In August 2025, GSA’s OneGov agreement with Anthropic priced Claude for Enterprise and Claude for Government at a nominal $1 for up to a year across all three branches, and GSA’s AI training series remains free to anyone with a .gov or .mil address. Access stopped being what a public-sector team is short of.
What it lacks is a way to buy four hours of hands-on work tied to its own files, under an instrument that clears review this fiscal year. AI training for government teams is not short of catalogs. It is short of a path to purchase, so this is about which instrument fits, what the free material covers, and what to watch afterward.
Access got cheap, capability did not
Licenses stopped being the obstacle in the public sector before they stopped being one anywhere else. The OneGov deals pushed per-seat cost toward zero for a year, and GSA stood up a no-cost sandbox, USAi, for teams to try the tools first. Those offers are term-limited and get renegotiated, which is the point: license pricing is now a footnote, and the capability gap is the line item. An agency can be fully licensed and fully untrained at once, the gap behind the case for treating enablement as work rather than as a software buy.
The consequence shows up in oversight reporting. In a March 2026 review of AI at the IRS, GAO found that “IRS officials said they had not identified skills needed to support AI or developed a plan to address the skills gaps,” and made eight recommendations. That is an agency with the tools, the mandate, and no answer on who can use them.
So the live question in a department is narrower than the vendor catalog suggests. Not whether to adopt, which staff settled months ago on their own phones, but what each function needs to do, and what may never be pasted into a chat window.
Use the free material, then say where it stops
The free federal material is good and you should use it. GSA’s series runs three tracks, acquisitions, leadership and policy, and technical, built with Stanford, George Washington University Law School, and Princeton. Registration costs nothing for any employee with a government address. OPM publishes ten SCORM module packages an agency can load into its own learning system, also free, with the honest note on the page that OPM “does not offer technical support” for them.
That material does two things well. It gives a department shared language, so the people arguing about risk argue about the same thing. And the acquisitions track speaks to a duty that already exists: the AI Training Act, Public Law 117-207, signed October 17, 2022, obliges OMB to run an AI training program for the acquisition workforce.
Where it stops is at your documents. GSA reported “over 12,000 registrants, with an average 94% satisfaction rating” for its series, and satisfaction measures the session. No course built for all of government opens the permit file, the grant report, or the correspondence queue one department works in daily, which is where fluency either transfers or fades.
One agency, several functions, several curricula, a split we work through in the role-based enablement playbook. Treat the free coursework as the floor and buy the part that touches your own work.
Four ways to buy a session without an RFP
Here is the detail that changes the conversation: a market-rate workshop fits on a purchase card. FAR 2.101 sets the micro-purchase threshold at $15,000 and the simplified acquisition threshold at $350,000, and private sessions run roughly $2,000 to $15,000 at market rates. Most of that band sits below the ceiling where competitive quotes and most clauses begin.
| Instrument | Ceiling | What it asks of you | Best fit |
|---|---|---|---|
| Purchase card, micro-purchase | $15,000 | A reasonable-price determination and a record | One department, one session, this quarter |
| Simplified acquisition | Above micro-purchase to $350,000 | Quotes from an adequate number of sources | A program across several functions |
| Schedule order or blanket purchase agreement | Schedule terms, FAR 8.4 procedures | Ordering from a holder of the training line | Recurring need, federal buyers |
| Federal grant funds | 2 CFR 200.320 micro-purchase, self-certified up to $50,000 | Written justification, audit-ready file | Grant-funded programs and subrecipients |
Three cautions on that table. FAR 2.101 also holds a $2,500 ceiling for services subject to the Service Contract Labor Standards; instruction by professional staff usually sits outside that statute, but the call belongs to your contracting officer rather than a vendor’s proposal.
Grant-funded buyers get more room than they think. Under 2 CFR 200.320 a recipient may self-certify a micro-purchase threshold up to $50,000 a year, with documentation showing low-risk auditee status, an annual internal risk assessment, or that the higher figure does not conflict with state law. What a two-person development team can cover runs the same math on the grantee side.
State, local, and district buyers cannot shortcut through a GSA schedule. Cooperative Purchasing opens only the information technology and the security and protection categories to non-federal buyers, and professional training is in neither. The route is your own small-purchase rule, a term contract, or a cooperative agreement you already hold.
Write the quote request so it survives review
A one-page request beats a scope document nobody can price. Five lines carry it.
- The task list, from your own files. The grant report, the staff report, the records-request acknowledgment, the correspondence draft, the aged-queue triage.
- Fixed price, all in. Travel, materials, and follow-up in one number, so the price determination takes a paragraph.
- A written data rule as a deliverable. Two columns, which account and which document category, signed off by whoever owns records.
- Tool neutrality. The session teaches whichever assistant you are cleared to use. A quote that needs a new platform is a software purchase wearing a training label.
- Named measures and a date to read them. This is the line that protects you at the next budget hearing.
Leave out seat counts for a platform nobody will log into.
Bring your contracting officer in early. Thanks to the AI Training Act, acquisition staff have had a path to this material longer than most program offices have, and a warranted officer who knows the vocabulary will name the right instrument in ten minutes.
Settle the data rule before anyone prompts
Decide what may be typed where before any hands-on work, because here the downside is legal rather than operational. “Be careful with sensitive data” is not a rule anyone can act on. A rule is a two-column list taped next to the monitor: which account, which document category.
Two facts shape the columns.
An authorization is narrower than a logo on a vendor page. Google’s own deployment guidance for Gemini for Government says resources must sit inside an Assured Workloads folder, and that “the authorized data stores for FedRAMP High and IL4 are Cloud Storage buckets and BigQuery datasets.” It instructs users plainly not to connect anything else. So check the FedRAMP Marketplace listing and your own tenant settings, not the marketing page, and remember that a personal login an employee pays for sits outside all of it.
Records law turns on behavior. NARA’s guidance on applying the Federal Records Act to AI materials, issued August 21, 2026, treats prompts and outputs as records once they are captured in an agency system and used for official purposes. An account history counts when it is used to conduct official business, while material left in a third-party tool that goes nowhere is read more narrowly. The staff-facing version is one sentence: if you save it, send it, or rely on it, treat it as a record.
Then run the drill. Each person brings twenty real items from their own week, a constituent email, a draft staff report, a pre-decisional memo, and sorts them into three piles: goes in the cleared account as-is, goes in once identifiers come out, never goes in. Ten minutes of argument about the edge cases teaches more than an hour of policy reading, and in the sessions we run that argument is where the rule gets written. The gap between a written policy and what people do is the subject of our guide to safe use.
One test shows whether it stuck. Ask three staff separately what may be pasted into which account. Three answers mean you have a policy document, not a rule.
What the options cost once you price the hours
The invoice is the small number. Staff hours are the large one, and no course page prices them for you. BLS puts employer compensation cost for state and local government workers at $66.45 per hour worked in June 2026, $25.78 of it benefits. A 25-person department costs about $1,661 an hour to put in a room.
| Option | Published price | Hours each | Price plus the hours it burns |
|---|---|---|---|
| GSA series and OPM modules | $0 | 2 to 7, if anyone finishes | $3,300 to $11,600 of staff time for 25 |
| Graduate School USA leadership course | $675 a seat | 4 live | about $940 for one manager |
| GovTra instructor-led course | $1,630 online, $2,230 classroom | 7 | $2,100 to $2,700 a seat |
| Coursera Teams catalog | $399 a license a year, 2 to 499 | 10, if anyone finishes | about $26,600 in year one for 25 |
| Private session on your own documents | $2,000 to $15,000 at market | 3 to 4 | the fee plus $5,000 to $6,600 of staff time |
Two readings of that table. The free options are not free, and a self-paced rollout nobody finishes is the costliest line per unit of capability gained. Catalog subscriptions are the worst fit for a 25-person department, because the per-seat math only works at high completion, the one thing nobody measures.
Some vendors will not give you a figure at all. DataCamp’s government page offers a demo request and learner-count bands, no price, which is useless to a buyer testing a number against a threshold. Ask for a fixed quote in the first email, as the buy-versus-build playbook argues.
Four measures to read in 90 days
Pick the measures before the session, because afterward everyone has a view to defend. All four live in systems you already run.
- Cycle time on one recurring document. Same type, same reviewer: a quarterly grant report, a staff report, a records-request response. Time ten before and ten after.
- Aging on one queue. Permit intake, records requests, invoice processing, whichever queue has a published target. Queue discipline moves first, as it does for operations teams elsewhere.
- First-pass acceptance. How often a drafted item clears review without rework. This tests the review rule rather than the tool, which is why supervisors trust it.
- The recall test. Three staff, asked separately, stating the same data rule.
Login counts and satisfaction scores are not on the list. A 94% rating tells you the sessions were well run, not that a grants analyst drafts faster, and turning a plausible gain into one a budget office accepts is what our measurement guide is for.
What to buy first for a 25-person department
Four hours, split by function, on the department’s own documents, with a written data rule as the deliverable. That is the whole first purchase in AI enablement for a department this size. At market rates it lands under the micro-purchase threshold: one purchase card and a file memo rather than a solicitation.
Run two groups of about twelve rather than one room of 25. The functions fail differently: public-facing staff fail in public and at once, back-office staff fail in private and six weeks later, and a combined room drifts toward whichever group talks more. Bring everyone together for the last 45 minutes to settle the data rule and name the owner. Send a division manager to a priced public course afterward so somebody internal can field week-six questions, and for the people who approve the next tranche of spend use the shorter format in programs for leadership teams.
Then leave it alone for 90 days and read the measures. If cycle time has not moved and nobody can state the rule, the problem is the review step rather than the training, and that is an hour’s work to fix.
Frequently asked questions
Is there free AI training for government employees?
There is, and it is worth using. GSA’s AI training series is free for anyone with a .gov or .mil email address and runs acquisitions, leadership and policy, and technical tracks. OPM publishes ten SCORM modules agencies can load into their own learning systems, as-is and without support. Neither works on your own files, so treat both as the floor.
What is the easiest way to buy a one-day workshop for a public-sector team?
A purchase card under the micro-purchase threshold, which FAR 2.101 sets at $15,000. A private session at market rates usually fits, so the file needs a reasonable-price determination and a record of the buy rather than a solicitation. Ask the provider for one fixed, all-inclusive number.
Can state and local agencies buy training off a GSA schedule?
No. GSA’s Cooperative Purchasing program opens only the information technology and the security and protection categories to state, local, and tribal buyers, and professional training is in neither. Use your own small-purchase rule, a term contract, or a cooperative agreement you already hold.
How much does AI training for a government team cost?
Private sessions run roughly $2,000 to $15,000 at market rates depending on length and group size. Published per-seat options sit between $675 for four live hours and about $2,230 for a seven-hour classroom course, and catalog subscriptions about $400 a person a year. Add staff time at $66.45 an hour, usually the larger figure.
Does a workshop satisfy the AI Training Act?
Not on its own, and do not let a vendor imply otherwise. Public Law 117-207 puts the duty on OMB to provide a training program for the acquisition workforce, and GSA’s acquisitions track is how agencies meet it. A session on your own documents sits alongside that duty.
Which assistant can our team use on agency data?
Whichever one your authorization and your tenant settings cover, which is narrower than the vendor page suggests. Google’s guidance for Gemini for Government, for one, requires resources inside an Assured Workloads folder and limits authorized data stores to Cloud Storage buckets and BigQuery datasets. Check the FedRAMP Marketplace listing and your own configuration first.
Are AI prompts public records?
Sometimes, and the test is behavioral. NARA’s August 2026 guidance treats prompts and outputs as federal records once they are captured in an agency system and used for official purposes, and reads material left in a third-party tool more narrowly. State and local records law varies, so train the simpler rule: if you save it, send it, or rely on it, it is a record.
Who should own this inside the agency?
A named deputy or division manager with hours in their workload, not a committee. Programs stall when ownership is implied. The job is narrow: keep the prompt library current, answer week-six questions, bring the four measures to the 90-day review.
Key takeaways
- Access got cheap. GSA’s 2025 OneGov agreement put Claude for Enterprise and Claude for Government at a nominal $1 for up to a year across all three branches, and term-limited or not, license pricing is no longer the constraint. The gap between what staff may use and what they can do is.
- The free federal material is a real floor. GSA’s series and OPM’s modules cost nothing and build shared language, and neither touches the permit file or the grant report your team works in.
- FAR 2.101 sets the micro-purchase threshold at $15,000, so a market-rate session fits on a purchase card with a price determination and a file memo. Grant-funded buyers may self-certify up to $50,000 a year under 2 CFR 200.320, and state and local buyers cannot use GSA Cooperative Purchasing for training.
- Price the hours. At $66.45 an hour of employer cost, a 10-hour self-paced course is a $16,600 payroll item for 25 people before tuition.
- Measure in your own systems: cycle time on one recurring document, aging on one queue, first-pass acceptance, and three staff stating the same data rule.
If you are deciding what to put in front of a public-sector team this quarter, book a free AI readiness call. We will go through your document categories, the instrument that fits your ceiling, and the four measures, and say which function to start with.
Arthur Wandzel
