The Pentagon is testing a model for government AI procurement that makes contractors cover deployment costs and pays them a share of the savings they deliver, as US agencies face rising bills for AI once introductory deals expire.
In a November 17 commentary for Federal News Network, Michael Garland, founder of the procurement consultancy Garland LLC, and Gaurav “GP” Pal, founder and chief executive of the federal compliance firm stackArmor, argued that shared savings contracts could give agencies a way to modernize without paying the full cost of new systems upfront. Both advise or sell to government buyers, which gives them operational knowledge of federal contracting and a commercial interest in how it changes.
Their proposal comes as McKinsey warns that introductory offers helped accelerate federal AI adoption, while giving agencies little indication of what the same tools will cost at regular commercial rates.
When Contractors Get Paid for Savings
The Pentagon’s Chief Digital and Artificial Intelligence Office signed a one-year agreement on September 8 with Red Cell Partners, a venture studio that builds technology companies in healthcare, cyber and national security,with a ceiling of $100 million. The pilot puts a question about government AI procurement into practice. Can a contractor cover the cost of introducing AI, reduce the cost of an existing government service, and earn enough from the difference to make the work worthwhile?
“No savings, no payment,” Garland and Pal wrote.
Under their proposal, a contractor would recover its costs from a share of the verified cost savings. Federal agencies use comparable arrangements to pay for energy upgrades, and Medicare rewards participating providers that hold costs below a benchmark under its own savings program.
Red Cell estimates that Air Force contracts for audit preparation, valued at $62 million, could be consolidated into an AI-assisted operation costing about $24.8 million. In that scenario, the vendor would receive 30% of the verified reduction. The figures are a projection from the company involved in the pilot. They illustrate how a contract could affect government AI spending if its assumptions hold.
Red Cell has said the pilot services will come from companies in its own portfolio, so the venture studio stands to benefit both as contractor and as investor.
For shared savings contracts to work, both sides need to agree on what the service already costs. Garland and Pal suggested measuring a unit of work, such as the cost of deciding a claim or completing an audit sample. An independent review of that starting figure would give agencies and vendors a common basis for comparing costs after deployment.
The authors argued that government AI procurement should give agencies ownership of the code, workflows, and data produced under the agreement. They proposed keeping existing systems running until savings can be demonstrated. Once the vendor receives its agreed share, the agency could use the new operating cost as a benchmark and invite bids for maintenance.
What Happens When the Discounts End
McKinsey said federal agencies could initially access some AI tools for a dollar or less under introductory agreements launched in 2025, including offers from Anthropic and OpenAI that priced their assistants at $1 for government users. As those offers expire and federal AI adoption expands, agencies will have to budget for work that was cheap to test and could become expensive to run across entire departments, which makes government AI spending harder to forecast.
Under AI usage-based pricing, a bill can increase with the amount of text a model reads and generates or the number of steps an automated tool takes to finish a job. A system that checks a document, consults other records, and revises its answer may call a model several times for a single request.
“While unit token costs are falling, the total cost is rising fast,” McKinsey’s authors said.
The analysis calls for AI cost controls that include spending caps for individual tools, alerts when usage spikes, and the use of cheaper models for routine work when they meet the same quality requirements.
Those bills also matter when a vendor is paid from the savings its software produces. Garland and Pal outlined two options for government AI procurement. A contractor could absorb the AI charges and recover them from its share, or the agency could pay those charges directly and subtract them before calculating the savings to share.
The Red Cell agreement avoids standard procurement rules by using an Other Transaction Authority, a mechanism that lets the department strike agreements outside the usual acquisition process. That is why expanding this form of government AI procurement beyond short pilots would require congressional action, Garland and Pal argued.
They pointed to the E-Government Act of 2002, which authorized share-in-savings contracts for federal technology projects. That authority expired in September 2005 without a single contract being awarded, according to the Government Accountability Office. Garland and Pal proposed giving agencies access to specialists who could help them build and verify cost benchmarks before a deal is signed.
For the Pentagon, the question is whether the pilot can show that an AI-assisted service costs less to run once the software bill is counted. Agencies considering similar deals would also need to decide whether the service still works as well for the people who depend on it.
Inside Telecom provides you with an extensive list of content covering all aspects of the tech industry. Keep an eye on our Intelligent Tech sections to stay informed and up-to-date with our daily articles.