Hi Claude, can you help me with my taxes?

A tax worth fifty hours

What can you do in 50 hours?

Many things. Practice a skill, a backcountry hike, write several articles, cook healthy meals for the week, exercise… My personal favorite: with that time, I could read ten additional minutes to my kids every weekday night for a whole year.

Fifty hours may not seem much, and yet it can be a meaningful amount. In a recent conversation with Tyler Cowen, Annie Lowrey estimated that 50 hours is what Americans spend, on average, dealing with government paperwork. An estimate she draws from the American Time Use Survey and which she considers conservative. Her new book, The Time Tax, argues that the time spent in bureaucratic chores is a tax to citizens, and a particularly regressive one.

Forever the optimist, Tyler Cowen wondered whether we will soon get all this taxed time back:

How long will it be before AI agents are just built to do almost all of this for us? You just say to your agent, “Hey, fill out these benefits form,” and then you go away, you go to sleep and it’s done? Won’t that happen within five years?

Lowrey’s answer comes with another question:

I imagine so. The questions that I have are, one, are the lower-income people, who are the users of the most burdensome programs, do they know about those agents? Do they feel comfortable using them? I don’t know.

AI won’t just give us our time back

Some of this is already happening. Chris Schmitz, Lewis Hammond and Alan Chan have compiled 84 cases across 11 jurisdictions in which a government service saw a surge in demand that the service itself, or a credible source, attributed to AI use by the public. They find that AI tools can significantly reduce some of the learning, compliance, and psychological costs that Don Moynihan, Pamela Herd and Hope Harvey have conceptualized as Administrative burden.

Yet AI won’t automatically, “agentically” we could say, give us our time back. Reducing administrative barriers will only lead to greater access if governments have the capacities to absorb the increased demand unleashed by AI. Without them, this flooding can easily strain and overwhelm an already stretched bureaucracy.

If Cowen’s five year horizon is right, this is one of the most critical challenges that governments will face in the decade, and the problem is that we do not have good ways to measure it.

Governments only see the arrivals, but the Labs see the departures

The approach taken by Schmitz and colleagues is enlightening but has clear limits, which they acknowledge. First, it measures AI impact on these services indirectly, by cataloguing those services that are experiencing increased demand and where “the affected government body itself – or a reputable, relevant source – attribute a change in demand patterns to AI use by the public”. Second, they focus on cases where LLMs can generate the text and humans navigate the process manually. They cannot capture what they call “more sophisticated use of agents, such as autonomous website navigation,” which “is not yet evident”. This is, precisely, the type of task Tyler Cowen was imagining.

I see at least two sources of data that - to my knowledge - have not been sufficiently exploited to study this issue and offer great potential. The first one would be governments themselves, who could set up their interfaces to measure when services are being interacted by AI agents. This would not capture, however, the requests created through LLMs yet navigated and submitted by humans (e.g., a complaint drafted with an LLM and then filed by a lawyer contributing to flooding a court). Governments see the arrivals, but they don’t see where the journey started.

The second source comes, precisely, from the companies owning the chats that most citizens are using to interact with governments. If only we could analyze the data from the conversations people have in ChatGPT, Claude and Gemini we would have a pretty good picture of how people can - and indeed use - AI to interact with governments…

The cool thing is that we can.

To some extent.

A new source of insights

The study by Google’s AI and Economy ATLAS, built on 15 million de-identified interactions from April 2026, indicates that nearly 1% related to help with “government services”. This may not sound like much. But the ATLAS team also compared the share of each activity in Gemini conversations with the share of Americans’ time it takes up in the American Time Use Survey – the same survey behind Lowrey’s 50 hours estimate. Government services and civic obligations come out over-represented “by a factor of almost twenty”: relative to how little of their day people spend dealing with the government, they turn to AI for it far more than for almost anything else. And nearly half of the medical, legal, financial and government consultations in the sample happen outside 9-to-5 regular working hours.

Citizens are already trying to dodge the time tax with AI.

Anthropic has also published some really interesting data on AI use in collaboration with Stanford, Oxford and METR through their Anthropic Insights pilot initiative (a sample of nearly 250,000 conversations from April and May 2026 made available - through privacy protecting methods - to researchers).

None of these three studies looks at government interactions specifically, but we can glean some insights from two of the released datasets (here and here).

The Stanford study, for example, classified conversations based on the type of user requests (185 types). Three of these types concerned government services: (1) government navigation1, (2) immigration and visa guidance2, and (3) tax advice and compliance3. These three categories summed 2,685 conversations. This is just 1.07% of the 249,834 sampled conversations in the study, strikingly in line with Google’s figures.

Though a small subset, this data lets us compare government interactions against the overall sample to assess AI assistance levels and user friction. For example, humans led in over 85% of the government related conversations, compared to roughly 70% in the whole sample.

Across the 2,685 conversations about government services, immigration, and taxation, friction was slightly more often classified as productive than in the overall sample (21.6% versus 19.8%) and less often as mixed (7.0% versus 11.9%). These are classifications made by the AI model and do not establish whether the advice provided by Claude was accurate or whether citizens got their permit, benefit or refund.

An important research agenda

The categories Anthropic and Stanford chose to analyze the data don’t let us dig deeper, but the questions are obvious to anyone who works on administrative burden: what learning, compliance or psychological costs were users trying to overcome with AI? How successful were they in overcoming them? What cues can we derive about their perception of and trust in government from these interactions? And how do these vary across types of users and government services (Lowrey’s question to Cowen)?

These are critical questions for all of us committed to improving public services and giving people back not just their time, but the health, employment, education, and other services they are entitled to and need to live a meaningful and fulfilling life.

We have a new, rich and powerful data source to understand how citizens interact with the government, and we should collectively explore better ways to use it. This probably starts seeking collaborations with the Labs, but there may be other ways too.

In the EU, every user has the right to export their own chat history, and researchers have already built tools for people to donate this kind of digital trace to a study, under privacy safeguards. A data-donation study of how citizens use chatbots to deal with the government would need no one’s permission but the citizens’ own.

Somewhere right now, someone is asking a chatbot how to deal with the state. Roughly once in every hundred conversations. One of them may well be typing “Hi Claude, can you help me with my taxes?”

This post was originally published in the Datapolis substack.

***

1 Conversations “involved users seeking guidance on government administrative processes, including driver’s licenses, vehicle registration, housing programs, social welfare benefits, unemployment eligibility, and civic documentation. Users also sought information on electoral procedures, voting data, marriage registration, and community social services.”

2 Conversations “centered on providing guidance and information related to international immigration, including visa applications, passport procedures, residency questions, and citizenship eligibility. Users also sought advice on work visas, international relocation planning, tax residency, and navigating border control requirements.”

3 These conversations “encompassed a wide range of tax-related inquiries, including filing returns, calculating liabilities, and understanding deductions across personal, business, real estate, and cryptocurrency domains. Users also sought guidance on VAT compliance, international tax rules, resolving disputes with authorities, and verifying tax identification numbers.”