OpenAI's Dots Needs Your Inbox to Work, That's Exactly the Problem Privacy Advocates Keep Raising

OpenAI's Dots, the agentic assistant the company unveiled about a month ago at a San Francisco developer event, is built around a different premise than most AI tools on the market: it doesn't wait to be asked. Rather than responding to prompts, Dots is designed to monitor a user's goals, inbox, and work tools and initiate tasks on its own. OpenAI showcased the product again at its London Dev Day, rolling it out first to enterprise customers and premium U.S. subscribers, the highest-spend tier of users, before any broader release. It isn't available in Europe yet, pending regulatory approval.

The consent problem nobody asked Dots about

Signal president Meredith Whittaker has been one of the sharpest critics of what proactive agents like Dots actually require to function. The core issue she raises is straightforward and hard to engineer around: making an assistant like this genuinely useful means giving it deep access to a user's communications, but the people on the other end of those communications, everyone who emails or messages a Dots user, never agreed to have an AI system ingest their messages. Whittaker has also flagged the specific risk to people in sensitive situations, warning that exposing contact information and communication metadata to an AI system this deeply could put vulnerable groups, including human rights activists, at real risk.

OpenAI's answer is narrower permissions, not a different architecture

OpenAI's public response to the privacy critique leans on gradual trust-building and granular user controls rather than rethinking the proactive model itself. Romain Huet, the company's head of developer experience, described setting his own boundaries with the product directly: he told his Dot to categorize his email constantly but never to send messages on his own behalf. That's a meaningful control for the user doing the configuring, but it does nothing to address Whittaker's underlying point, that the third parties whose messages get read never had any say in the matter at all.

The other problem: this is genuinely expensive to run

Privacy isn't the only constraint OpenAI is up against. Adam Fry, who leads ChatGPT's consumer business, has acknowledged that Dots sits on the compute-heavy side of OpenAI's product line, and the company isn't alone in facing that math. Elon Musk has said xAI's own Grok Bot carries costs that are unsustainably high relative to the revenue it generates, a rare admission from a major AI lab that the current generation of always-on, proactive agents may not be economically sustainable at today's usage patterns and pricing.

The workaround: quietly routing you to a cheaper model

OpenAI's cost fix is model routing. Certain Dots tasks get sent to Terra, a smaller GPT-6.1 model, instead of the larger GPT-6 Astra that powers the flagship experience, and according to Huet, Terra performs comparably to Astra in specific areas at a tenth of the price. That's a sensible engineering response to a real cost problem, but it also means the quality of a user's Dots experience may vary invisibly depending on which model happened to handle a given task, a tradeoff most users will never see explained.

A crowded field racing toward the same tradeoffs

Dots isn't launching into a vacuum. Meta has Muse, xAI has Grok Bot, and Anthropic has Duet, all competing in the same proactive-agent category, and enterprise AI spending has reportedly reached 177 billion dollars annually, with agents alone now consuming an estimated 80 percent of tokens processed on OpenRouter, a platform that provides access across multiple AI models. Every company building in this category is racing toward the same unresolved set of tradeoffs: how much access does an agent need to be genuinely useful, who consents to that access, and whether the compute bill for running it proactively, all the time, for every user, actually works at scale.

Why marketers and comms teams should watch this closely

For any business thinking about building on top of these proactive agents, whether for customer service, outreach, or internal workflow automation, Dots' launch is a preview of the governance questions that are coming regardless of which lab's product wins. An agent that reads and acts on communications on a company's behalf raises the same third-party consent question Whittaker is describing at OpenAI's scale, just applied to customers, prospects, and partners instead of personal contacts. The compute economics argument matters too: a proactive agent that's expensive to run today is a cost structure vendors will eventually pass along, directly or through usage caps, once the free-trial phase of this product category ends.