In the few months since the General Services Administration set up contracts to make it easy and cheap to access artificial intelligence tools from major commercial providers, 43 agencies have signed up.

“We can get an enterprise license for $1 or less for the frontier models of the large language models (LLMs), which is incredible, and that gives agencies this amazing opportunity to test out a new technology, either through doing the acquisition themselves, or using USAi and having a chance to try out something that previously might have been cost prohibitive, or if they had tested it out, they were going to be facing only a really very limited pool of users,” said Laura Stanton, the deputy commissioner of the Federal Acquisition Service at GSA, recently at ACT-IAC’s Executive Leadership Conference. “This gives a chance for exploration and testing in a way that we’ve never seen before.”

As more agencies jump on board the AI train, they will have to add new clauses and acquisition requirements to their contracts starting March 11.

Under a new memo from the Office of Management and Budget, agencies have 90 days to update acquisition polices to ensure the LLMs they purchase are truth seeking and ideologically neutral.

“LLMs shall be truthful in responding to user prompts seeking factual information or analysis. LLMs shall prioritize historical accuracy, scientific inquiry, and objectivity, and shall acknowledge uncertainty where reliable information is incomplete or contradictory,” OMB Director Russ Vought wrote in the Dec. 11 memo. “LLMs shall be neutral, nonpartisan tools that do not manipulate responses in favor of ideological dogmas. Developers shall not intentionally encode partisan or ideological judgments into an LLM’s outputs unless those judgments are prompted by or otherwise readily accessible to the end user.”

The memo fulfills one of the requirements from the July executive order, Preventing Woke AI in the Federal Government.

The memo details requirements for new contracts for AI as well existing ones. Additionally, agencies should apply these principles to commercial and internally developed LLMs.

OMB says as part of the contracting effort for LLMs, agencies must obtain information from the vendor to determine that the technology meets these unbiased AI principles.

“The amount and type of information available will vary depending on the vendor’s role within the software supply chain and its relationship with the LLM developer itself, with more information generally being available from sources closer to the original LLM developer,” OMB wrote. “When an agency transacts with such a third-party LLM provider, the availability of product information and potential for direct product interventions will depend on the willingness of the actual AI developer to collaborate through the third-party distributor. Agencies should consider these nuances of LLM procurement in determining how to apply the requirements of this memorandum to ensure that an LLM offered for procurement complies with the unbiased AI principles.”

Memo builds on AI guidance

At the same time, OMB emphasized that agencies shouldn’t require vendors to disclose sensitive technical data, such as model weights.

“Documentation requests should seek enough information for an agency to assess a vendor’s risk management actions at the model, system, and/or application level, as appropriate, to establish compliance…” OMB wrote.

OMB has issued previous memos to agencies for how to ensure they are buying AI tools that meet their needs. In April, the administration laid out several deadlines for agencies, including updating processes and contractual terms to address the use of government data and clearly delineate the respective ownership and IP rights of the government and the contractor.

In October 2024,  OMB tried to address current and emerging questions to ensure agencies are applying the right rigor to the AI tools.

Jose Arrieta, founder and CEO of Imagineeer and a former federal acquisition and technology executive, said the memo is focused on making AI operational, and that is really important at this point and time.

“I read this as guidance that enables AI. It is not about banning models at all. It’s about requiring truth, provenance, and accountability at the contract level,” said Arrieta, who served as chief information officer at the Department of Health and Human Services and as director of the GSA IT Schedule during his more than 15 years in government. “I think it enables agencies to treat AI like critical infrastructure very quickly. It creates a structure that rewards disciplined AI platforms, particularly because it is grounded in enforceable governance. As with anything, the devil is in the details of implementation, and there are certainly ways this could be implemented that introduces risk. But overall, I found it refreshing. In many ways, it feels like an early Christmas gift for procurement and technology professionals who are trying to responsibly deploy AI in federal agencies, as well as for AI firms that are actually building with guardrails.”

He said the memo lays out for the acquisition community the rules of the road for how to evaluate, govern and award contracts that are based on metrics.

For example, OMB says for the minimum threshold for transparency, agencies must request four different data sets, including the acceptable use policy and information about the model, system and/or data as it relates to the LLM.

Workforce training still to come

Arrieta added that the guardrails outlined in the memo also give acquisition and technology professionals much needed top cover to ask vendors harder questions about transparency and model provenance.

“OMB didn’t issue a rule with this memo. They used acquisition policy, and historically acquisition authority has been used to drive a wide range of policy initiatives across government,” he said. “What’s different here is that this isn’t just a policy statement. It’s a very operational set of feedback for agencies. The open question is how agencies will react to it. Will they behave differently even though it’s not a rule? Does this guidance intersect with existing regulations in unexpected ways? Does it give agencies enough certainty to execute, or will it prompt more questions back to OMB?”

Some of those unanswered questions, Arrieta said, center on how best to equip the acquisition workforce to use the memo as practical guidance when buying AI tools.

“This is going to require new acquisition templates,” he said. “Someone is going to have to develop a vendor due diligence playbook. How do vendors respond to bias? In the age of AI, what triggers a termination for convenience versus a termination for cause? The answers may seem obvious at first, but they get much more complex when you work through real AI scenarios. If there’s no explicit audit policy, for example, how do you audit AI models and provide meaningful feedback? These are hard questions agencies are going to have to confront. The memo leaves space for that thinking, which is good, but someone needs to lead by creating playbooks and taking a center-of-excellence approach. Agencies are going to need that kind of support.”

The post OMB sets procurement guardrails for buying AI tools first appeared on Federal News Network.