OPM.gov /

Use, Sentiment, Impact

Guest post by Adam Starr, U.S. Office of Personnel Management

AI adoption is measurable with three metrics — and the numbers tell you what to do next.

Nobody at your agency has a Microsoft Word adoption strategy. Word is simply how documents get written, the way email is how messages move. My theory is that AI becomes part of the operating system of how work gets done. Broadly used, barely remarked on.

So, how do we get there? We need metrics to understand where we stand and guide our next steps. Our signals come from three questions. Do people use the tools? Do they like them — do they feel the tools help them do the work? Is the business better? Use, sentiment, impact. That is the whole framework.

Three numbers guide investment

Use is the simple count to see if people are even trying the tools. My advice is to count from the logs, not just the survey.

Sentiment is where you can start to understand deeper barriers to adoption. Maybe people use them out of necessity, but there is room to improve. I like the Net Promoter Score (NPS) question that brands use to measure loyalty: would you recommend our approved AI tools to a colleague? Then calculate promoters minus detractors, on a scale from −100 to +100. Above zero means more promoters than detractors. Fifty is excellent. Seventy is outstanding. It is a hard metric that pushes programs to constantly improve.

Impact requires no new dashboard (ideally), and you should be suspicious of anyone selling you one. The impact measures are the business metrics the agency already runs on: days to process a retirement case, the cost curve of the health insurance program we administer, time to hire. All of these efforts must be in service of the mission.

What the numbers said

In our June workforce pulse survey, we not only asked about culture but also about AI adoption. Scott talked about some of the results earlier this week. We had a 56% response rate with over 1,100 responses. Let’s dig into what the numbers show us. 81.4% of employees said they use our approved AI tools, with self-reported use climbing by seniority: 77.4% of non-supervisors, 93.5% of supervisors, 97.3% of senior executives. Our telemetry puts actual use closer to 70% overall. The gap could be sample bias or reflective of user self-perception. Again, this is one reason to prefer usage data from the tools themselves.

As for sentiment, OPM’s overall NPS value is +17.9. More promoters than detractors. That said, the spread of results across teams was massive.

At the top is the Office of the Director at +95.7: of 23 respondents, 22 are promoters and not one is a detractor. That is the front office — unstructured knowledge work: synthesis, drafting, analysis, the blank-page problems — and for that work the current AI chatbots and agentic co-work tools are a force multiplier. Our technologists in OCIO scored +65.0; the coders found these tools early and would be up in arms if we ever tried to take them away. Our human capital office is close behind at +64.1.

The value shows up in how people’s daily work has changed. One member of our Organization Design practice used to spend hours, sometimes days, hand-coding focus-group feedback in an Excel workbook so themes could surface. Now he attaches the service catalog, the session protocol, and the raw responses, assigns the model an expert role, and gets the categories and themes in minutes. His work now is validating and using the data. Another member of our engineering team has a team of agents, all with different personalities, that help decode and modernize mainframe programs. Another team member has changed how her team tracks progress with custom dashboards. Across OPM, AI is changing how people work.

We still have a long way to go. At the other end: the CFO organization at +5.4, our legal office at −4.8, and Retirement Services at −38.1, where nearly six in ten respondents are “detractors.” In Boyers — home to our retirement operations mine — “AI tools are improving the way my work unit performs” scored 4.6 out of 10. While new retirement applications may be digital, we still have a paper backlog and a structured process where we can rethink and simplify a newly digital process. Still, people in Retirement Services are experimenting with AI tools and finding novel applications. One employee built a tool that automates a premium-shortfall calculation his unit does by hand roughly 11,000 times a year — what took 5 to 10 minutes of manual math per case now comes out as an auditable worksheet in seconds.

The pattern is the work, not the people

Don’t read these results as a measure of enthusiasm. The pattern tracks the shape of the work. Where work is unstructured knowledge work, a general-purpose assistant fits it on day one. Where work is structured — paper processes, fixed case flows, rigid systems of record, procedures refined over decades in finance, in legal, in retirement case processing — a chat window in a browser tab is beside the point. Without meaning to, we had asked those employees to bring their work to the tool.

The fix runs the other way: bring the tool to where the work happens — into the case system, the financial workflow, the document pipeline.

Employees said as much when we asked directly. The top barriers to more effective use: lack of time to learn or experiment (47.5%), accuracy and reliability concerns (34.3%), and security or data-sensitivity concerns (23.8%). Only 1.4% said their supervisor discourages AI use. The barrier is not the tool or lack of interest. The blockers are capacity and trust — both fixable.

And we are acting to overcome these hurdles. We are giving $1,000 bonuses to employees who innovate with AI. We are partnering engineers with program offices in our AI Catalyst program to help make adoption easier. We have clarified our policies in our all hands meetings and in emails to all employees. And when employees have found unclear or inconsistent policy language, we have clarified it within a day.

One more number worth staring at: the lowest-scoring item on the entire survey was “OPM approved AI tools are improving the way my work unit performs,” at 7.3 out of 10 agency-wide. Use is running ahead of felt impact. We have seen many examples of tasks that are easier or processes that are simpler. Still, we have the lagging indicator of impact to our mission. We are not done. Closing that gap is the whole job, and these metrics guide us.

How we are guiding our action

Everything above — the Catalyst partnerships, the bonuses, the policy fixes — came straight from the numbers. Each could be its own post.

But measurement comes first because it is the part any leader can start this quarter. If you run an agency, a bureau, or a business line, everything above fits in one survey cycle: count use, score sentiment, and tie impact to the business metrics you already report. The score is not the point. The score sets a baseline and tells you where to look. Then invest in overcoming the barriers in your organization.

Mind the AI GAP

OPM stood up the AI Governance and Adoption Program — AI GAP — to help every agency in government measure adoption, understand its barriers, and close its gap. The acronym is not an accident: every agency has one. Write us: AIGAP@opm.gov.