What Millions of AI Conversations Reveal About Where AI Actually Works

Most claims about what AI can do for an organisation come from people selling something — us included. So it is worth paying attention when one of the world's frontier AI labs simply publishes the data. The Anthropic Economic Index is an ongoing research program that analyses, with privacy-preserving methods, millions of real conversations with Claude — the AI model family behind a large share of the world's serious AI deployments — together with a survey of 81,000 people about how AI is changing their work. It is the closest thing the industry has to a census of actual AI use: not what AI might do, but what people and businesses are genuinely doing with it, measured quarter after quarter.

The latest report, published in March 2026 and covering usage in February, contains findings that should interest anyone running a council. Three stand out.

Finding one: customer service is where business AI lives

When the Index looks at how businesses deploy AI through the API — that is, AI wired into systems and workflows rather than a person typing into a chat window — customer service workflows are a mainstay, with automated support for payment and billing issues called out specifically as one of the most common patterns. Across the broader economy, the fastest-growing automated workflows are the high-volume, well-bounded, repetitive ones.

Read that list back as a council. Payment and billing enquiries are rates calls. High-volume, well-bounded, repetitive interactions are bin days, dog registrations, pothole reports and "are you open on the public holiday?" The global market, voting with its money, has converged on exactly the category of work that dominates a council contact centre. This is not a coincidence, and it is not vendor marketing — it is what AI is demonstrably best at today: work with clear boundaries, factual answers and unforgiving volume. Devonport City Council's experience — 43 per cent of calls resolved end-to-end by Cassie — is the same finding, measured from the council side.

Finding two: augmentation is rising, and the mix matters

The Index distinguishes augmentation — AI working collaboratively with a person, checking, drafting, teaching — from automation, where AI completes the task itself. In everyday use, augmentation is rising. In business API traffic, automation is. These sound contradictory; they are actually a division of labour. Routine, structured work flows toward automation. Judgement work stays with people, increasingly AI-assisted.

That division is precisely the design brief for AI in a council. Automate the calls that are really lookups — balances, bin days, opening hours — and the after-hours triage that otherwise wakes a duty officer for a barking dog. Augment everything else: the planner summarising forty submissions, the customer service officer who receives a transferred call with a full transcript so the resident never repeats themselves. Councils that pick one mode and force everything through it — all automation, or all augmentation — are fighting the grain of the technology. The data says the organisations getting value run both, deliberately, with a clear line between them.

Finding three: the learning curve compounds

The report's most strategically important finding is about people, not models. Users with six or more months of experience succeed at their tasks roughly 10 per cent more often than newcomers, collaborate with AI on more sophisticated work, and are better at choosing the right tool for the job. The researchers describe a learning channel of skill-biased change: the benefits of AI flow disproportionately to those who have already put in the hours.

For a council workforce, this turns the timing question on its head. The instinct to wait — for the technology to settle, for the framework to land, for a neighbouring council to go first — quietly assumes that starting later costs nothing. The data says otherwise. An organisation's AI capability behaves like compound interest: the council that started a contained, well-governed deployment this year is not one year ahead of the council that starts next year — it is one learning curve ahead, in staff fluency, in governance maturity, in knowing which problems AI should not touch. The gap between experienced and inexperienced organisations is widening, and it widens fastest early.

The quiet conclusion

There is a reassuring undercurrent in the Index that is easy to miss. Real-world AI use is not science fiction. It is overwhelmingly ordinary: answering questions, processing payments, drafting documents, handling the front desk of organisations at all hours. The frontier may be moving at an extraordinary pace, but the value being captured right now is being captured in the most familiar places. For local government, that is good news twice over. The work AI does best is work councils have in abundance. And the evidence for starting is no longer anecdote — it is the largest dataset of real AI use ever published, and it points, with unusual clarity, at the phone on the counter.