For most of the internet's life, a government website was a page problem. You wrote the copy, you built the navigation, you tested the forms, and you shipped something a citizen could scroll through to find a permit, a benefit, a phone number. California just changed that math. The state has launched an AI assistant meant to help residents navigate state services, and the underlying model layer is getting more capable by the week: Anthropic has added its Fable 5.1 model to Claude for Government, putting a newer reasoning engine directly into the hands of agencies that used to procure static portals.
Nobody redesigned a homepage here. The whole deliverable moved: pages designed once got replaced by prompts tuned continuously against whatever a resident happens to type into a box. The information architecture doesn't disappear, but it stops being the thing a citizen sees. It becomes the thing the model is instructed to consult before it answers.
The Deliverable Changes
When an agency buys a chatbot instead of commissioning a portal, the vendor relationship shifts from "build this and hand it over" to "keep tuning this indefinitely." Content strategy doesn't vanish, it gets absorbed into prompt design: the same decisions about tone, hierarchy, and what to say first now live inside a system prompt instead of a homepage. The people writing those instructions are doing a version of the same job a UX writer used to do, except the output is invisible until a resident asks the right question, and there's no way to proofread every possible conversation before launch the way you could proofread a page.
That's a real loss of a certain kind of quality control, and a real gain in responsiveness. A form can only anticipate the questions its designer thought to ask. A tuned model can, in principle, meet a resident somewhere closer to their actual confusion. Whether it does that reliably is the whole question agencies are now buying into without much of a track record to check it against.
Who Answers For the Answer
Security officials inside government are already saying the industry side of this bargain isn't holding up its end. The nuclear agency's top IT official has put it plainly: industry has not done enough on security. That's a warning worth sitting with before anyone hands a conversational front end the job of telling residents what benefits they qualify for or what a legal deadline is. A wrong page can be corrected with an edit. A wrong answer, given confidently in a chat window and then closed by the resident before anyone reviews the transcript, may never surface as an error at all.
There's a useful comparison in how the FTC handles a much more sensitive front door problem: its consumer guidance on what to do if your intimate image is shared online without consent is written as a precise, sequenced script, because getting the order of steps wrong has real consequences for someone in crisis. That's the standard a state chatbot is implicitly being asked to meet the moment it starts answering questions about benefits, deadlines, or legal rights. A page can carry a "last reviewed" date and a named author. A model's answer, generated fresh each time from a tuned prompt, doesn't come with that same paper trail unless someone builds one deliberately.
The Front Door Needs a Foundation
None of this works if the resident can't get online. Researchers estimate about 1 million BEAD-eligible locations still remain unserved, and Comcast is expecting more broadband losses this quarter than it saw a year ago. Washington State is moving along on its own BEAD grants at the same time other agencies are standing up AI front doors, which means two procurement tracks are running in parallel: one laying fiber, one tuning prompts, both aimed at the same resident, neither one useful without the other.
This isn't a new tension for public infrastructure. America once tried putting telephones inside the post office, on the theory that if the building already served as the place people came for official business, it might as well be the place they made a call too. It didn't stick, but the instinct behind it is the same instinct behind an AI assistant bolted onto a state services site: put the newest communication technology at the point where citizens already show up, rather than asking them to learn a new one. The lesson from the post office experiment is that the front door only works if the infrastructure behind it is actually reliable, not just conveniently located.
What Agencies Are Actually Buying
Even the FCC's own procedural habits show how granular this gets once you formalize a communication channel: its Telecommunications Relay Service rules specify an ASCII format requirement, a detail that exists because someone had to legislate exactly how text gets transmitted to make a service usable for people who rely on it. A chatbot procurement is going to need the same level of specificity eventually: what format answers are returned in, what happens when the model can't find an authoritative source, how a transcript gets logged, who signs off on a change to the underlying prompt. Right now, most of that is being worked out quietly, agency by agency, without the kind of public scrutiny a redesigned website would draw simply by looking different.
Buying a chatbot doesn't just change the interface. It changes who does the work, what gets reviewed, and what a citizen can point to later when the answer they got was wrong. Agencies that treat this as a procurement swap, portal out, chatbot in, are skipping the part where they decide who's accountable for a conversation nobody wrote down in advance.