Anthropic says its AI is building its replacement (and the number is surprising)
Anthropic measured how much of its R&D its AI already does. What changed, why it matters, and how to assess it in business.
Introduction
Anthropic and the AI measuring its own work highlight a practical question for any company: it is not enough to “use AI”; you also need to know how much of the work it is already doing. In its own report, Anthropic measures how much of its research and development is performed by its AI, shifting the discussion from intuition to measurement.
What changed in the report
On September 17, 2026, Anthropic published an internal report about its research and development. The focus is not a vague promise, but a concrete metric: what share of the work its AI already performs. According to the report itself, the figure rose from under 1% in February to a much higher percentage in August. The company also acknowledges, with numbers, that clear limits still remain.
Why it matters for business
When an organization measures how much work AI already does, it stops relying only on impressions and starts making decisions based on evidence. That matters for prioritizing investment, identifying repetitive processes, and separating tasks AI can support from tasks that still require human oversight. In business terms, this can apply to sales, support, or finance: not as an immediate replacement plan, but as a way to understand where real opportunities for partial automation or assistance exist.
What “measuring” means in practice
Anthropic’s approach is also an evaluation framework. If a company wants to know whether AI is creating value, it can define by function what share of tasks the technology already performs. For example, support teams may measure ticket classification; sales teams may measure draft writing; finance teams may measure initial report preparation. The key is to distinguish fully automated tasks from assisted tasks and review what still depends on human review.
Limits and risks
The report itself suggests caution: more AI work does not mean perfect work or operation without control. There is also a risk of measuring poorly if “done by AI” is not clearly defined. An ambiguous metric can overstate progress or hide critical dependencies. That is why quality, supervision, traceability, and exceptions should be reviewed before drawing conclusions.
Evaluation checklist
- Define one metric per function.
- Separate automated tasks from assisted tasks.
- Identify where AI still does not reach.
- Review quality and the need for supervision.
- Compare the metric with real operating goals.
How to apply it in your business
Start with a simple measurement: what percentage of tasks AI already does in each area. Then compare that number with business value, risk, and the need for human control. That AI assessment is the starting point for deciding where to scale, where to adjust, and where not to move forward yet.