Anthropic's 2026 State of AI Agents: where deployment actually stands

In 2026, AI agents began moving from experiment to production. Anthropic's 48-page report brings together a survey of 500+ US technology decision-makers with real company cases. Here we avoid over-reading the numbers and look at how they connect to practice in Japanese companies.

Anthropic is the AI company behind Claude and also works on safety research. It is at the same time a vendor of AI agents, so the report's findings are useful but should be read knowing it is vendor-run research.

The survey was conducted with Material in late 2025. The respondents are 500+ US technology leaders. The proportions therefore cannot be transferred directly to Japanese companies or to the world as a whole.

The report says 57% use agents in multi-step workflows and 16% have moved to operation spanning multiple teams. That reads as a transition in progress: from a personal assistive tool to a mechanism that connects an organisation's work.

The most advanced area of use is software development. Over 90% use AI coding, 86% have deployed coding agents against production code, and 42% let them lead development work under human supervision. The important part is not the absence of supervision but delegating after designing how supervision works.

Use is not limited to development. Data analysis and reporting is at 60%, internal process automation at 48%, and 56% plan to extend into research and report writing over the next 12 months.

80% report measurable economic impact and 88% expect returns at least as good going forward. But 80% is the share of responding organisations — it does not mean profit or productivity rose by 80%. What each company counted as ROI also differs.

There is no single deployment model: 47% mix buying and building, 21% lean on off-the-shelf products, and 20% build entirely in-house. Building only the differentiating part of the work and using existing products for the common part is the realistic design.

The barriers cited were integration with existing systems at 46%, data quality at 42%, and change management at 39%. Picking a higher-performing model does not solve those; you need safe access to the right data and teams able to change how they operate.

The report includes cases from Novo Nordisk, Doctolib, L'Oréal and Shopify. What they share is that none tried to automate the whole company at once — each started from a specific job whose value could be measured.

To apply this in a Japanese company, start by choosing one job in one department. Write down the data the agent can read, the operations it can execute, the point where a human approves, the audit log, and who is responsible for stopping it when something goes wrong.

Scaling AI agents is not about removing people; it is about redesigning the boundaries and the accountability of work. The next step is measuring completed outcomes, rework, exceptions and the number of cases a human must judge — not the number of deployments.

That AI agents have moved from one-off assistance into multi-step real work, with 80% of responding organisations reporting measurable economic impact, while integration, data quality and change management remain the main open problems.

No. It means 80% of responding organisations reported measurable economic impact. It is not a figure for how much margins or productivity increased.

Start with one job in one department where value is easy to measure, and define data access, execution permissions, the human approval point, the audit log and who can stop it before you build.