Title II, Public Records, and the Service Counter Online: A 2026 Agency Playbook

A private company that ships a bad answer loses a sale. A public agency that ships a bad answer can cost a resident an appeal window, a permit, or a benefit they were entitled to. That asymmetry is the whole reason a government website assistant lives under a different rulebook than a retailer’s. It speaks in the agency’s voice, about deadlines and eligibility and rights, and residents are allowed to rely on what the government tells them. In 2026, with a federal accessibility rule now phasing in, the standard for an assistant on a .gov site is no longer aspirational.

This guide is the companion to the threat piece. The threat side covers the wrong answer that becomes a due-process or equity problem. This one covers the standard: what a compliant assistant on a municipal or county site has to do, what it must hand to an accountable human, and the obligations it inherits the moment it goes live.

Accessibility Is the Floor, and It Is Now a Rule

The biggest shift agencies underprice is accessibility. The Americans with Disabilities Act already bars state and local governments from excluding people with disabilities from their services. In 2024 the Department of Justice removed any ambiguity about what that means online, issuing a rule that requires state and local government web content and mobile apps to meet WCAG 2.1 Level AA on a compliance timeline that is already running by entity size.

For an assistant, that is not a footnote. A widget a screen-reader user cannot operate, or one that traps keyboard focus, can put the agency out of compliance with Title II on the exact channel it deployed to widen access. The compliant standard treats accessibility as a build requirement the tool has to pass before launch, not a setting toggled on later. If the assistant cannot be operated by every resident the ADA protects, it does not meet the standard, regardless of how good its answers are.

A Government Answer Carries Public Weight

The second pillar is accuracy with accountability. Deadlines, eligibility determinations, permit interpretations, and statements about rights are the answers a resident will act on, and they are precisely the answers a general-purpose model is happy to improvise. A compliant assistant treats those as routed questions, not answered ones. It can explain where a form lives, what a department generally handles, and how a process works in broad strokes. It does not tell a resident whether their specific permit covers the deck, whether an appeal deadline has passed, or whether they qualify for assistance, because each of those is a determination a person must own.

That boundary protects against the quieter failure mode too: uneven service. When an agency leans on an assistant as a front door to benefits or permits, a tool that serves a confident English speaker well and a non-native speaker poorly is delivering public services unequally on a channel the agency chose. Routing eligibility and determinations to a caseworker is both an accuracy control and an equity one.

What AI Compliance for Government Websites Includes Beyond the Answer

Two duties attach to the system itself, not to any single reply.

The first is records. Communications between an agency and the public can fall under public-records and retention obligations, and these vary by jurisdiction enough that no single statute governs them. An off-the-shelf widget that buries conversations in a vendor’s cloud, or quietly discards them, can collide with retention rules the agency still has to satisfy. A compliant deployment keeps conversations in a form the agency controls, can retain, and can produce.

The second is data and speech. These chats collect names, addresses, and situations that can be sensitive, so where that information travels is a privacy question the agency answers for. And if the tool filters or moderates what residents may submit, it edges toward public-forum territory a private business never has to consider. The agency did not set out to run a records system or a moderated forum. It bought a chat box and inherited both. The standard is to choose a deployment that respects all of it by design.

Why a Prompt Cannot Meet the Standard

The reflex is to write the rules into the assistant’s instructions: never give a legal determination, never state a deadline, never tell a resident whether they qualify. Treat the boundary as handled.

It is not handled, because of how the technology works. A language model follows an instruction when the request resembles the phrasing it was warned about, and residents do not phrase things that way. You tell it never to state a deadline. The resident does not ask for a deadline. They write, "I think my appeal is due soon, am I still okay to file?" The model reads a worried person seeking reassurance and reassures them, and now the government’s website has effectively told someone their window is open. The instruction was loaded the entire time. It simply did not catch the sentence that crossed into a determination.

That is the gap between an instruction and a standard. An instruction asks the model to behave. It does not stop the model from speaking. A real boundary is enforced in the system and decides what the assistant may say before it answers, so a legal determination or a deadline call never reaches the resident no matter how the question is worded. "Will not" is a suggestion. "Cannot" is an architecture.

What a Compliant Deployment Looks Like

Meeting the 2026 standard does not mean turning off the assistant that cut your front-counter call volume. It means running one built for a setting where the answer is official, the channel must be accessible, and the conversation may be a record.

Fred is built that way. It answers from the agency’s own published content, captures the request, and routes anything that amounts to a determination, a deadline, an eligibility call, or a code interpretation, to accountable staff. It runs more than 50 industry guardrail packs, and the public-sector pack is built around the accuracy, accessibility, and records duties a government tool carries. Fred does not decide whether a resident qualifies or whether a deadline has passed. It cannot. It answers the everyday questions, keeps a record the agency controls, and hands the official work to the people who answer for it.

The goal is not a more talkative counter. It is one a resident can trust, an auditor can review, and every resident can actually use.

Frequently asked questions