AI Forces a Rethink of Federal Contracting Models

Artificial intelligence is rapidly becoming part of the federal acquisition process, not simply as a technology agencies want to buy, but as a tool that is changing how government and industry approach the business of buying and selling technology. 

Federal contractors are using AI to analyze solicitations and draft proposals. Agencies are experimenting with AI to conduct administrative reviews of responses. And increasingly capable AI agents are beginning to perform work that once required teams of human developers, analysts and other professionals.  

The technology is moving quickly. Federal acquisition, by contrast, remains built around processes that can take years. That mismatch is becoming one of the central challenges for agencies trying to adopt AI. 

Anil Chaudhry, senior advisor for AI, autonomy and national security at the Transportation Department, described application development as being at an “inflection point,” driven in part by agentic AI’s ability to accelerate coding, development and testing. 

But the question is how to reconcile that pace with a contracting system in which five-year agreements remain common.

“If you have requirements that were developed five years ago and were based on the reality of five years ago,” Chaudhry said in an interview with GovCIO Media & Research, “how do you adapt your contracts to the reality of now?”

The answer emerging from federal acquisition officials and industry is not to abandon established controls. It is to make the process more modular, transparent and focused on outcomes.

AI is Already Changing the Acquisition Process

Chaudhry said that previously companies might have had the bandwidth to seriously consider roughly 10 opportunities, for example. Now, AI could allow them to respond to “10 times that amount.”

Ricky Clark, NITAAC director at NIH, warned companies against relying too much on the new tools. Federal evaluators can often recognize when proposals have been heavily generated by AI, he said, especially when multiple submissions contain strikingly similar language, or even the same mistakes.

“It was easy for us to see because of the volume of proposals that we have,” he said at GovCIO Media & Research’s Health IT Summit. “There were always syntax errors. There were always, you know, misplaced orphans, hallucinations. … So if you’re using AI, you want to use it responsibly. You want to have someone that can overlook that. I always want to keep the human in the loop.”

Clark advocated for keeping a subject matter expert in the loop to verify what the technology produces. That is especially important when a company is paying substantial funds to have a proposal prepared.

On the federal side, agencies are using AI to manage the flood of additional proposals. But federal acquisition leaders make clear that AI tools are not making decisions; they are helping agencies streamline the review process. At NITAAC, Clark’s team developed an AI-powered proposal tool for algorithm validation and document consistency.

“By employing AI at every step, we created an avenue, if you will. We used AI to identify the gaps and deficiencies. Those gaps and deficiencies were then forwarded to the CEO, who can make a decision,” he said.

Chaudhry said AI tools are useful for conducting an initial “administrative sweep” to filter proposals.

“Not a technical sweep, which is a human endeavor, but things like if the contract was meant for small businesses, is the company certified to be a small business? That’s something that an AI agent can do very well,” he said. “You don’t need to have a contracting specialist spend 11 hours going to different websites to figure out if that company’s still in business and still registered as a small business.”

AI Brings New Efficiencies Along With New Challenges

The prospect of AI writing solicitations and AI evaluating those proposals raises an obvious question: Are federal acquisitions heading toward a world where AI is talking to AI about AI?

Ben Cushing of Red Hat said that scenario presents some challenges. Large language models’ knowledge reflects their training, and newer research or information may conflict with what a model has previously learned. That creates the possibility that an AI evaluator could flag a contractor’s proposal as incorrect simply because the proposal reflects newer information than the model has encountered.

Cushing’s recommendation is not to try to game an AI evaluator, but to make the reasoning behind a proposal especially clear.

“It is very important that you take your solicitation and run it through a series of LLM services to see what they say back to you,” he said at the summit. “If it comes back and there’s a contradiction, and the weights of the existing model are unable to reason with the new research you’re providing. You need to explicitly say the old research said this, the new research says this, and then have a citation to the new research. Otherwise, a large language model is going to spit out a bunch of problems with your solicitation.”

Another challenge surrounding AI in acquisitions, according to Quentin McCoy, executive director in the Office of Mission Acquisition Solutions at HHS, is making the mistake of treating AI as a replacement for acquisition expertise rather than as a tool for making that expertise more effective.

McCoy said that AI can augment judgment, but it cannot assume responsibility for that judgment. That distinction matters particularly in government, where acquisition decisions involve taxpayer dollars and can be challenged through protests and other oversight mechanisms.

“There’s accountability for fiscal dollars,” he said at the summit. “You can’t take the AI to jail, right? But I can go to jail if I misuse, you know, fiscal dollars, and so that’s the important piece that we are telling our folks as we train them up to this AI generation … how to use those smartly and to also not lose the judgment piece that you’re supposed to make and that you’re entrusted to have.”

Rethinking Contracts for an AI-Speed World

The current contract structure often struggles to keep pace with the speed of AI technology. Experts suggest changes could include modular contracts, regular assessments of value and meaningful communication with industry.

Chaudhry said modular contracting is one potential answer because the government could break work into smaller phases rather than lock agencies into requirements developed years earlier, potentially allowing agencies to reassess technology and requirements every six months.

McCoy said agencies should think in terms of prototyping, iteration and scaling rather than automatically defaulting to a five-year contract. A longer agreement might still make sense in some circumstances, but it should allow the government to evaluate value and adjust course as technology changes.

Measuring AI by Outcomes, Not Headcount

Federal acquisition teams will likely need to define AI-related labor categories, performance metrics and pricing models in order to distinguish between the cost of operating an AI system and the value of the work it produces.

If AI allows a smaller team to accomplish what once required a much larger workforce, then staffing numbers become a poor measure of value. McCoy said HHS is now asking vendors not simply how many people they can hire or what their rates are, but what technologies they are using to augment their work.

For Chaudhry, the addition of AI is making it harder to build a budget for federal agencies. There are two main ways to buy things in government: firm-fixed-price contracting and structured negotiation, which requires some back and forth to get the best price for the government.

“This is where things get tricky because now in the age of agentic AI, are you really measuring apples to apples or apples to oranges?” he said. “Vendor A might have a proposal where their delivery team is 80% people and 20% AI agents, and they give you a certain price, and vendor B gives you a team where it’s 10% people and 90% agents. How do you compare and contrast to make sure government is getting the best value?”

GSA currently uses labor categories to determine pricing based on education and experience. Chaudhry said agencies should start developing similar categories to evaluate AI agents as they increasingly become part of the delivery team. He advised against using token consumption as a pricing metric because it’s more of an “activity, not an outcome.”

Clark agreed that outcomes and return on investment should play a larger role in contract considerations.

“The competitive landscape for us is changing,” Clark said. “Now it should be driven more about the return on investment, ROI. So before in the federal government, we just make a buy, and as long as you get the job done, that’s good. Now I think as we evolve, the ROI period is huge for us.”