Incident sheet · /enterprise-ai
- System
- Enterprise AI
- Filed
- Read
- 4 min · 883 words
- Severity
- Routine
Pentagon's $100M Shared-Savings AI Pilot: Can Government Buy Results, Not Hours?
A Pentagon pilot with a $100 million ceiling pays AI vendors only from verified savings. Backers cite precedent — and a 2005 failure that killed the same idea for IT.

Event log
- The Pentagon's CDAO signed a one-year, $100 million-ceiling shared-savings agreement with Red Cell Partners in September.
- Red Cell projects $62 million in Air Force audit contracts consolidating into a $24.8 million AI model, vendor earning 30% of verified savings.
- GSA set a 50% OpenAI token discount for 27 months and a 40% Anthropic discount as of Oct. 27.
- GAO found in 2005 that share-in-savings IT authority from 2002 produced zero contracts and expired unused.
The Pentagon signed a one-year, $100 million-ceiling agreement in September with Red Cell Partners to test shared-savings contracting for AI services. That pilot, brokered by the Chief Digital and AAI Office's Chief Digital and AI Office, is the sharpest signal yet that some federal buyers want to pay for measured outcomes instead of labor hours. The authors of the proposal, procurement consultant Michael Garland and technology executive Gaurav "GP" Pal, argue the model can move government AI adoption from renting to results. The mechanism has precedent. It has also failed in government before.
Why the hurry now?
The numbers driving the argument are stark. Anthropic's IPO filing shows revenue grew from $41 million in 2023 to $16 billion in the first half of 2026 alone — the company calls it the fastest enterprise adoption of any technology in history. OpenAI is on a comparable curve: Bloomberg reports its annualized revenue hit roughly $50 billion at the end of September, about 2.5 times its run rate at the end of 2025.
Government finances point the other way. The Congressional Budget Office puts the fiscal 2026 deficit at $2.0 trillion. The authors estimate federal spending on commercial software licensing at about $20 billion a year. Pew Research found the federal workforce fell roughly 10% in 2025. A retirement wave is draining long-tenured expertise. Agencies must do more with less.
History counsels patience with such urgency. The Cloud First policy arrived in 2010, FedRAMP in 2011, and more than 15 years later agencies are still migrating workloads. The authors attribute the delay to thin in-house expertise and budgeting models built for purchase, not rental.
How does shared savings actually work?
The contractor funds the deployment. The government pays a percentage of verifiably delivered savings. No savings, no payment. It extends performance-based contracting from service levels to measured results.
Red Cell's own projection shows the shape of the deal: $62 million in Air Force audit-readiness support contracts consolidated into an AI-augmented model costing about $24.8 million, with the vendor earning 30% of the verified reduction. The authors concede the projection is aspirational. They note it has the right shape: the vendor has skin in the game, and the government stops acting as an ATM.
Precedents exist in adjacent domains:
- Energy Savings Performance Contracts have let agencies pay contractors from guaranteed energy savings for decades.
- Medicare's Shared Savings Program rewards providers who hold costs below a benchmark.
What about the incentive problem the model fixes?
Much federal integration work is bought on time-and-materials or cost-reimbursement terms. Those terms reward providers for taking their time. In the most cynical outcome, a flawed rollout can be more lucrative than a seamless implementation, the authors write — whatever anyone's stated intentions.
AI input costs are no longer opaque, they argue. GSA's OneGov program recently set a 50% discount on OpenAI token usage under a 27-month agreement and, as of Oct. 27, a 40% discount on Anthropic's Claude models. Contractors can buy at those prices under a delegation of authority, or negotiate better. One governmentwide price gives every bidder the same input cost to plan around.
Where does the risk actually sit?
Everything depends on the baseline. The government must insist it is auditable and independently verified — that is where the bulk of contracting work must be done, and where financial risk resides for both parties. A sound baseline measures the cost of a unit of government work: a claim adjudicated, an audit sample closed. The authors urge agencies to publish cost inputs transparently enough for industry to trust them.
Safeguards they recommend include:
- Legacy systems stay live until savings are proven.
- Government owns code, workflows, prompts, evaluation suites and data.
- Stage gates let either side stop early.
- Performance bonds and past-performance ratings make walking away costly.
Success needs planning too. Once the vendor recovers its investment and share, the new operational cost becomes the baseline and O&M should be recompeted. Without government ownership of the work product, the builder becomes the only possible maintainer — vendor lock-in by another name.
Has Congress tried this before?
Yes, and it went nowhere. The E-Government Act of 2002 authorized share-in-savings IT contracts. GAO reported in 2005 that implementing regulations were never issued, only 21 acquisition employees had taken GSA's training course, and no contracts had been awarded. The authority expired that September. The authors' verdict: a good idea needs infrastructure to go with it.
The legal gap remains. A multiyear savings payout risks obligating funds before appropriation, which the Anti-Deficiency Act forbids. The Pentagon pilot sidesteps this as a one-year other transaction agreement. Scaling beyond pilots needs statutory authority modeled on Energy Savings Performance Contracts, plus a small office — possibly at GSA, akin to the Energy Department's Federal Energy Management Program — to design cost baselines and templates.
Cloud moved government from owning technology to renting it. The authors say shared savings can move it from renting to results. The skepticism they themselves have applied to AI in government, from lock-in to opaque pricing, is exactly why they back this model. The further government falls behind industry, the more it will have to answer for.
via federalnewsnetwork.com (Source)
More from Daniel Okafor
Show full bio
Staff writer covering technology and product news in enterprise information technology and the leaders who run at CIO Monitor.
203 articles
The newsroom digest
One email a day with the stories that matter. No noise.