Billions of people have a personal AI agent working on their behalf by 2031

I think that it’s extremely unlikely if you look out five years from now, for example — whatever period of time you want — that you don’t have billions of people with a personal agent that understands your goals and that is just working on your behalf 24/7 to achieve your goals in whatever the domain is that you care about.
Mark Zuckerberg, 2026-07-29 — Meta Q2 2026 earnings call

Restated 2026-09-24: Everyone is going to have an agent that really intimately understands your goals and everything around you and can help you get stuff done.

Sealed at 27%

Locked 2026-09-28. Confidence 62%. Uncertainty ±8 pts. Resolves 2031-12-31. Status: PENDING.

How the engine got there

This house agrees with the direction of this statement and was on record before it was made. We carry TEC-04 / CAL-13 at 75% — that a frontier lab documents an AI agent completing a full week of professional work by end-2027 — and booking, paying, scheduling and messaging are minutes-long tasks far inside that horizon. Our technology-adoption base case, published before this sentence existed, has US adults using generative AI at 74% by 2030. So capability is not the question, and neither is whether people use AI at all. The question is the step this statement actually makes: from billions of people who have an AI to billions who let one act for them with their money, their health data and their relationships. That is delegation, and delegation diffuses on trust rather than on capability — online banking, digital wallets and voice assistants each took a decade or more to reach majority use, and the first public reaction to this very statement was about privacy and dependency. Decomposing, on the steelmanned reading below: P(agents reliable enough for routine consumer errands by 2029) = 0.90, anchored to our 75% on the far harder workweek bar two years earlier; P(an action-taking agent is switched on in front of two billion or more people by 2030, given that) = 0.75 — Meta alone can put one inside WhatsApp and Messenger, which is a product decision rather than an adoption curve; P(two billion people actually delegate at least monthly, and it is measured as such, by the end of 2031, given availability) = 0.40, because use lags availability and a platform reporting conversations is not a platform reporting actions. The product lands at 27%. The disagreement is about when and how many, not whether: we expect personal agents to become mass-market, later than five years from July 2026 and on a slower curve than chat assistants.

What counts as correct

CORRECT if, on or before 2031-12-31, it is documented that at least two billion people worldwide use a personal AI agent at least monthly — an AI that keeps a persistent memory of the user’s goals and takes actions on their behalf (sending messages, booking, buying, paying, scheduling or managing accounts) rather than only answering questions or generating content. Documented means either (a) a single provider — Meta, Google, Apple, OpenAI, Microsoft or any other — publicly reports at least two billion monthly users of such an agent, with the metric defined as action-taking use, or (b) independent measurement (Pew Research, Gallup World Poll, Reuters Institute, GSMA, ITU or equivalent) supports at least two billion such users. Steelmanned on three axes: "billions" is read at its lowest value, two billion; monthly use counts as having one; and any provider’s agent counts, not only Meta’s. A monthly-user figure for a chat assistant does not resolve this CORRECT unless the provider states that it counts action-taking use.

We are wrong if

no provider reports two billion monthly users of an action-taking personal agent, and no independent measurement supports two billion such users worldwide, by 2031-12-31.

Causal chain

Capability is no longer the constraint

This house already seals TEC-04 at 75%: a documented autonomous professional workweek by end-2027. Routine consumer errands are minutes-long tasks far inside that horizon.

Agents able to run a person’s routine errands exist well before 2031.

Distribution can be switched on

Meta reaches more than three billion people a day and rolled business agents out to over a million businesses on WhatsApp and Messenger in the quarter of this statement; Apple and Google control the phone operating systems.

Access for billions is plausible by 2029-2030 — having an agent available is not the bottleneck.

Delegation runs on trust, and trust diffuses slowly

Letting software act with your money, health data and relationships is a different act from asking it a question. Every earlier delegation technology took a decade or more to reach majority use, and one large fraud wave can reset the curve.

Use lags availability; the gap between billions who have an AI and billions who let it act is where this seal sits.

Historical precedent

2013-2018 — WeChat Pay inside a messaging app

The strongest case for the claim: a messaging app became the payment and errand layer for hundreds of millions of people in about five years. It proves a chat app can carry delegation at national scale — in one country, with little card habit to displace.

2011-2020 — Voice assistants

Siri, Google Assistant and Alexa reached billions of devices, but delegation stayed shallow — timers, music, weather. Ubiquity without agency is the default outcome this seal is weighing.

2022-2026 — The generative-AI diffusion curve

The fastest consumer adoption recorded — a majority of US adults using generative AI within about three years (our tech-adoption series). It shows how fast people will talk to AI; it says less about how fast they will let it act.

Directly related seals

More on this desk

Browse the Technology & AI desk, the full projection registry, the calibration ledger, or the record by subject. Scoring is explained in how a prediction is made.