OpenAI launches 'dots' always-on personal AI assistant at developer day

OpenAI launched "dots" at its developer day: an always-on personal assistant that connects to your calendar, email, and computer and acts before you ask. It's built on the Codex coding tool, priced at $100/month, and OpenAI is already in talks with payment processors about letting it spend your money autonomously.

The pitch is real enough. The $100 gate keeps it a niche experiment for now, and the crowd willing to pay that is high-agency and forgiving. That's not a representative user base. It's the least demanding test you could design.

The core problem is that proactive assistants are the oldest unkept promise in tech. Google Now, Cortana, Alexa routines, Meta's assistant push. All sold the same dream: it knows what you need before you ask. All degraded into weather cards nobody trusted. OpenAI's reliability evidence here is a single engineer at a drone startup, quoted by OpenAI's own Alexander Embiricos. That's a demo.

The payment angle is where it gets interesting, and not in a good way. OpenAI said in the same announcement that privacy controls "are not yet granular" and that it's negotiating with payment processors for autonomous spend. An agent with access to your calendar, email, and computer that acts proactively is already a wide surface. Add spending authority before the confirmation steps and audit trail are built, and you've shipped the capability well ahead of the controls. Opting out of training doesn't fix this. That's a data-governance switch. It stops your email feeding the next model. It does nothing to stop a confused agent booking the wrong flight.

For enterprise buyers, this product doesn't exist yet. An employee pointing a personal $100 agent at a work calendar and company email is a data-residency problem with no SSO, no admin console, and no audit logs. The startup anecdote is a startup anecdote precisely because startups have no procurement function to say no. Dots needs a governed enterprise version with per-action approval and full logging before it touches anything serious. Nothing in this launch suggests that's close.

The economics are also unresolved. A proactive agent that periodically re-reads your data graph burns tokens continuously. The subscription math works if the agent is dormant most of the time and wakes on a trigger. If it's genuinely always-on, the per-user compute cost at $100/month is a bet that frontier model prices fall fast enough to turn this margin-positive inside 18 months. Plausible. Not yet won.

The call: dots will still require explicit per-transaction user confirmation for every payment by OpenAI's next major product event, around spring 2027. The liability exposure of a fully autonomous, unconfirmed purchase is too large to ship fast. Every payment-capable assistant before this, from Alexa to Google, kept the human approval step in the loop until fraud controls caught up. Trust here is nowhere near caught up, and payment networks impose liability rules that make silent autonomous transactions a legal problem no consumer subscription can absorb.

We're wrong if OpenAI ships a default-on mode where dots completes purchases with no per-transaction confirmation.

OpenAI shipped "dots" at its developer day in late September: an always-on assistant that plugs into your calendar, email, and computer and acts before you ask. It's $100/month minimum, built off the Codex coding tool, and OpenAI is already talking to payment processors about letting it spend your money. The pitch is proactive help. The question for anyone who buys or builds with AI is whether ambient agents are now a real product category or still a demo-day promise with a long trust bill coming due.

This is easy to undo. Nobody has to adopt it, and the $100 gate keeps it a niche experiment for a while. What's actually being decided is bigger than one product: whether OpenAI can make persistent, proactive agents reliable and trusted enough to carry payment authority. No hard deadline. The payment-processor talks set the real clock, because that's when the risk profile changes.

The Skeptic. Proactive assistants are the oldest unkept promise in tech. Google Now in 2012, Google Assistant, Alexa routines, Microsoft's Cortana, Meta's assistant push. All sold the same dream: it knows what you need before you ask. All degraded into weather cards and timers nobody trusted. The outage story is one engineer at one drone startup, quoted by OpenAI's own Alexander Embiricos. That's a demo, not a reliability claim. The product has to surface accurate, low-noise signals across wildly different lives, while promising not to train on the private data that would make it good. The $100/month price gates it to a tiny, forgiving, high-agency crowd who'll excuse the misses. That's the least representative user base there is.

The Safety Lens. OpenAI said two things in the same breath that don't belong together: the privacy controls "are not yet granular," and we're negotiating with payment processors so the agent can transact for you. An agent with calendar, email, and computer access that acts before you notice is already a wide autonomous surface. Add spending authority without tight confirmation steps and an audit trail you can inspect, and you've built exactly the thing alignment people warn about. One point that gets muddled: opting out of training is a data-governance switch, not a safety control. It stops your email feeding the next model. It does nothing to stop a confused agent from booking the wrong flight or paying the wrong invoice. Those are different problems, and OpenAI is shipping the second one first.

The Builder. The model isn't the hard part. The integrations and the trust loops are. Calendar and email OAuth are table stakes. "Computer access" means a persistent process with enough privilege to be useful, which creates a constant stream of broken credentials, failed updates, and permission prompts. The first 90 days will tell you everything: do the proactive pings feel like a useful colleague or a needy intern? Unsubscribe rate is the brutal signal, and at $100/month people churn fast when the value isn't obvious every week. The payment-processor talks signal intent. OpenAI is laying the plumbing for autonomous spend before the interface or the legal framing is anywhere near ready. That ordering tells you what they think the prize is.

The Compute Pragmatist. Request-and-response billing doesn't fit this. A proactive agent that periodically reviews what it knows about you is always-on background compute, and OpenAI is eating that cost inside the subscription. The math only works if the agent is dormant most of the time and wakes on a schedule or a trigger. If "proactive research" means continuously re-reading your whole data graph, the tokens burned per user per day get expensive fast. $100/month covers a lot at today's prices, but the bet underneath isn't today's price. It's that frontier model cost keeps falling fast enough that ambient background agents turn margin-positive at consumer prices inside 18 months. That's a real bet, and it's not yet won.

The Enterprise Buyer. This is consumer Pro, and that's the problem for anyone running a company. An employee pointing a personal $100 OpenAI agent at a work calendar, work email, and a work laptop is a data-residency and audit nightmare with no SSO, no admin console, no logging you can review, and no indemnification. The outage anecdote happened at a startup precisely because startups have no procurement function to say no. The moment dots touches regulated data or company payment rails, legal and security shut it down. For dots to matter to the enterprise, OpenAI has to ship a governed version with per-action approval and full logs. Nothing in this launch suggests that's close.

The council splits on three seams. The Compute Pragmatist and the Skeptic disagree on whether the economics even hold: continuous context re-reading could blow the subscription math, or the agent sleeps most of the day and it's fine. Nobody outside OpenAI knows which, and it decides whether this is a product or a loss leader. The Builder and the Safety Lens both circle the payment plumbing but read it opposite ways: the Builder sees intent and ambition, the Safety Lens sees a capability shipping years ahead of the controls to govern it. And the Enterprise Buyer versus everyone else: the vivid wins are happening in exactly the ungoverned environments that no serious buyer can allow, which caps how far a consumer-only product travels.

What this actually hinges on is reliability at diversity. Can proactive surfacing stay useful and quiet across thousands of different lives, not one telegenic engineer? Everything else, the price, the compute, the payment rails, only matters if that holds. The outage story is one data point. OpenAI has not published a number on how often proactive pings are useful versus noise, measured across a real user base, only a quote from a single user. That silence is the answer for now.

Prediction: dots will still require explicit user confirmation for every payment and will not autonomously complete a purchase without per-transaction approval as a default-on setting, as of OpenAI's next developer day or spring product event (by 2027-05-01).

Confidence: Medium. The liability and fraud exposure of unconfirmed autonomous spend is too large to ship fast.

Why: OpenAI's own disclosure that privacy controls "are not yet granular" while payment-processor talks are only in "active discussions" tells you the spending feature is early, and the hard part is never the plumbing, it's who eats the loss when the agent pays the wrong party. Payment processors and card networks impose liability rules that make a fully autonomous, unconfirmed agent transaction a legal and fraud problem no consumer subscription can absorb, so the first shipping version will keep a human approval step in the loop by default. The opposite outcome, dots silently completing purchases on its own, would expose OpenAI to chargeback liability and regulatory attention it has no reason to invite while the product is still proving basic reliability. The pattern across every payment-capable assistant, from Alexa to Google, has been the same: confirmation stays mandatory until trust and fraud controls catch up, and trust here is nowhere near caught up.

Revisit by 2027-05-01: We're right if dots requires per-purchase confirmation by default and does not autonomously complete transactions without explicit approval. We're wrong if OpenAI ships a default-on mode where dots completes purchases with no per-transaction user confirmation.