AI changes tasks: what do we know about its impact on work?

Debates about AI and work often jump from a capability to a conclusion: if a model can write a report, the job of the person who wrote it will disappear. Between those statements lie organisational choices, quality checks and other responsibilities that cannot fit inside a single demonstration.

Anthropic’s Economic Index, published on 15 January 2026, analyses conversations from November 2025 and adds measures of complexity, autonomy and success. It is a window into Claude usage, with limitations the report acknowledges, not a picture of the entire labour market. [Anthropic]

A task is not an entire profession

Imagine someone who prepares quotations. Writing the document is only part of the job: they may also need to clarify requirements, negotiate exceptions and check that commitments can be met. A tool helping with writing does not automatically resolve the rest of that process.

The ILO’s May 2025 study measures task exposure to generative AI. It considers job transformation more likely than complete replacement in many occupations, because they continue to involve human input. Exposure means potential for change, not observed redundancies. [ILO]

Measuring usage also means reading the methodology

Anthropic notes that people may choose tasks they believe the model will handle well. They can also break tasks down and correct results. This matters when comparing real conversations with fixed tests: they do not observe exactly the same situation. [Anthropic]

Our view is that technical capability, adoption and organisational effects should remain separate. A company can test a tool without deploying it. It can deploy it and redistribute work. It may also need more review than expected. None of these steps follows from a conversation count alone.

It would also be inappropriate to present estimated time savings as a direct measure of net benefit. An internal evaluation would need to decide whether preparation, corrections and final verification are included. Without that boundary, two productivity figures may describe different jobs.

The useful question inside a business

We suggest observing a specific process over a defined period. Which tasks change? Who reviews the result? Which errors need correcting? Which hours can be used for other work? That record would be more useful for internal decisions than extrapolating a global headline.

A support team offers an example: the time needed to prepare a reply could be separated from the time needed to resolve the case. If the message goes out sooner but the problem remains open, an improvement in one stage does not explain the whole customer experience.

The studies provide language and signals for investigating change. They do not determine how benefits will be shared or which policies each company will adopt. The question we suggest tracking is what changes in actual work and what evidence supports it, without turning a technical possibility into an inevitable future for employment.