Within a few years, I expect most people will meet their government and their NGO service provider through an AI agent, and they will expect that agent to finish the job.
That is my bet. Today most people still type their own answers into portals.
The best new measurement of the shift is RADAR, from Luke Jordan, Tiago C. Peixoto and Manuel Ramos-Maqueda. It tests whether an AI agent can reach a public service up to the login screen. Its authors say friction at login and payment may be doing the job it should.
That boundary is the conventional wisdom I want to push past. Governments and NGOs design services on the assumption that a human takes over at authentication. The citizen with an agent will not accept that handoff for long.
Sign Up Now for more digital government insights
Agents Cannot Even Reach the Door
The RADAR team tested four AI models on 20 services in 166 countries, then sent agents after a birth certificate. In every country, models explained services better than agents could reach them, by an average of 2.37 points on a 10-point scale. Wealth made no meaningful difference.
The UN E-Government Development Index explains 69 percent of the variation in chat scores and 24 percent of the variation in agent scores. The ranking governments use to judge their portals barely sees the problem.
RADAR also scored machine-readable access, giving credit for APIs or service buses such as Estonia’s X-Road. That score was zero in 84 percent of country-model observations.
Commerce Already Lets Agents Finish
Chris Schmitz, Lewis Hammond and Alan Chan catalogued 84 AI-linked demand surges at government services. In 87 percent of cases, people used chatbots to write text, then clicked through the portal by hand. I think the clicking is the temporary part.
Commerce shows why. Mastercard says its Agent Pay system completed authenticated agentic transactions in nine Asia-Pacific markets in 2026, each one an AI agent booking and paying for a ride. These are pilots, a handful of transactions.
The design is what matters. Mastercard issues each agent its own token and captures the consumer’s consent explicitly. The bank never asks the agent to prove it is human. It asks who authorized it.
An aside: While checking sources for this post, an AI agent tried to fetch that Mastercard press release announcing AI agents that shop on your behalf, but Mastercard’s newsroom didn’t allow the agent to read the press release about AI agents.
Login Is Where the Citizen’s Work Starts
RADAR stopped at login. For the citizen, that is where the burden begins: form fields, scanned documents, fees, appointments, status checks. Those are the tasks people least want to do and most want to delegate.
Let’s be honest about what we’re really proposing here. A service that lets an agent find the birth certificate page and then demands a human finish the form has moved the queue online. My bet is that people will route around services their agent cannot complete.
Authenticate the Mandate and Let the Agent Work
The strongest case against me is flooding. Schmitz reports that some German social courts attribute a 55 percent year-on-year rise in 2025 caseloads largely to AI-generated claims. Agents that complete transactions could multiply that.
My answer is to keep authentication and change what it checks. A citizen authenticates once with national digital ID and issues a scoped, revocable mandate. The agent completes the transaction under it.
Schmitz and his co-authors recommend integrating digital identity into exposed services because it lets governments enforce per-claimant limits. A mandate tied to one verified person gives a ministry that limit and gives the citizen a finished service.
We have done a version of this before.
In 2010, the Central Bank of Kenya issued its Guideline on Agent Banking. It let banks deliver services through third-party agents once the regulator approved both the agent network and each specific agent. Those agents worked for the bank. The citizen’s AI agent works for the citizen. The regulatory move transfers: approve the agent and bound what it may do.
The identity layer is mostly built. The donor-funded digital public infrastructure push left dozens of countries with digital ID. What is missing is a way for a citizen to lend that identity to an agent, and the trust to do it. Citizens who see digital ID as a surveillance tool will not delegate it to anyone.
NGOs Deliver Services and Fail Too
NGOs run clinics and enroll families in cash transfers. On October 4, I pointed a scripted headless browser at two NGO front doors.
- Marie Stopes Kenya, MSI Reproductive Choices’ Kenyan affiliate, answered with 403 Forbidden from a US cloud server. It refused both a plain request and the browser. The agent never saw the site.
- Kiva’s US loan program let the browser reach its application in two clicks with no CAPTCHA. Then the eligibility page said Kiva is accepting applications only through community-partner referral links because of high demand. The agent got through the door and found the service rationed behind it. That is the same demand-suppression move Schmitz documents in governments.
In a LinkedIn comment, Peixoto put the full RADAR run at “a few thousand dollars in tokens.” Any donor funding a government portal or an NGO service platform can require an agent test that runs past the login screen, at baseline and at completion.
Nobody ever wanted a login. They want the birth certificate.
