What ChatGPT and Claude changed about online work

Explainer / analysis · Prepared with AI assistance. Published by The AI System Institute. Editorial approach and corrections.

A faster draft is a visible result. A better business is a much harder thing to measure. Much of the confusion around AI and independent work comes from treating those two outcomes as though they were the same.

Research on writing and customer support offers more specific evidence. It also shows why a single productivity percentage cannot describe every task, worker or organisation.

Two findings, two different settings

A 2023 experiment by Shakked Noy and Whitney Zhang studied professional writing tasks. Participants with ChatGPT access finished faster on average and received higher quality ratings. The reported changes were 40% less time and 18% higher assessed quality in that setting.

A separate study published in the Quarterly Journal of Economics in 2025 examined an AI assistant used by 5,172 customer-support agents. It found a 15% average increase in issues resolved per hour, with substantial differences across workers. Less experienced workers benefited more; the most experienced group saw smaller speed gains and some quality declines.

Neither result is a universal productivity multiplier. One setting involved bounded writing tasks; the other involved support work inside an organisation. Neither study establishes how much revenue a solo operator should expect from an AI-assisted resource.

What usage research adds

The initial Anthropic Economic Index approached the question through anonymised Claude usage. It classified observed tasks and distinguished forms of collaboration from more automated use. The report is useful evidence about activity within its dataset, with the limits that come from observing one service.

Usage data helps identify what people ask an assistant to do. It does not by itself show whether the output was accurate, whether it was used or whether the task produced a useful commercial result. Those questions require additional measurement.

The unit of analysis is the task

For an independent operator, a task-level comparison is often more informative than a general impression. A draft can be assessed for time spent, factual accuracy, missing steps and the amount of revision it requires. A spreadsheet can be checked against known inputs and expected outputs.

A faster first pass can be a poor trade if checking it takes longer than doing the task directly. Conversely, a tool that helps someone produce a clearer outline may be valuable even when the final writing still takes care. The relevant comparison includes the review work.

The work that remains outside the draft

Deciding what a customer needs, obtaining permission to use an asset, understanding an ambiguous brief and responding to a dissatisfied reader are not eliminated by a writing assistant. Tools can help organise the work, but responsibility for the result stays with the operator.

The same is true of confidentiality. A task may involve private customer information or material subject to a contract. Whether it belongs in a third-party tool depends on the agreement, the service’s current controls and the circumstances. A convenient workflow is not sufficient permission to share data.

A useful reading of the evidence

The research supports careful experimentation with defined tasks. It also supports caution about broad claims. “This task was completed faster in this study” is a much more precise statement than “a one-person business now does the work of three people.”

The economic consequences depend on what happens after the time is saved: more review, a different service, a better resource, a lower price or simply fewer hours worked. That next step belongs in the analysis. It cannot be inferred from the speed of a draft alone.