In-House Agentic AI for Government Services

Proposal to the United States Congress, 119th Congress

Submitted to: House Committee on Oversight and Government Reform — Subcommittee on Cybersecurity, Information Technology, and Government Innovation
Copied to: Senate Committee on Homeland Security and Governmental Affairs; House Committee on Energy and Commerce; House Committee on Small Business
Submitted by: Leslie Lanagan, citizen proposer
Date: November 20, 2026
Action requested: Hearing, GAO study, and drafting of a charter bill as set out at the end of this proposal

Why the agent on a federal benefits, health, or small-business website must answer only to the public — and two ways to build it so no outside stakeholder can steer it.

Look, a person applying for a small-business loan, health coverage, or disability benefits is not a sales lead. They are a person, in a bad week, trying to get the government to do the thing the government said it would do. And the software helping them should have no other customer. None. It cannot rest on private money, because whoever pays directs, even when nobody means to. Public funding is the point. That is the whole proposal, right there. The rest of this is just me showing my work.

Why this committee

This belongs in Oversight, because it is about who controls government itself. Not health policy, not small-business policy — who is driving the car.

The House Committee on Oversight and Government Reform is the House’s chief oversight committee, with government-wide jurisdiction over government operations, information technology, and federal procurement. Its Subcommittee on Cybersecurity, Information Technology, and Government Innovation held a hearing on "The Federal Government in the Age of Artificial Intelligence" on June 5, 2025, and Chairman James Comer described the committee as "uniquely positioned" on federal AI because of that broad jurisdiction. This proposal is written to that committee and that subcommittee first.

Its Senate counterpart is the Committee on Homeland Security and Governmental Affairs, which has carried federal IT management legislation including the Federal Information Technology Acquisition Reform Act (FITARA). Copies go to Energy and Commerce and Small Business not as the lead committees, but because two of the first services at stake — health coverage and small-business help — sit in their subject areas. The question is whether the agent that explains those programs works for the citizen or for someone else.

What this proposal is — and is not

It is a proposal for the public-facing agentic layer: the conversational agent that greets a citizen on a federal website, understands what they need, gathers information, fills and checks forms, tracks a case, and hands off to a human. It is not a proposal to ban agencies from buying software, cloud hosting, or foundation models. Tools can be bought. Direction cannot be outsourced. It is not a request to let an agent decide benefits. No agent proposed here makes an adverse determination alone. It helps, explains, prepares, and routes — a human decides. The robot does the paperwork. The human does the deciding. That is the deal.

How to read the two options. Option A builds the agent fully in house. Option B makes every major AI company a piece of it, so no one company has any clear direction over it. They are genuine alternatives, and the Committee could choose either. They share one floor, stated in my own terms: at minimum, the plug-ins need to be written in house. Whoever supplies a model, the connectors that touch citizen data, agency systems, and eligibility rules are written by the government. That part is not negotiable. Everything else is up for discussion.

Executive summary: no other customer

Here is the thing about agentic AI. It is different from a search box or a static chatbot, and the difference is the whole ballgame. An agent does things over time: it remembers what a citizen has already provided, works through a multi-step application, checks one program against another, follows up, and keeps the thread when the citizen cannot. That persistence is exactly why it is valuable — and exactly why who controls it matters. A thing that remembers everything about you should not have a second boss.

If that agent is built and run by a company that also sells advertising, insurance, lending, data services, or premium upgrades — or that is paid per lead, per conversion, per minute of engagement, or per upsell — then every answer sits inside a conflict. The conflict does not require bad faith. Nobody has to twirl a mustache. A system optimized, even slightly, for a vendor’s other business will steer, rank, omit, and nudge differently from a system optimized only for whether the citizen got the benefit, loan, or coverage they were eligible for. It will, because that is what optimization means.

Congress should therefore charter a Federal Agentic AI Service and require that the public-facing agent on federal service websites be stakeholder-free. This proposal offers two ways to meet that standard:

Option A — Fully in-house build. A career federal service builds, owns, and operates the platform, the agent logic, and the plug-ins. Outside models, if used at all, are replaceable parts behind a government wall.

Option B — Multi-vendor federation. Make every AI company a piece of it — Microsoft, Meta, Anthropic, Google, and every other major AI company — so no one company has any clear direction. Government-written plug-ins, orchestration, and routing decide which model answers, and any model can be removed without breaking the service.

The agent’s job is simple to state: make the government website more helpful — help the citizen get things done on the site, from finding the right program to finishing the form correctly the first time. A helper with that job cannot rest on private money. Public funding is therefore not a detail of this proposal; it is the foundation. Either option must meet the statutory floor below: no advertising, no referral or lead fees, no sale or secondary use of citizen data, no steering toward a private product, full audit logging, and a guaranteed route to a human. The pilots start where steering would hurt most: the Small Business Administration (SBA) and HealthCare.gov. Funding totals are left for Congressional Budget Office scoring; this proposal specifies what to fund and how — not invented totals. I am not going to make up a number and pretend it is real. That is the CBO’s job, and I am happy to let them do it.

The problem: a benefits moment is not a market moment

Consider what a citizen is doing in the moments an agent would help them. On SBA.gov, they may be trying to keep a business alive and asking which loan or disaster program they qualify for. On HealthCare.gov, they may be comparing plans after losing a job, with a pre-existing condition and a deadline. On a benefits site, they may be documenting a disability, a layoff, or a death in the family. In each case they disclose sensitive facts, under time pressure, to the one institution that is supposed to have no angle. This is not a shopping trip. Nobody is browsing. They are scared, they are tired, and they need the form to work.

Five stakeholder conflicts a vendor agent cannot design away

1. Steering. Ranking one lender, plan, or preparer above another — or presenting a private option before a public one — changes outcomes even when every sentence is technically true. Technically true is not the same as honest.

2. Engagement. A vendor paid or judged on time-on-site, return visits, or conversation length benefits when a task takes longer. The citizen benefits when it ends quickly and correctly. One of those two is the customer. Pick one.

3. Data. Conversations about health, income, disability, and business finances are commercially valuable training and profiling data. Contract language can restrict that use; ownership removes the temptation and the enforcement burden. A lock on the door is nice. Not having the thing worth stealing in the house is better.

4. Lock-in. If the agent’s prompts, plug-ins, knowledge connectors, and case history live in a vendor’s proprietary system, changing vendors means rebuilding the service — so oversight weakens with every renewal. Every year the vendor gets harder to fire, and every year the Committee gets less say. That is not a partnership. That is a hostage situation with a renewal clause.

5. Invisible change. A commercial model update can change tone, thresholds, and recommendations overnight across every agency site at once, without any public record a citizen or this Committee can inspect. The agent you approved in January is not the agent answering citizens in March, and nobody sent you a memo.

The test for any design the Committee considers: If the company behind the agent disappeared tomorrow, or doubled its price, or was acquired by a lender or insurer — could the government keep serving citizens the next morning, with the same answers and the same records? If not, the government does not control the service. The stakeholder does. And if the stakeholder does, then the stakeholder is running a federal benefits program, and nobody voted for that.

Where things stand: fragmented today, vendor-heavy — and the in-house bench was just cut

Federal agencies already report thousands of AI use cases in the public inventory required under current Office of Management and Budget (OMB) policy, including hundreds described as chatbots or virtual assistants. The pattern is piecemeal: one agency buys a chatbot, another pilots a different vendor’s assistant, a third builds a small tool in house, with no shared agent a citizen can recognize from one site to the next. In the SBA inventory, for example, the external customer chatbot for SBA.gov is listed as purchased from a vendor, while other SBA tools are listed as developed in house — the mix is decided procurement by procurement, not by any rule about stakeholder conflict. It is a yard sale, not a system.

At the same time, the government’s best-known in-house digital team is gone. 18F, the General Services Administration (GSA) unit founded in 2014 inside Technology Transformation Services that built and advised on federal digital services, was eliminated on March 1, 2025, and its staff dismissed. GSA’s retained priorities — Login.gov, SAM.gov, FedRAMP, cloud.gov, and Centers of Excellence — show that shared federal platforms can survive as public infrastructure. What did not survive is the standing team that helped agencies build instead of only buy. Any in-house proposal that does not rebuild and protect that bench is a slogan. I am not here to sell slogans.

Current policy helps, but does not settle this question. OMB Memoranda M-25-21 (use of AI) and M-25-22 (acquisition of AI), both issued April 3, 2025, require agency Chief AI Officers, governance boards, public use-case inventories, and minimum risk practices for "high-impact" AI, and they bar vendors from using non-public government data to train commercial models without explicit consent. Those are necessary guardrails for buying AI. They do not answer the prior question this proposal asks: for the citizen-facing agent itself, should the government be the builder — or, at minimum, the author of every plug-in that connects a model to a citizen’s case?

The statutory floor: what "free of stakeholders" must mean in law

A principle that cannot be audited will not survive a procurement. So here it is, in plain language, as a floor the charter bill should define — a stakeholder-free public agent — with all of the following, whichever option Congress chooses:

Requirement What it forbids / requires
No commercial steering No advertising, sponsored ranking, referral fees, lead generation, or presentation of a private product ahead of an eligible public program.
No secondary data use Citizen conversations and case data are never sold, shared for marketing, or used to train a commercial model without explicit, separate consent — codifying and strengthening the M-25-22 protection.
No conflicted pricing No payment per lead, per conversion, per enrollment, per loan, or per minute of engagement for a public-facing agent.
Government-owned record Prompts, plug-ins, decision logs, and case history are federal records, stored in government systems, portable, and inspectable by the agency, GAO, and — in their own case — the citizen.
Disclosure in plain language Every session opens by stating the citizen is talking to a government AI, what it can and cannot do, what it remembers, and how to reach a human instead.
Human determination The agent never alone denies, reduces, or terminates a benefit, loan, coverage, or service. A human makes and signs that decision, with a real appeal route.

An agent that fails any row is not stakeholder-free, whatever its vendor or host calls it. This floor is the part of the proposal that is not optional between the two options. You can pick A or B. You cannot pick neither and call it done.

Option A: fully in-house — the government builds it, owns it, and runs it

Design. Congress charters a Federal Agentic AI Service as a shared service inside the General Services Administration, rebuilding the function lost with 18F — this time in statute, not at administrative discretion. Because here is the lesson of 18F: if it can be eliminated overnight by memo, it will be. Put it in law, or do not bother building it. Career and term federal engineers, designers, accessibility specialists, and agency detailees build one shared platform: identity through Login.gov, search and knowledge through agency-verified sources, form-filling and case-tracking components, audit logging, and a human handoff queue. Each agency adopts the same agent, configured with its own programs and rules, so a citizen learns it once. Learn it once. Use it everywhere. That is the whole pitch.

What in-house buys

Single loyalty. Staff are federal employees with no other product, advertiser, or shareholder. Their performance measure is task completion and correctness for citizens. Continuity. Code, plug-ins, and institutional knowledge stay when administrations and contracts change — if the charter protects the service from overnight elimination. Inspectability. Prompts, ranking rules, and changes are versioned federal records. The Committee can read exactly why the agent answered as it did on a given date. Cost discipline over time. The government pays salaries and infrastructure once for a platform every agency reuses, instead of paying each agency’s vendor to rebuild the same agent.

Honest costs of this option

It is slower to stand up, it must compete for scarce AI engineers, and it risks falling behind the best commercial models if it tries to build its own foundation model — which it should not. It may run vetted models behind the government wall, but it does not need to train a frontier model to control its own direction. If Congress judges that hiring and pace risk too high, that is the case for Option B — not for handing direction to a single vendor. Slow is a real problem. It is not a reason to hand the keys to one company and hope for the best.

Option B: make every company a piece of it, so no company directs it

In my own framing: the other option is to make every AI company a piece of it so that no one company has any clear direction. An in-house model could be built off of Microsoft, Meta, and Anthropic to spread the wealth around — or, at minimum, the plug-ins need to be written in house. I mean that. Spread the wealth around. If everybody owns a slice, nobody owns the pie. This proposal takes that option seriously, and completes the list as I directed: Microsoft, Meta, Anthropic, Google, and every other major AI company that can meet federal security and accessibility requirements — not just the first three named. Do not forget Google. Do not forget anybody. The whole point is everybody.

How the federation works

Many models, one government agent. The citizen meets a single federal agent. Behind it, a government-owned router sends each task — plain-language explanation, document reading, form checking, translation — to whichever approved model fits best, and can cross-check a high-stakes answer with a second model from a different company. No company sees the whole citizen. Tasks are split and minimized so no single vendor receives a complete case file. No vendor’s model is the sole path to any service. Spread the wealth, spread the risk. Contracts are deliberately plural and time-limited. Market share rotates on performance — accuracy, accessibility, cost — never on incumbency. If any company changes its terms, prices, or politics, its piece is removed and traffic shifts the same week. Plug-ins stay in house. Every plug-in — the code that connects a model to SBA loan rules, HealthCare.gov plan data, an agency case system, or a citizen’s documents — is written in house by federal developers. A model without a government plug-in cannot touch a citizen’s case. Full stop.

Option B: vendor-neutral and plug-in-agnostic, by design

The principle behind Option B is freedom on the government side of the wall. A developer in house at the government should be able to use whatever plug-in from whatever tool they want. The government side stays vendor-neutral and plug-in-agnostic: plug-ins follow an open federal interface standard, models are interchangeable behind it, and no single company’s toolset is ever the only way in. A new company — including a small or open-source entrant — qualifies by passing the same published tests as an incumbent, not by winning a sole-source platform contract. If the little guy can pass the test, the little guy gets to play. That is how you know it is a standard and not a velvet rope.

Question Option A: In-house Option B: Federation
Who directs? Federal service, directly Federal router and plug-ins; vendors supply interchangeable parts
Speed to launch Slower; hiring first Faster; uses existing models
Model quality Depends on vetted models it runs Best available per task, cross-checked
Main risk Hiring and pace Complexity; quiet drift toward one dominant vendor — prevented by share caps and exit tests
Exit test Service is the government’s own Remove any one vendor in a week with no citizen-facing break — tested live, twice a year

Guardrails that keep Option B honest: a cap on any one company’s share of tasks; a ban on exclusive or sole-path arrangements; twice-yearly vendor-removal drills with results reported to this Committee; and the same statutory floor above. Without those, a federation collapses into a single-vendor service with extra logos. With them, it is a real alternative: the government keeps direction precisely because every company is only a piece. A federation without guardrails is just a monopoly with better branding. Do not let that happen on your watch.

Recommendation. I present both options for the Committee’s judgment. If forced to rank them, start with Option B’s federation for the two pilots — it can launch sooner — on Option A’s institutional base: a chartered in-house service that owns the router, the plug-ins, and the record. That pairing is not a compromise that blurs the line; it enforces it. Direction lives in house either way.

Why helpfulness is accessibility: the agent holds the threads so the citizen does not have to

A government form is rarely one form. It is a chain — eligibility in one place, documents in another, a deadline in a third — and the citizen is expected to hold every thread at once. For people with cognitive disabilities, including autism and ADHD, and for anyone in crisis, that thread-holding is the barrier, not intelligence or effort. I know this because I live it. A well-built agent distributes that cognition: it remembers what was already provided, says what happens next, keeps the place between visits, and does so without the guilt or resentment that can build when a person must lean on another person for the same help. That is cognitive relief, and on a government website it is an accessibility function. It is not a nice extra. It is the ramp.

(I laid out the accommodation argument in The Benjamin Franklin Corollary.)

The charter should therefore require, beyond Section 508 compliance and the accessibility, consistency, and user-centered standards of the 21st Century Integrated Digital Experience Act (Public Law 115-336): plain language by default, voice and text equally, the ability to pause and resume without penalty, a summary the citizen can save or print, and a human alternative that is never slower or worse for having chosen it. An agent that only helps people who already navigate websites easily has failed its purpose. If it only works for the people who did not need help, it is not help. It is a mirror.

Pilots — prove helpfulness where it counts

Pilot 1 — SBA.gov: help a small-business owner find the right loan, grant, or disaster program, assemble documents, and complete the application. Success is measured by completed, correct applications and time saved — never by loans steered to any lender.

Pilot 2 — HealthCare.gov: help an applicant compare plans on total cost and their own doctors and prescriptions, meet the enrollment deadline, and report life changes. Success is correct enrollment, not enrollment in any particular plan.

Each pilot publishes its measures before launch, runs for a defined period, and stops automatically if accuracy or complaint thresholds are breached. Only then does the service extend to cross-benefits discovery and to the Social Security Administration and Department of Veterans Affairs. Prove it small, or do not scale it at all. That is the rule.

The practical payoff: help in the browser cuts the support burden everywhere else

Today, help arrives after failure. A citizen gets stuck on a website, abandons the application, then joins a phone queue, visits an office in person, asks a family member or paid preparer, or files the wrong form and starts over — and the agency pays for every one of those contacts while the citizen pays in time. An agent working with the citizen in the browser itself reverses that order: it walks them through the site right there — highlighting the next step, catching a missing document before submission, explaining a term on the page it appears on. Help before the failure, not after it. What a concept.

Fewer phone calls and office visits for questions the website could have answered in context. Fewer abandoned and incorrect applications, so agency staff spend their time deciding cases instead of repairing them. Fewer repeated requests: the agent remembers the thread, so the citizen does not re-explain their case at every contact.

Plainly: the alternative to the agent is the phone queue, and the phone queue is failing people. That is what "support" means today for a citizen stuck on a government website — a long hold, being bounced between departments, repeating the same story to each new person, and starting again when the call drops. It is annoying for everyone. For many disabled people it is the barrier itself. I am terrified of phone calls; a system whose only real help requires one has, for people like me, no help at all. The in-browser agent matters because it meets the citizen before the queue — on the page, in the moment, in writing. You want to cut support costs? Stop manufacturing the need for support.

That payoff is also why the funding model matters. Support savings belong to the public that funded the agent — they must not be converted into a vendor’s per-contact or per-conversion revenue. The savings are the public’s. Keep them that way.

Workforce, oversight, and public funding

Workforce. The charter must authorize direct hiring for engineers, designers, and accessibility specialists, mix career staff with term appointments and agency detailees, and protect the service’s existence in statute so it cannot be eliminated overnight as 18F was. Oversight. The service reports to this Committee every six months; publishes its use-case inventory, system cards, accuracy and complaint data, and (under Option B) vendor task shares and removal-drill results; submits to GAO audit and pre-deployment testing, impact assessment, and continuous monitoring for high-impact uses under OMB M-25-21; and logs every material model or prompt change in a public changelog. Funding. The platform is funded by direct appropriation — public money for a public helper. Free is the most expensive price there is: if the public is not paying, the citizen is, in ways they cannot see — investment, sponsorship, advertising, lead fees, or a "free" vendor service paid for in data and direction.

Risks: the risks argue for control, not for outsourcing it

Risk Mitigation in this proposal
Wrong or invented answers Answers grounded only in agency-verified sources with citations shown; cross-checking for high-stakes steps; published accuracy thresholds and automatic pilot stop rules.
Bias in guidance Pre-deployment impact assessment, testing across languages and disabilities, public results, human determination of any adverse outcome.
Security / privacy breach Government hosting and records, data minimization, no vendor training on citizen data, FedRAMP-authorized infrastructure, audit logs.
Quiet vendor capture (Option B) Task-share caps, no sole-path design, plug-in-agnostic standard, twice-yearly live removal drills reported to the Committee.

Eight actions requested of the Committee

  1. Hold a hearing in the Subcommittee on Cybersecurity, Information Technology, and Government Innovation on stakeholder conflict in citizen-facing federal AI.
  2. Request a GAO study of current vendor-built citizen-facing agents: pricing structures, data terms, and exit costs.
  3. Direct CBO scoring of a chartered Federal Agentic AI Service under both Option A and Option B.
  4. Draft a charter bill codifying the statutory floor: no steering, no secondary data use, no conflicted pricing, government-owned record, human determination.
  5. In that bill, require government-written plug-ins and a vendor-neutral, plug-in-agnostic interface standard.
  6. Authorize the SBA and HealthCare.gov pilots with pre-published measures and automatic stop rules.
  7. Require semiannual reporting, a public changelog, and GAO audit access.
  8. Protect the service in statute against elimination without notice to Congress.

Closing

Put a genuinely helpful agent on the government’s own websites, in the browser beside the citizen, so people can get things done — and pay for it publicly, because a helper paid for privately will, in the end, help whoever paid. Build it in house, or build it from every company at once with the plug-ins written in house. Do not let any single stakeholder own the doorway to public services. The doorway belongs to the public. That is the whole thing.

Respectfully submitted,
Leslie Lanagan, citizen proposer
November 20, 2026

Sources staff can verify: House Oversight hearing, "The Federal Government in the Age of Artificial Intelligence," June 5, 2025; OMB M-25-21 and M-25-22, April 3, 2025; 21st Century IDEA Act, Public Law 115-336; federal AI inventory (SBA); reporting on the elimination of 18F, March 1, 2025. No funding total is asserted pending CBO scoring.