OpenAI launches always-on Dots into a market where 2% of firms have scaled agents

Capgemini's adoption split is more useful when the groups are combined. Organizations running agents either at scale or in part come to 14% ^[1]. The four reported groups add up to 98%, so a small slice of the survey is missing from the published breakdown ^[2]. The same firm puts the economic opportunity from agents at up to $450 billion by 2028 ^[10].

Gartner, Capgemini and Bain measure different things, and only Gartner is counting spending. It expects global spending on AI models and platforms of $64.25 billion in 2026, up from $39.31 billion in 2025 ^[9]. That is an extra $24.9 billion in one year ^[5]. Bain puts the US software opportunity in cross-system work at $100 billion and says more than 90% of it is untapped ^[12]. By Bain's count, less than $10 billion of that pool is being captured today ^[4].

The first dot is free with eligible Pro and Business Premium plans ^[3]. OpenAI is charging for the seat and including the agent in it. Business Premium costs $100 per user per month on an annual contract or $125 month-to-month ^[4]. Over a year that comes to $1,200 against $1,500, so the right to cancel costs $300 a seat, a 25% premium ^[3]. A company that expects to drop the agent if it misbehaves pays the $300.

OpenAI's own task list is narrower than its pitch of agents working with little supervision ^[1]. It says Dots can investigate bugs raised in Slack, turn designs into working apps and run read-only proactive research while users are away ^[5]. So the work OpenAI describes doing while nobody is watching is read-only research. Meta's Muse launched 21 days earlier ^[7] and asks for authorization before it sends emails or makes purchases ^[6]. The OECD says agents now coordinate harder jobs over long periods with little human help, but are not yet fully reliable ^[13].

The two companies are aiming at different buyers. Meta built Muse into WhatsApp and says it is meant to help "billions of people" ^[7]. OpenAI is building on Frontier, the platform it launched in February for agents that work across a company's apps, data and tools ^[14]. It has not dated SMS access for Dots ^[2], and Pro access is not yet live in the European Economic Area, Switzerland or the UK ^[3].

Business clients bring in about 40% of OpenAI's revenue, and CFO Sarah Friar says that share could reach 50% by the end of 2026, according to Cryptopolitan ^[8]. For that to happen, business revenue has to grow from about 0.67 times the rest of OpenAI's revenue to parity with it, a rise of half ^[6].

On the optimistic path, Enterprise admins switch the beta on widely ^[3], agents take on paid work and Friar's 50% arrives early. In the middle case, customers take the free dot, keep it on read-only research and pay for seats they would have bought anyway. In the bad case, an agent fails in the way the OECD warns about ^[13] and buyers go back to pilots. I think the middle case is the likeliest through 2026. A free first agent and a read-only task list both describe a seat sale with an agent attached ^[3][5]. The counter-case is distribution. A company that already takes about 40% of its revenue from business clients ^[8] can switch Dots on for them through an admin setting ^[3] without having to sell them a new product. I would be wrong if Capgemini's at-scale share reaches double digits in its next count, or if the business share of revenue hits 50% before the end of 2026 ^[8].