Home · Industries · Government
Industries · Government

Claude, working on the backlog your agency can't hire its way out of.

Anthropic sells Claude to state and local agencies directly and through the cooperative vehicles your procurement office already uses. What comes after is implementation: getting Claude into the eligibility file, the records request, and the grant report, with the determination still sitting with your staff. That is our half of it.

Book a call See workflows
The problem

The work a public agency doesn't get to defer.

Most states have an AI pilot running somewhere. Very few have a workflow a program depends on.

Eligibility casework
Verifying documents, interpreting policy that shifts at the federal, state, and county level at once, and working a queue that refills faster than it clears. The volume is set by statute, so it is not something the agency gets to manage down.
Public records requests
Search, review, redaction, and a statutory clock running the whole time. One broad request can absorb a records officer for weeks, and the deadline does not move just because the office is short-staffed.
Grant applications and reporting
Chasing the outside funding a program depends on, then reporting against it in whatever format each funder requires. Deadlines that fall to someone who already has a full-time job.
Procurement documents
Writing the solicitation, building evaluation criteria that match it, and scoring what comes back. Months of a program manager's year spent on paperwork before a vendor starts work.
Resident and legislative correspondence
Legislative inquiries, complaints, and questions arriving in more languages than the office staffs for. Response time is the part of agency performance the public notices.
Systems nobody left documentation for
Eligibility running on COBOL, permitting on aging .NET, a tax system whose institutional knowledge retired with the person who built it. The modernization everyone agrees on and nobody can schedule.
Where we fit

Anthropic sells Claude to the public sector. We configure it for your agency.

The buying side is further along than most agencies assume. Anthropic sells to state and local government directly and through its distributor on state and local cooperative vehicles, so in a number of states Claude is already available on a contract your procurement office recognizes. For most state and local work Anthropic's own public sector guidance points buyers to Claude Enterprise, which carries a third-party NIST attestation covering controlled unclassified information. If your state has a vehicle, use it. Access attaches to your agency through Anthropic, not through us. We do not sell, broker, or discount it.

A signed contract and a working system are not the same thing. An agency that signs one gets seats, a training offer, and a capable tool on every desk, which still leaves the deputy secretary with the same queue as last quarter and nobody whose job it is to build the eligibility-review workflow. That is implementation, and in government the bar is higher: the workflow has to survive a public records request, an audit, a legislative question, and the staff turnover that happens between the pilot and the second budget cycle.

Claude does the assembly. Your staff make every determination that affects a resident, and we build the workflow so that is enforced rather than assumed. Outputs carry the policy section they came from, so a caseworker can check the citation before anything goes out and, months later, an auditor can see what a recommendation rested on.

What we build

Six workflows inside a public agency.

Specific workflows, configured for the programs your agency runs and the rules it runs them under.

01
Eligibility and benefits casework
Policy questions answered against the federal, state, and county rules that govern the case, with the citation attached to every answer. The caseworker gets the governing language and its source, and still makes the determination.
02
Public records requests
Search across the record set, a first pass for responsiveness, and proposed redactions with the exemption cited for each one. The records officer reviews and signs, working from a set that has already been marked up.
03
Grant applications and funder reporting
Applications drafted against the funder's stated criteria out of the program documents the agency already has, and the recurring report rebuilt in whatever format each funder wants. The program lead still owns the narrative and the numbers.
04
Procurement documents
Requirements turned into a draft solicitation in your jurisdiction's template, with evaluation criteria that follow from the requirements instead of being written afterward. Scoring stays with the evaluation committee. Where helping draft a solicitation would compromise our role in what follows, your procurement office draws that line and we work to it.
05
Resident and legislative correspondence
Legislative inquiries, complaints, and resident questions drafted from the agency's own policy and its prior responses, in plain language and in the languages residents write in. Staff review and send.
06
Systems nobody left documentation for
Reading an undocumented system and writing down how it works, generating the tests nobody wrote, and scoping a change so the risk is understood before it is scheduled. The agency can put a scope and a risk estimate in front of the budget owner.
Proof
Why us, specifically

We implement Claude for public agencies. It's the only thing we do.

Implementing Claude is the whole business. In a public agency that means the eligibility queue moves, a broad records request reaches the records officer already marked up, and every recommendation carries the policy citation it came from, so the file holds up when someone asks how a decision was made.

Shannon Advisory is a member of the Claude Partner Network.

How we work

Start with one workflow.

For most agencies, the right first step is Build: one high-leverage workflow, turned into a working system the program runs, on a fixed timeline.

Recommended entry point
Build

One specific problem, turned into a working system. We build it, test it against your real work, and hand it over running. Pricing by proposal.

See the Build page

Not sure where to start? Adopt is built for that: team training, a workflow audit, and an opportunity map that surface the workflow worth building first.

FAQ

Questions agencies ask before the first call.

We're a state agency, not a federal one. Does this apply to us?

Yes. This page is written for state and local government: a state agency, a county, a city department. Anthropic sells to state and local buyers directly and through its distributor on state and local cooperative vehicles, so in a number of states Claude is already available on a contract your procurement office recognizes. Our own work is a separate and much smaller engagement, and we are used to being put on whatever instrument your office prefers, whether that is a professional services roster, a small purchase, or a sole source justification. Claude access attaches to your agency through Anthropic. We do the implementation that comes after it.

What about security and compliance?

Anthropic runs the compliance program behind Claude, and the government authorizations attach to the environment Claude is deployed in. For most state and local work its own public sector guidance points to Claude Enterprise, which carries a third-party NIST attestation covering controlled unclassified information, and a HIPAA-ready configuration with a business associate agreement is available for programs handling protected health information. If your security standards reference FedRAMP, the Claude for Government application is authorized at FedRAMP High, as is Claude through Amazon Bedrock on GovCloud or Google Vertex with Assured Workloads. FedRAMP is a federal authorization program, so it binds a state or county buyer only if your own standards adopt it. We hold none of those authorizations ourselves and do not claim to. What we do is map the data flow before anyone writes a prompt: what stays inside the agency, what goes to Claude, and which posture your program's obligations require.

Public records law applies to us. What happens when someone requests the prompts?

Assume they are records. Where courts and records offices have taken the question up, the answer has usually been that the prompts an employee writes and the outputs they receive in the course of their duties are records made or received in the transaction of public business, though at least one state has come out the other way on exemption grounds. So we build for that: workflows that keep the prompt and the source citation attached to the work product, retention that matches the schedule your agency already follows, and outputs a records officer can produce without reconstructing what happened months later. It is far cheaper to build this in at the start than to reconstruct it after the first request arrives.

Where does Claude actually help, and where should it stay out?

It helps most in the reading and the assembly: policy research across layered federal, state, and county rules, first-pass document review, drafting against a template, and turning program data into the report a funder or a legislative committee is waiting on. It stays out of the determination itself. Anthropic's usage policy independently prohibits several uses that come up in government, including determinations about criminal justice matters such as parole or sentencing eligibility, predictive policing, and tracking a person's location, emotional state, or communications without their consent, which is where facial recognition and biometric categorization sit. Where an output goes to a resident, we build the workflow so a qualified staff member reviews and sends it, and the file records that AI helped produce it. Where a resident would work with Claude directly, they are told so at the start. We scope those limits into the workflow at the beginning.

Our procurement cycle is long. How does that affect the timeline?

Procurement usually takes longer than the implementation does. Once access is in place, a typical Build engagement runs four to six weeks from kickoff to a working system your team has been trained on. If your state or a cooperative vehicle already carries Claude, that part is much shorter. We are glad to scope the work while procurement runs in parallel, so the build can start the week the contract is signed.

Pick one workflow. Ship it in six weeks.

Bring the queue that is not moving, or the request you know is coming. We will tell you if this is the wrong tool for it.

Book a call