The Short Answer
Government AI procurement is shifting from choosing a vendor to owning the architecture: the code, the model layer, the data and the deployment. OMB's M-25-22 already asks contracts for portability and lock-in protection, but its code-rights term covers code produced under the contract. On ibl.ai you own all the code and the data, so the platform itself stays with the agency.
The rules are in OMB Memorandum M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government, signed April 3, 2025. The evidence for why they matter is in a GAO review of restrictive vendor licensing published November 13, 2024.
This post sets out what the memo asks for, what it leaves to the agency, and a clause table for each layer of an AI system.
Why is government AI procurement moving from vendor selection to architecture ownership?
Government AI procurement is moving to architecture ownership because AI stopped being a feature and became shared infrastructure. A chatbot on one website can be swapped out. Agents that read case files, answer the public and support decisions across departments cannot.
When that much of an agency runs through one system, the vendor relationship becomes a dependency.
M-25-22 names the risk in its first section: agencies must watch "vendor sourcing, data portability, and long-term interoperability to avoid significant and costly dependencies on a single vendor."
Government also carries obligations that outlast any contract. Under FOIA, a "record" includes information maintained for an agency by an entity under government contract for records management, per 5 U.S.C. § 552(f)(2)(B).
So the evaluation question changes. It is less "which vendor has the best model this year" and more "what does the agency still hold, and still run, on the day the contract ends?"
What does OMB M-25-22 require of federal AI contracts?
M-25-22 applies to AI contracts awarded under solicitations issued 180 days or more after April 3, 2025, and to options and extensions exercised after that date. It does not apply to AI acquired as a component of a National Security System.
For a buyer worried about lock-in, five of its requirements matter most:
- Use of government data. Contracts must "permanently prohibit" using non-public agency inputs and outputs to further train publicly or commercially available AI, absent explicit agency consent.
- Vendor lock-in protections. These "include requirements for vendor knowledge transfers, data and model portability, providing agencies with rights to code and models produced in performance of a contract, and transparency in licensing and pricing."
- Availability over time. Agency processes should ensure "components necessary to operate and monitor the AI system or service remain available" to the agency for as long as necessary.
- Independent testing. The agency evaluates with data it defines, which the vendor should not be able to access, and contracts may not stop the agency from internally disclosing test results.
- Contract closeout. Once an agency decides not to extend, it should agree the format and usability of its data and plan any transfer of "data or other derived assets."
Two further points are easy to miss.
Every AI system operated by or for an agency still needs an authorization to operate before deployment. And the memo's footnote 25 lists open-source licenses for a vendor's models, systems and datasets as an evaluation criterion agencies can use.
We have covered the memo's provisions as a vendor checklist in an earlier post on federal AI acquisition rules. This post goes one level down, into the architecture those provisions imply.
What does architecture ownership mean in a government AI contract?
Architecture ownership in a government AI contract means the agency, not the vendor, controls four layers: the code, the model layer, the data and the deployment environment. A fifth, the audit record, follows from the other four.
M-25-22 speaks to every layer, but it does not settle every one.
The gap that matters most is in the first row: the memo's code-rights term names code "produced in performance of a contract." A commercial platform the vendor built before the contract is a licensing question the agency has to raise itself.
The table below is the one we would want in hand before a contract is signed. It pairs each layer with what the memo asks for, what it leaves open, and the clause that closes the gap.
| Layer | What M-25-22 asks for | What it leaves to the agency | Clause that closes the gap |
|---|---|---|---|
| Code | Rights to code and models produced in performance of the contract; IP terms scoped "to avoid vendor lock-in" | The vendor's pre-existing platform code | Source code of the deployed platform delivered under a perpetual license, with build and deployment documentation |
| Model layer | Data and model portability; well-defined APIs that promote interoperability (footnote 25) | Which model providers the system can use at all | Any commercial or open-weight model can be configured by the agency, and the agency's evaluation set runs against each |
| Data | Permanent bar on training with non-public agency data absent consent; data collected by a vendor only when reasonably necessary | Where the data physically lives during the contract | Data stays inside the agency's boundary, and the export format is fixed at award, not negotiated at closeout |
| Deployment | Components needed to operate the system remain available as long as necessary; an ATO before deployment | Whether the system runs without the vendor | The system runs in an environment the agency controls and keeps running after the contract ends |
| Audit record | Agency-led testing with agency-defined data; results the agency may disclose internally | Where agent logs are stored and for how long | Every agent interaction logged in agency systems, with retention set by the agency's records schedule |
None of the right-hand column contradicts the memo. Each clause is a stricter answer to a question M-25-22 already tells agencies to ask, and footnote 25 invites agencies to reward exactly that kind of proposal.
What did GAO find when federal agencies tried to move off a vendor?
GAO found that restrictive vendor practices are common and that no agency had fully assigned anyone to manage them. Its review, GAO-25-107114, looked at six randomly selected agencies and 11 cloud investments; five of the six agencies reported restrictive practices.
The examples read like a list of what architecture ownership prevents.
A vendor required agencies to repurchase the same licenses to use them in the cloud. Another charged extra to run its software on other providers' infrastructure, and one required or encouraged agencies to use its own cloud.
The sharpest case is about data. A contractor that migrated an agency's data into a vendor's cloud "required the agency to pay to regain ownership of the data at the end of the contract."
GAO also found that none of the six agencies had fully assigned responsibility for identifying and managing restrictive practices, and made 12 recommendations, two per agency.
That review covered cloud software in general, not AI, which is why it is a warning rather than a measurement of AI contracts.
Does per-seat pricing fit government AI at scale?
Per-seat pricing is the wrong shape for government AI at scale, because it multiplies with headcount whether or not anyone uses the tool. M-25-22 asks for "pricing transparency," and a per-seat bill says nothing about the AI an agency actually consumed.
Microsoft lists Microsoft 365 Copilot at $30 per user per month, paid yearly, at commercial list price. Government pricing is negotiated separately, but the shape is the same: the bill tracks the number of people with a seat.
| Staff with a seat | Per-seat bill at $30/user/month | Usage-based or owned stack |
|---|---|---|
| 1,000 | $360,000 / year | Tracks tokens consumed or GPU hours run |
| 10,000 | $3,600,000 / year | Tracks tokens consumed or GPU hours run |
| 50,000 | $18,000,000 / year | Still tracks use, not headcount |
Introductory pricing can hide the shape for a while.
In August 2025 GSA announced ChatGPT Enterprise for every participating agency at a nominal $1 for one year. The announcement says nothing about pricing after that year, which is the term an agency is actually committing its workflows to.
A usage-based or owned stack does not make AI free. It makes the cost follow the work, which is what an agency can defend in a budget justification.
Can managed AI vendors run in disconnected or sovereign government environments?
Yes, some managed vendors can now run disconnected, so where the system runs no longer separates the options. Microsoft documents disconnected operations for Azure Local, and in February 2026 it announced support for large AI models running fully disconnected.
That is real progress, and it means "air-gapped" alone is not a reason to choose one vendor over another. The questions that still separate offerings are the ones in the table above: who holds the source code, who chooses the models, and what keeps running when the contract ends.
A disconnected deployment of a licensed product can still leave the agency dependent on one vendor for every upgrade, model and fix. Ownership is the property that answers that, and it is a contract term, not a network diagram.
What does Germany's MissionOSBw source-code handover show about AI procurement?
Germany's handover shows ownership being written into a defense AI program from the start.
On November 4, 2026, Helsing gave BAAINBw, the Bundeswehr's procurement office, the source code, documentation and SDK for MissionOSBw, its combat-drone mission software, with ownership staying with the German state.
The detail that matters for civilian buyers is the second half: the Bundeswehr can develop the software itself or have other companies do it. We covered the MissionOSBw handover in detail, including what the public reports do not yet say about license terms.
Owning code nobody else is allowed to work on is ownership on paper. The clause that makes it leverage is the right to bring in another team.
Why does model switching matter for a government agency?
Model switching matters for an agency because model prices, terms and versions change faster than contract cycles. A system wired to one provider turns each of those changes into a migration project.
Duolingo shows the size of the stakes. Its CEO said on the company's November 3, 2026 earnings call that one AI Video Call cost about $0.30 in inference and now costs "well under $0.01," mainly by moving to open-weight models, a drop of more than 30 times per call.
Duolingo kept frontier models for the cases where they still win. Our breakdown of the Duolingo figures covers what the calls support. For an agency, the lesson is that the right to choose the model per task is worth more than any one model.
What should an agency CIO ask before signing an AI contract?
An agency CIO should ask seven questions before signing an AI contract, each answerable in writing:
- Do we receive the source code of the platform we deploy, or only rights to code written under this contract?
- Can our team change the model provider by configuration, including to an open-weight model on our own hardware?
- Where do prompts, outputs and agent memory live during the contract, and who outside the agency can read them?
- What is the export format for our data, and is it fixed now or at closeout?
- Does the system keep running, unchanged, on the day after the contract ends?
- Where are agent logs stored, and does our records schedule set their retention?
- Does the price track use, or the number of people with a seat?
Each question maps to a row of the clause table, and each answer can be tested before award, which M-25-22 asks agencies to do "to the greatest extent practicable."
How does ibl.ai fit the architecture-ownership model for government?
ibl.ai is built for the ownership column of that table. On ibl.ai you own all the code and the data: the agency receives the platform's source code under a full code license and runs it in its own cloud, on-premise, in GovCloud or fully air-gapped.
The platform is model-agnostic, so an agency can run commercial models or open-weight models on its own hardware and switch between them. Pricing has no per-seat component, and the platform supports PIV/CAC sign-in with NIST 800-53 aligned controls.
ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on. More on deployments for agencies is on our government solutions page.
Sources: OMB Memorandum M-25-22 (April 3, 2025); GAO, Cloud Computing: Selected Agencies Need to Implement Updated Guidance for Managing Restrictive Licenses, GAO-25-107114 (November 13, 2024); 5 U.S.C. § 552; GSA OneGov announcement (August 6, 2025); Microsoft 365 Copilot pricing; Microsoft Azure Local disconnected operations.