Big Tech sets out its pitches on AI agents

In the years before the iPhone, there was general agreement in the tech world that mobile access to the internet would be the next big thing. But it took many false starts, from Windows Mobile to Nokia’s Symbian operating system, before Apple showed the way, in the process resetting the direction for much of the tech industry.

Something similar seems to be happening now with the development of AI agents. Many tech leaders believe software that can take actions on your behalf, removing the drudgery of routine work and personal tasks, represents a huge opportunity. Exactly what form this will take, however, is less clear. The rewards for finding the right formula are likely to be huge: personal agents could become a new entry point to the digital world and turn their makers into the next gatekeepers.

So far these include Muse from Meta and dots from OpenAI, while Microsoft has revamped its Copilot assistant to act more autonomously, Apple added agentic features to its Siri assistant in its latest iOSG and Google launched a new work assistant in Gemini. Several start-ups are also angling for attention, most notably Instinct, which launched an invitation-only test over the summer.

The contrast between Meta’s Muse and OpenAI’s dots shows how far apart the agent visions are. Both feature cute, personalisable mascots and come from companies with a large number of consumer users, but their intentions could hardly be more different.

Meta pitched its agent as a tool for handling day-to-day tasks of all kinds. By contrast, OpenAI chief executive Sam Altman brushed such things off as almost beneath his company’s interest, dwelling instead on the kind of complex work tasks you would normally give to a “high-agency engineer or chief of staff”. Other companies’ agents generally fall somewhere in between.

To some extent, the positioning appears to be as much a reflection of the business models of the companies concerned as a sign of where agentic AI is likely to have its most immediate impact.

For Muse, that means extending Meta’s advertising business into taking a cut from transactions its agent instigates, making itself more central to ecommerce. It may be a difficult evolution. At the moment, about two-thirds of travel, commerce and reservation sites on the internet block Muse in some way according to analysts at Jefferies, putting pressure on Meta to forge more direct links with individual sites. As these links harden into commercial deals and traffic is steered inside a new ecommerce ecosystem, the buying habits of billions of people could be at stake.

OpenAI, by contrast, hopes to monetise the more intensive use of its work agents through high-priced subscriptions, with dots at first available only to customers on ChatGPT subscription plans of at least $100 a month.

Regardless of positioning, all face the same challenge of persuading billions of people to learn entirely new digital habits. Central to that will be instilling trust. As Microsoft showed two years ago, when an AI feature designed to remember everything a user did on their PC provoked a privacy outcry, powerful new capabilities need to be handled with care.

Recommended

Central to the question of product design is the extent to which companies need to display the capabilities of their new agents, rather than just leaving people to find their own way. Muse takes the hand-holding approach, peppering users with suggestions of things to do, while OpenAI maintains more of a blank canvas. Microsoft’s latest Copilot offers a choice between three different tabs, including a new Autopilot mode for agents that can take on longer work tasks.

Behind the issue of design lies a deeper question: will people want a single, infinitely malleable personal agent to handle everything they do, or will they turn to more specialised agents, many of them embedded in different applications? For now, many companies are hedging their bets. Walmart has its own shopping agent, called Sparky, but has also allied itself with Meta to make it easier for Muse users to browse its site. Amazon, by contrast, blocks agents from Anthropic, OpenAI and Meta, at least for now.

A world of specialists, or one agent to rule them all? The latter would confer huge power on the companies involved, particularly if it led to the sort of market concentration already seen in search, social networking and ecommerce. But any company aspiring to achieve that will first have to find the right formula to cause billions of people to transform their digital behaviour.

richard.waters@ft.com