Start small, measure well: a better path to federal AI adoption

The Office of Management and Budget’s 2025 inventory reported more than 3,600 artificial intelligence use cases across federal agencies. As agencies decide which applications deserve greater investment, the number of AI deployments matters less than whether those deployments improve the work agencies are doing.

That puts responsibility on government and industry alike. Before proposing a new AI capability, contractors should help agencies answer a basic question: What result would make this technology worth expanding?

The best place to start is a well-understood task with an existing baseline. Give AI a defined role in that task, compare the results with the existing process, and use that evidence to decide what comes next.

Start with a specific problem

Department of Commerce Acting Chief Technology Officer Matheus Passos recently shared how a Census Bureau employee used AI to build a tool for tracking and deploying AI across the department, which was quickly adopted by other bureaus.

The people doing the work often see the best opportunities first. They know where time gets lost, which tasks consume disproportionate effort, and where automation might help.

Software management is a good example. IT teams regularly review large amounts of information about applications, licenses, and usage. Much of that work involves repetitive analysis, and agencies can already measure outcomes such as review time, accuracy, and software utilization. A focused pilot could test whether AI reduces manual review or identifies records that need attention without giving the technology control over the broader process.

For government contractors, that means starting with the workflow rather than the AI platform. Understand how an agency performs the work today, establish a baseline, and identify the specific outcome the technology should improve.

Measure against what already matters

Agencies often already have the metrics they need. A service desk may track mean time to resolution or ticket backlog. Other teams may focus on processing time, accuracy, or cost. Those measures create a baseline, but agencies must account for the full workflow.

We see this often in proposal responses. Generative AI can produce a lengthy draft in seconds. But if a subject-matter expert then spends hours correcting errors or rewriting the response, AI has simply shifted the work downstream. A narrower application may work better: AI can organize requirements or create a structure, while the expert remains responsible for the substance.

Government IT teams should apply the same test. An automated system might close service requests faster, but teams still need to know whether those issues stay resolved and how much time an employee spends reviewing inaccurate recommendations. Prompt counts and usage rates can measure adoption, but agencies should judge performance against the operational outcomes that mattered before AI entered the workflow.

Define the limits before deployment

Measurement also requires clear boundaries. In any workflow, an AI recommendation depends on accurate information about specific applications, licenses, and usage. Employees need enough context to evaluate the recommendation before acting on it.

Agencies and contractors should define the technology's authority before a pilot begins. Summarizing information carries less operational risk than changing a production system. As AI takes on more responsibility, teams need checkpoints for human review and a record of the actions the system takes.

Those boundaries help agencies evaluate whether a pilot can scale. If employees are regularly correcting AI’s recommendations agencies should account for that effort when determining whether or not the tool improved the workflow. A pilot that succeeds only because it has unusually clean data or extensive expert oversight may not produce the same result elsewhere.

Turn pilots into acquisition evidence

A successful pilot should make the next investment decision easier. If AI improves a defined workflow without increasing errors or review time, leaders can test the same approach with another team or related workflow to see whether the improvement holds.

That requires capturing lessons along the way. In April 2026, the Government Accountability Office found that several agencies were not systematically collecting lessons learned from AI acquisitions, limiting their ability to apply successful practices and avoid repeating mistakes.

Contractors can help close that gap by building measurement into proposals, pilot plans, and implementation. They should identify the baseline, expected improvement, deployment limits, and criteria for deciding whether to expand, adjust, or stop the effort. A focused use case that produces measurable improvement gives an agency stronger evidence for its next investment.

Some pilots will justify further investment, while others will show that AI is a poor fit for the task. Both results are useful if agencies and contractors agree at the outset on what success looks like.

Egon Rinderer is the senior vice president of federal and enterprise growth at NinjaOne. With over 35 years of experience spanning the Department of Defense, intelligence community, and private sector, Egon Rinderer has built a career tackling the most complex technical challenges facing modern enterprises.