The old digital divide was often described as a gap in access to computers, broadband or information. Autonomous agents may create a stranger version. Consider two applicants confronting the same bureaucratic form. One can ask a basic chatbot for instructions. The other can delegate the job to an AI system that gathers records, checks deadlines, books appointments and negotiates with service providers. If that second person can run dozens of agents simultaneously, they might be able to pursue opportunities the first never has time to discover. The asymmetry is no longer only about who knows more. It is about who can send more effective representatives into the world.[1]
Matthew Sharp, Omer Bilgin, Iason Gabriel and Lewis Hammond describe this phenomenon as agentic inequality in a Perspective published by Nature Machine Intelligence on 20 October 2026. Their proposed framework has three dimensions: availability, or whether a person can use an agent at all; quality, or how well that agent performs; and quantity, or how many agents the person can deploy. The research is a conceptual analysis. It does not measure the current global distribution of agents or establish that any proposed inequality has already emerged at a particular magnitude.[1][2]
Why agents are different from better search
With a search engine, differences in money or skill can affect what information people find. An agent is intended to act: it can monitor a deadline, send a message, gather evidence, apply a rule and possibly commit resources. It therefore competes directly with other agents for limited things: customers, appointments, attention, favourable prices or political influence. A firm with excellent proprietary data, faster models and a larger agent budget could win a greater share of such opportunities even if its human staff were no more knowledgeable.[1]
The framework's third dimension, quantity, is particularly revealing. Two people could subscribe to models of equal quality but enjoy very different outcomes if one can run a single delegate and the other can deploy a large fleet. The relevant unit of advantage becomes sustained, parallel action. That does not mean that more agents always produce more value. They can duplicate effort, make inconsistent decisions or become constrained by real-world supply. Still, the ability to attempt more tasks creates a plausible mechanism for unequal outcomes.[1]
There is an equalising scenario too
The same technology could lower barriers. A small business might obtain the document research, scheduling and first-draft contract preparation that once required an additional employee or expensive adviser. A household could receive help understanding a benefit entitlement, translating a form or challenging an administrative mistake. Oxford researchers explicitly discuss such potential benefits. A free or low-cost agent capable of reliable basic bureaucracy might reduce inequality even if premium agents remained better at complex negotiations.[1]
The effect depends on how the systems are distributed and governed. Open weights, competition and public-service access could expand availability; data lock-in, expensive inference and private platform privileges could restrict it. If an online service prioritises requests from major paying agent operators, a seemingly equal right to use an agent may not produce equal practical access. If agents can transact without meaningful authorization, their expansion may instead magnify fraud or mistakes.[1]
How to make an abstract warning measurable
A useful public metric would compare performance across price tiers on ordinary tasks with actual stakes: correctly filing an application, finding a deadline, completing a purchase without unwanted fees or obtaining the same appointment under the same rules. Quality should include successful outcomes, error rates and the amount of human correction required. Quantity should record the limits on concurrent delegations and sustained use. Availability must encompass language, device, disability access and eligibility, not just the presence of a free sign-up button.[1]
The researchers raise policy ideas, including a public or subsidised baseline of capable agency. Such proposals deserve scrutiny: a universal public agent could improve access but might also become a single vendor's gateway to government services. Another approach would be interoperability and enforceable limits on privileged treatment. The proper choice requires empirical evidence that the field has only begun to collect.[1]
A new denominator for digital opportunity
The paper's value lies in giving policymakers a question that cannot be answered by counting AI accounts. How much useful action can different people actually delegate, at what reliability and price? If a free agent frequently fails the tasks that matter most, its nominal availability may conceal a serious inequality. If efficient public agents handle routine paperwork well, the technology could help people who need assistance most. The future is not predetermined by the capability of the most expensive model. It will also depend on who is allowed to put that capability to work.[1]
Sources & further reading
- Agentic inequality Nature Machine Intelligence