthropic published this week dismantles it with real production numbers. They analyzed 1.2 million Claude Cowork sessions — their autonomous agent tool — from more than 600,000 organizations between May and June 2026. The result: only 8.7% of sessions involve software development. The largest use, more than double the next category, is business operations: reports, checklists, and reconciling scattered spreadsheets.
The most common agent use isn't coding — it's the work nobody wants to own
The most common tasks within the 33.4% business operations block are exactly the kind of work that exists in any company without a tech department: pulling scattered updates into a single report, building employee onboarding checklists, and reconciling spreadsheets arriving through different channels. None require code. All require time — the same time that costs money. In second place at 16.4% is content creation: drafts, slide decks, proposals. Software development, which many assume is the core use case, ranks third with just 8.7%.
6.4 hours a week: what each employee recovers in production
Four independent Q1 2026 studies converged on a consistent range: employees using agents in production recover between 6.1 and 7.2 hours per week. McKinsey reported 6.4 h, Salesforce State of Service 6.7 h, Slack Workforce Index 6.1 h, and Anthropic's enterprise telemetry 7.2 h. In the highest-friction roles — customer service (8–9 h) and senior practitioners (10–12 h) — savings run even higher. Those hours aren't recovered from creative or strategic work: they come from the same type of work the 33.4% of sessions reveals as dominant — processing, organizing, consolidating.
- Consolidating weekly project updates from 8 different sources into a single document.
- Generating a first draft of a commercial proposal from meeting notes.
- Comparing two versions of a contract and listing differences clause by clause.
- Building an onboarding agenda for a new hire based on the existing process.
- Classifying and prioritizing a support request inbox.
The cost of doing that work — and the cost of not doing it
When an agent handles customer service, the cost per ticket drops from $4.18 to $0.46 — nine times cheaper. 74% of executives who deployed agents report measurable return within the first year, and 52% already have at least one agent in production, according to Google Cloud data. But the cost that rarely appears in the calculation is the time quietly accumulating without visibility — the report nobody consolidates on time, the checklist that doesn't exist, the spreadsheet living in the email of the employee who no longer works there.
"AI agents can be applied to so many use cases, the number of businesses adopting them should be 100%. I can quickly point to dollars saved." — Fiona Tan, CTO of Wayfair (Google Cloud Blog, 2026)