Shared savings: How government can adopt AI quickly with little financial risk

Federal News Network

Source published: Site published:

Original article

In the past, the government has usually been slow to adopt new technology. For example, the Cloud First policy was introduced in 2010, and its security program, FedRAMP, followed in 2011. Sadly, these good policy goals moved faster than the actual work. More than 15 years later, government agencies are still moving their computer systems to the cloud. We think one reason for this slow progress is that agencies lack people with the right skills. They also have trouble learning how to budget for and buy services that work more like rentals than traditional software purchases.

The move to cloud technology happened slowly, and delays were not a major problem. But the artificial intelligence revolution is moving at incredible speed. A recent filing for Anthropic's initial public offering (IPO) reports that its revenue grew from $41 million in 2023 to $16 billion in just the first half of 2026. That is amazing growth, although the company's losses are also very large. Anthropic says its technology is being adopted by businesses faster than any other technology in history. That seems believable. OpenAI is growing at a similar rate. Bloomberg reports that its annual revenue, based on its September earnings rate, was about $50 billion. That was roughly two and a half times its rate at the end of 2025. Clearly, private companies are already changing how they work by using these tools.

The government needs to join this change too. We believe it can often adopt AI quickly, with little financial risk and very little money needed upfront. It just needs to change the way it pays for the work.

We estimate that the government spends about $20 billion each year just to license commercial software. Adding new AI licenses on top of those costs could cause technology licensing expenses to rise sharply, at least for a while, before older systems can be shut down. That would be a problem, especially since the Congressional Budget Office estimates that the federal budget deficit for 2026 will be $2.0 trillion. At the same time, the government workforce is getting smaller. Pew Research found that the federal workforce shrank by about 10% in 2025. A large wave of retirements is also taking away workers with many years of valuable experience.

Clearly, agencies must get more work done with fewer resources. Unfortunately, the way they often buy technology services makes this harder. Much of the government's technology integration work is paid for based on the time workers spend or the costs companies report. These payment methods can give contractors a reason to take longer than necessary. In the worst case, a badly managed project could make a contractor more money than a smooth and successful one. This problem exists regardless of what the people involved intend to do.

Fortunately, there is already a proven payment method used in other parts of government that could lower the financial risk of AI projects and help start a much larger wave of AI adoption. It is called shared savings. Under this approach, the contractor pays for setting up the new system. The government then pays the contractor a percentage of the money it can prove the new system has saved. If there are no savings, the contractor gets no payment. This takes performance-based contracts one step further. Instead of paying for a certain level of service, the government pays for results that can be measured.

This idea has worked before. Energy Savings Performance Contracts have allowed government agencies to pay contractors using guaranteed energy savings for decades. Medicare's Shared Savings Program also rewards healthcare providers when they keep costs below an agreed target. In September, the Pentagon's Chief Digital and AI Office signed a one-year agreement with Red Cell Partners. The agreement has a maximum value of $100 million and will test this payment approach for AI services.

Red Cell's own analysis shows where this approach could work well. The company estimates that $62 million in Air Force contracts supporting audit readiness could be combined into an AI-supported system costing about $24.8 million. The contractor would receive 30% of the savings that are confirmed. This is an ambitious estimate, but it shows the basic idea. The contractor has a real financial stake in making the project work, and the government is no longer simply paying the bills without a clear link to results.

Of course, this approach will not work for every project. A contractor that pays for a new system must cover all the costs. The biggest cost that can change is likely to be the AI itself. Fortunately, the General Services Administration's OneGov program provides known prices. GSA recently agreed to give a 50% discount on OpenAI token usage through a 27-month agreement. As of October 27, it also offered a 40% discount on token usage across Anthropic's Claude models. Contractors can buy AI services at these prices if they receive permission to do so. They can also try to negotiate even better prices. Having one governmentwide price means every company bidding on a project can plan around the same basic cost. In the simplest version of this approach, the contractor pays for everything, including the AI, and gets that money back from its share of the savings. Agencies could also pay for AI usage themselves, as long as those costs are included when calculating the savings shared with the contractor.

If the contract is designed properly, shared savings works like a fixed-price contract that pays for results. The contractor takes on the financial risk, so it has a strong reason to agree to a starting cost it can realistically meet. The government's job is to make sure that this starting point can be checked and independently confirmed. Most of the important contract work needs to happen here. A good starting point measures how much one unit of government work costs. Examples include processing one claim or completing one audit sample. The starting cost must be fair and reasonable for both the contractor and the government because both sides face financial risks based on it. The government must also make the cost information behind these calculations public and clear enough for companies to trust the starting figures.

The risks of putting the new system in place can be controlled. The old system stays in operation until the savings have been proven. The government owns everything created during the project, including computer code, work processes, instructions given to AI, testing tools and data. The project can have checkpoints where either side can stop early. Contractors can also face financial penalties through performance bonds and damage to their past-performance ratings if they fail or walk away. The government might lose time, but it will not lose its investment money.

The government must also plan what happens after the project succeeds, because savings cannot be shared forever. Once the contractor has earned back its investment and received its agreed share, the lower cost of running the new system becomes the new starting point. The government should then open the ongoing operations and maintenance work to new competition. This is why government ownership of everything created during the project is so important. Without it, only the original contractor might be able to maintain the system, leaving the agency dependent on that company. This is known as vendor lock-in.

Congress has tried this idea before. The E-Government Act of 2002 allowed government agencies to use shared-savings contracts for information technology. But in a 2005 report, the Government Accountability Office found that the required rules had never been created. Only 21 government purchasing employees had taken GSA's training course, and no contracts had been awarded. The legal authority ended that September. A good idea needs the right support and systems to make it work.

One important requirement is a law based on the rules for Energy Savings Performance Contracts. Those contracts allow agencies to pay contractors using savings over several years. Without similar legal authority, a contract that promises to pay a company from future savings over several years could commit government money before Congress has approved the funding. The Anti-Deficiency Act does not allow this. For now, the Pentagon's one-year other transaction agreement (OTA) pilot avoids this problem because it lasts only one year. Expanding the approach beyond one-year pilot projects and meeting other purchasing requirements will need new legislation. Another helpful step would be to create a small office with experts who can design and check cost starting points. The Energy Department's Federal Energy Management Program already provides similar support to agencies working on energy-saving contracts. Once the costs have been organized and understood, this office could create standard templates that government agencies could use when setting up shared-savings contracts. This office could be part of GSA.

The move to cloud technology changed government from owning technology to renting it. Shared savings could take government one step further, from renting services to paying for results. One of us has written in detail about the risks of using AI in government, including becoming too dependent on one company and dealing with prices that are hard to understand. That work comes from wanting AI to succeed, and it is exactly why this payment model is appealing. Moving quickly and making careful contracts can work together. The government can adopt AI quickly while keeping its financial risks low and gaining the huge benefits of this technology. Private industry is already making this change. The longer the government falls behind, the more it will have to explain why. We should not miss this once-in-a-lifetime opportunity.

Michael Garland is the founder of Garland, LLC, a consulting company that advises businesses and government clients on federal purchasing laws and the information technology business.

Gaurav “GP” Pal is a government technology executive and transformation expert. He focuses on helping organizations adopt technologies such as cloud computing and artificial intelligence safely and securely.