The Household Production Boundary

Sources#

Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy

Summary#

Nearly every economic study of AI has measured work. ATLAS v1.0's sharpest move is to point the same instrument at everything else, and find that that is where most of it is happening: 86.5% of conversational AI interactions occur outside formal work. Mapping ~10M Gemini App and AI Mode conversations onto the American Time Use Survey lexicon — 74% of ATUS categories, covering 98% of Americans' non-sleep time — turns usage logs into something close to a time-use diary.

The economic point is not that home usage is large. It is that home usage lands almost entirely on the unpaid side of the production boundary, where national accounts do not look. Whatever value it creates is, by construction, invisible to GDP — and in the specific case where AI substitutes a household activity for a purchased service, it makes measured productivity fall while actual welfare rises.

Evidence note. empirical for the usage distributions; the dollar valuations are scenario arithmetic, not measurement — ATLAS assumes a 0.5–5% time-saving range because no causal estimate of AI's household time savings exists. Treat the $15–149B band as a sized hypothesis. Non-work classification is ATUS Tier-1-accurate at 72.8% and Tier-3-accurate at 23.7% (Usage-Telemetry Classifier Validation).

Where the conversations go#

Global distribution across ATUS Tier 1 categories:

Category	Share
Socializing, Relaxing & Leisure	26.4%
Education	20.7%
Work & Work-Related	13.5%
Household Activities	11.0%
Personal Care	9.0%
Consumer Purchases	5.2%
Professional & Personal Care Services	4.8%

Nearly half of all global conversations sit in leisure and education alone. Practical household management — chores, purchases, personal care, services — accounts for about another 30%.

Time allocation predicts AI usage#

The result that earns the "general-purpose technology" claim: regress the US share of AI conversations on the share of American time spent, across ATUS categories, and human time allocation strongly predicts where AI questions go. At Tier 1 the slope is 0.767 and time use explains nearly 50% of the variance; at the most granular Tier 3 the slope stays above 0.4 with ~20% of variance explained.

That relationship is what distinguishes a general-purpose household technology from a niche tool. AI is not concentrated in a few workflows — it is distributed across daily life roughly in proportion to how people spend their days, with systematic deviations that are themselves the interesting part.

The deviations: cognitive over-indexing, physical under-indexing#

Relative to time spent (US, ATUS Tier 1):

Government Services & Civic Obligations — ~20× over-represented. The single largest outlier in the dataset.
Professional & Personal Care Services — >7×.
Education — 5.8× (and higher still outside the US).
Consumer Purchases — ~3×.
Eating and Drinking — ~18× under-represented, the starkest negative gap. Traveling, sports/exercise, and caring for household members also under-index.

At Tier 3, the most over-represented specific activities are research/homework, health-related self-care, comparison shopping, financial services, writing for personal interest, and appliance/tool/vehicle repair by self. The most under-represented are television, eating and drinking, washing and grooming, interior cleaning, and laundry.

The pattern is not "AI does cognitive things." It is that AI concentrates in the cognitive planning and organizational phase of activities whose execution stays physical. People don't ask AI to clean the kitchen; they ask it how to plan the week's meals, what to buy, and how to fix the dishwasher. ATLAS notes a custom classifier judged ~82% of ATUS Tier 3 activities non-automatable given current capability — and found substantial AI engagement inside them anyway. Adoption is being driven by demand for complementary decision and information support during physical execution, not by demand for task substitution.

High-friction bureaucracy, and the after-hours finding#

The government-services outlier decomposes into something concrete. The largest sub-shares are obtaining licenses and paying taxes, fines, or fees, then civic obligations (voting, jury duty, council meetings), then social services. Cluster summaries are dominated by the words requirements, compliance, and procedures, with themes clustering on law and police, social services, immigration and visas, and taxes.

And roughly half of all medical, legal, financial, and government-related conversations happen outside standard business hours. ATLAS's framing: accessing these institutions historically required substituting away from paid work during the exact hours you were supposed to be working — a shadow cost in lost wages. AI metaphorically keeps those offices open around the clock, and the cost it removes is not the price of a lawyer but the price of needing one during the workday.

This is the same signal AI Usage Cadences found in Claude's telemetry from the temporal side — tax queries spiking 8× around the April 15 deadline, off-hours usage carrying the rhythms of life. ATLAS supplies the mechanism: institutional access has opening hours, and AI doesn't.

Valuing what GDP cannot see#

Non-market activity has been outside the national accounts since Kuznets (1934), and if a technology moves activity from the market into the household, measured GDP falls even as welfare rises.

ATLAS applies the third-person criterion (Reid, 1934) — a household activity is "productive" if you could in principle pay someone else to do it — to isolate the valuable subset: cooking, shopping, housework, childcare, gardening, odd jobs, maintenance. Americans spend 17.8 hours per week on these. Valued at the BEA's deliberately conservative $12.02/hour replacement wage:

Assumed average time saving	Per person/year	US total/year
0.5% (≈5.4 min/week)	$56	$14.9B
2%	—	~$60B
5%	—	~$149B

ATLAS chose 0.5–5% as a deliberately conservative range against a literature that suggests far more: Blank et al. (2026) estimate 43–64% time savings for household shopping/browsing; Hertog et al. (2023) project 50–60% of unpaid domestic work automatable (though including robotics). For calibration, Nguyen et al. (2026) estimate ~$172B in US consumer surplus from AI in early 2026.

ATLAS also flags that time savings are the wrong metric for much of what is happening — "many improvements AI may bring at home will likely be manifested in the quality and breadth of household activities, rather than less time to achieve the same outputs." Better decisions about health, schooling, and legal exposure don't show up as minutes.

Distributionally, the upside is gendered: men spend up to 30% less time than women on productive household activities, so the potential gains accrue disproportionately to women — unless the documented gender gap in AI adoption cancels it, which ATLAS explicitly flags as the open risk.

Coyle's guest comment: the boundary itself is moving#

Diane Coyle (Cambridge) contributes a signed section arguing that the household side of the boundary has been ignored for a measurement reason rather than an economic one — statistical agencies collect far less data on unpaid than on market activity, even though the unpaid side is a large fraction of GDP when valued at equivalent wages.

Two reasons she gives for why this matters now:

The substitutable margin is growing. BEA data show eight headline categories of market work substitutable with household production (childcare, home healthcare, food services) rose from ~11% of non-farm payrolls in 1990 to 15.7% in 2025. More of the formal economy is now the kind of thing a household could do itself.
The technology-adoption timeline is compressed. Household durables took decades to reshape time use and labor-force participation; AI is moving on a much shorter clock.

Her conclusion is a warning about reading the productivity statistics: "innovations may seem to reduce productivity in economic statistics when they shift activity into the household because only the market sector is measured." If AI displaces demand for tax preparation, therapy, or form-filling services by making households self-sufficient, the national accounts will record that as decline.

Cross-country structure#

The composition of home usage varies systematically with national income:

Higher income → higher share of productive household tasks, and a lower share of work-related conversations (the work-share inversion).
Education dominates in the developing world — already over-represented in the US at 5.8×, and a larger share still in lower- and middle-income countries. ATLAS offers a mechanism: text-based AI responses can consume ~3,000× less data than average web search results, making AI a viable information channel in low-bandwidth environments at roughly 87% lower cost. It also notes the education-outcomes literature is mixed, with cheating and reduced time-on-task as live negatives.

Connections#

Google AI & Economy ATLAS — the program; this is the analytical block with no prior counterpart in usage-economics research
Task Saturation: Broad but Shallow AI Diffusion — the other 13.5%; the work-side finding this page is the complement to
AI Usage Cadences — the same after-hours and calendar-driven demand seen in Claude's telemetry from the temporal axis; ATLAS supplies the institutional mechanism (services have opening hours, AI doesn't) that explains why the rhythms look the way they do
Anthropic Economic Index — the rival program, which measured non-work usage far more thinly; ATUS mapping is ATLAS's clearest methodological lead
Conversation Artifacts — the output-side unit of analysis; household conversations produce artifacts (meal plans, budgets, forms) whose value is exactly what GDP misses
Organizational Complements to AI — the mirror argument on the market side: firms need complements to convert capability into measured output, and households are the case where the output isn't measured at all
Experimental Learning Impact of Generative AI — the education over-representation (5.8× in the US, higher in developing countries) meets measured evidence on whether AI-assisted learning actually sticks, and on the augmentation/automation split that decides it
Usage-Telemetry Classifier Validation — ATUS Tier 3 classification is 23.7% exact-accurate, so granular household claims carry the widest error bars in the report
Market-Priced AI Exposure (the AI Premium) — consumer surplus that never reaches a P&L is precisely the value an equity-price instrument cannot see; the two measure opposite sides of the boundary
AI Employee Framing · Human-AI Accountability Redesign — the workplace accountability literature has no household analogue; nobody has asked who is accountable when AI gives bad tax or medical guidance at 11pm

Open Questions#

The $15–149B range rests on an assumed 0.5–5% time saving because no causal estimate exists for AI in the household. What experiment would measure actual household time savings, and does the effect survive contact with one?
ATLAS argues gains skew to women (30% more productive household time) but could reverse given the AI adoption gender gap. Which effect dominates in current data?
If AI substitutes household self-service for purchased professional services (tax prep, legal advice, therapy), measured GDP falls while welfare rises. Is that substitution detectable yet in the market-services data Coyle cites?

Sources#

Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy — ATLAS v1.0 §4 Home Usage (§4.1 The Broad Picture, §4.2 Comparisons with Human Time Usage, §4.3 Non-Market Welfare), Figures 10–17, Table 3; guest comment by Diane Coyle, "Time, Productivity, and the Household Production Boundary"