AI agents are starting to act on our behalf, from dealing with companies to navigating healthcare and the law. In a Perspective published in Nature Machine Intelligence, research led by the Oxford Martin AI Governance Initiative argues that people with the best agents could pull further ahead, or that widely shared agents could narrow existing gaps, depending on choices being made while the technology is still new.

Some people may soon have a workforce of AI agents at their command. Others may have a single basic assistant, or none at all.

A chatbot waits for instructions and answers them. An AI agent is given a goal and works out how to reach it, planning the steps, taking action and seeing the task through on someone's behalf, from filling in forms to negotiating a deal.

That is what raises the stakes. When an agent acts for you, the gap between a basic agent and a powerful one, or between one agent and many, becomes a gap in real outcomes. The authors call this "agentic inequality", and they argue it is more than the digital divide in a new form.

Agents could just as easily narrow existing gaps. If capable systems are widely available, they could simplify complex processes and put expertise that once came with a lawyer's or accountant's fee within everyone's reach.

"The good news is that none of this is set in stone. Agents are still new, and the choices being made right now, about pricing, platforms and public services, will decide who benefits," said Matthew Sharp, Senior Postdoctoral Researcher at the Oxford Martin AI Governance Initiative. "But this isn't just about individuals. The same gaps can open up between businesses and between countries, so we need to get it right at every level, from how public services are designed to how countries cooperate on AI. Once these patterns set, they're far harder to change."

"The good news is that none of this is set in stone. Agents are still new, and the choices being made right now, about pricing, platforms and public services, will decide who benefits," said Dr Matthew Sharp, Senior Postdoctoral Researcher at the Oxford Martin AI Governance Initiative. "But this isn't just about individuals. The same gaps can open up between businesses and between countries, so we need to get it right at every level, from how public services are designed to how countries cooperate on AI. Once these patterns set, they're far harder to change."

Three ways the gap could open

• Availability: whether you can use an agent at all, which depends on cost, digital skills, language and whether one is offered where you live.
• Quality: how well your agent reasons, how reliable it is, and which other systems it can connect to and act in.
• Quantity: how many agents you can run. Running many at once, the paper argues, is less like using a better tool than directing a distributed workforce.

Access to a powerful AI model, they point out, is not the same as access to a working agent. Turning one into the other takes extra infrastructure, still concentrated among a few providers, plus specialist know-how.

Skills matter too. For now, setting complex goals and managing what you delegate to them can demand a high level of digital literacy. That barrier is likely to fall, the paper suggests, as people become able to instruct agents in everyday language.

The authors expect broad access to basic agents, but continued inequality in the most valuable uses. Paying customers can get agents that plan more effectively, make fewer mistakes and connect to more apps and services. Higher charges for running large numbers of agents could put that kind of scale out of reach for individuals and small organisations.

An uneven playing field

Agents will not only act for people. Increasingly, they may deal with each other, negotiating, bidding and strategising on their users' behalf. Contests once settled between people could be decided by networks of agents operating at machine speed, with weaker agents losing out to stronger ones.

Early evidence points the same way. In one study cited in the paper, a Stanford-led team had buyer and seller agents negotiate over real products. Buyers with weaker agents paid about 2% more, sellers with weaker agents lost up to 14% of their profit, and one buyer's agent committed its owner to a $900 phone on a $500 budget.

Companies could also use agents to scale up manipulative sales tactics, such as false scarcity claims and hidden information, beyond anything a human sales team could manage. If they are widely available, consumer agents that compare prices, flag deceptive pricing and negotiate on people's behalf could push back.

Widening the gap, or closing it?

Similar tensions run through public services, business and politics:

• Public services: affluent people could use premium agents to navigate healthcare or legal systems. If services are redesigned around agents, people without them risk being left behind. Managed well, though, agents built for the public good could help those who most often lose out. They could fill in forms and claim the benefits and support owed to them, easing administrative burdens that fall hardest on people with lower incomes.
• Business: dominant firms could use their own data to build more capable agents, win more customers and gather still more data, a loop smaller rivals may struggle to match. They could also run far more of them. But if powerful agents become widely available and prices keep falling, the reverse could happen: start-ups could run operations that once needed large teams, lowering the barriers to competing.
• Politics: well-resourced groups could deploy swarms of agents to run influence campaigns or overwhelm public consultations. But broad access could also strengthen the capacity of citizens and grassroots groups, with agents drafting policy proposals, managing campaigns and taking part in political processes on their behalf.

Shaping what comes next

There is no single fix. Instead, the paper sets out a research agenda, from measuring how far free agents trail premium ones to updating the law so people can seek redress when an outmatched agent leaves them worse off.

One idea they put forward is "universal basic agency": treating access to an agent good enough to make a real difference as something every citizen is entitled to, whether through a public-option agent or subsidised private ones.

They are clear that none of this is simple. Act too early and policymakers are working in the dark; act too late and the patterns are locked in. The paper also cautions against heavy-handed fixes: capping what agents can do could hold back innovation, and a single government-provided agent could lock in one supplier.

Dr Matthew Sharp led the paper with co-authors Omer Bilgin (University of Oxford), Iason Gabriel (Google DeepMind) and Lewis Hammond (University of Oxford and the Cooperative AI Foundation). The work was supported by the Future Impact Group and the Cooperative AI Foundation.