PublicTechnology reported that two in five websites run by the UK’s devolved administrations fail to meet accessibility requirements. The deadline passed almost six years ago. Shortly before, Patrick Vallance had been named chair of a new Prime Minister’s AI Taskforce, with a mandate to accelerate artificial intelligence adoption across the public sector.
The taskforce has a unique opportunity to kill two birds with one stone: make government services operable by AI agents and use the same investment to repair accessibility failures that years of monitoring have failed to fix.
The technical overlap is straightforward. Many browser agents rely heavily on a website’s accessibility tree, the structured representation of headings, controls, labels and states that screen readers also use. Playwright’s Model Context Protocol server works from that tree. So do many of the browser tools used by coding agents. The reason is cost: a modern page can serialize to more than 15,000 tokens of raw markup, while the accessibility tree of the same page runs 200 to 400, because it discards everything that is not a control, a label or a state. Google’s guidance makes the connection explicit: “Everything we suggest to make a site ‘agent-ready’ also makes sites better for humans.”
There are limits to the overlap. Accessibility and agent readiness are not the same thing. Colour contrast and captions may matter greatly to people without affecting an agent. An API may help an agent without helping a person using a screen reader.
But the shared foundation is substantial: labelled form fields, predictable controls, meaningful headings, clearly expressed errors and stable page structures. These are basic accessibility requirements. They are also what allows an agent to understand a service and complete a transaction without guessing what the government intended (and for goodness’ sake, no PDFs or JavaScript-rendered sites).
Tom Loosemore, who co-founded the Government Digital Service, demonstrated the problem in March when he tried to build a single application that would tell any UK resident when their bins were collected. He had Claude. He understood government technology better than almost anyone. He still gave up.
The problem was not the model. It was 365 councils structuring the same basic service in different ways: broken interfaces, bot-blocking, JavaScript-heavy forms and inconsistent address requirements. Westminster asked for a road name where other councils accepted a postcode. “Me and my LLMs have been beaten by the bins.”
If one of the people who built GDS cannot get an agent reliably through a bin-collection lookup, the constraint is not artificial intelligence. It is the condition of the underlying service.
That gives the Vallance taskforce a concrete objective. Make agent operability a central measure of public-sector AI adoption and define it at the level of the transaction.
A department can launch a chatbot, buy model licences and build another RAG system whose benefits are visible mainly to the team that built it. The test is whether an agent acting for a citizen can find a service, understand its requirements and successfully apply, renew, pay, report or appeal.
The agent would use the same authoritative rules, transaction service and audit trail as the citizen-facing interface. It may reach that service through the website, an API or an emerging agent protocol. A separate machine interface is not necessarily a problem. A separate government, operating through different rules and weaker accountability, is.
This requires two things, and today’s digital public infrastructure has neither.
The first is services that agents can read and operate. That begins with accessibility repair. Forms need labels. Buttons need meaningful names. Errors need to explain what failed. Procedures need to be expressed consistently enough that neither a screen reader nor an agent has to reverse-engineer the department’s intentions.
The second is a way for the state to know which agent is acting for which citizen and under what mandate. Call it Know Your Agent, the delegation-layer counterpart of Know Your Customer.
Identity alone is not enough. Government needs to know who authorised the agent, what it is allowed to do, how long that authority lasts, whether it has been revoked and what record exists of the actions taken. Without that layer, agents may be able to navigate government services but cannot safely act through them.
The taskforce could start with the highest-volume public transactions. The test would be whether leading agents can discover each service, understand the requirements, complete the process and produce an auditable record. The completion rate could sit next to the government’s existing accessibility-monitoring results, on the same page and on the same day.
An unlabelled form field currently produces an accessibility finding and, perhaps, an updated statement promising that the problem will eventually be fixed. Under an agent-mediated service, the same field will produce a failed transaction, a complaint, an information request and a demonstration that goes wrong in front of the prime minister.
I have watched enough digital-government programmes to know which of those moves a budget.
Accessibility failures have survived because their costs are dispersed, and dispersed across a territory under an electoral system that aggregates constituencies by location rather than by need. Monitoring identifies the problem, but the finding rarely becomes a sufficiently visible operational failure to force investment.
Agent operability changes that. The same defects begin to affect a prominent government programme with measurable completion rates, political sponsorship and attention from the centre. A problem previously treated as a compliance matter becomes a barrier to delivering the government’s AI agenda.
Budgets dedicated to repairing old websites and inaccessible forms are difficult to secure. Budgets for adopting AI are not. The government can use the second budget line to fund work that should already have happened under the first.
This does not mean accessibility becomes valuable only because machines now depend on some of the same infrastructure. Services must continue to work for people who use them directly, including those whose needs do not overlap with those of agents. Nor should agents be given a privileged route around safeguards that apply to citizens.
AI adoption creates a new constituency for fixing old systems. It turns inaccessible and badly structured services into visible obstacles to a ministerial priority. That makes funding more likely.
The alternative is that agent traffic grows while government websites remain broken. Agents will repeatedly attempt transactions that citizens abandon because they are too confusing, inconsistent or difficult. Friction was the rationing mechanism. Agents do not get bored.
The people identified by accessibility audits will still be on those pages.
Here is the opportunity for Vallance’s taskforce: make agent operability a core objective, test it through real public transactions, publish completion rates beside accessibility results and establish a common framework for identity and delegation.
That would allow the government to meet a target it has missed for almost six years, and one it has not yet thought to set itself: building the foundations of the agentic state.