Karen Whitfield, deputy director at a federal agency's data office, spent nine months and a lot of political capital hiring a brilliant machine learning engineer named Devon away from a tech company. Devon took a 30 percent pay cut to do mission work that mattered. Fourteen months later, Devon resigned, and the exit interview was not about money. Devon had spent most of those months stuck behind procurement delays and an aging laptop, was the only AI person in a 200 person office with no one to learn from, and could not see any job title above the one already held.

"I came to do AI," Devon said. "I ended up filing tickets." Karen had focused entirely on recruiting and almost not at all on what happens after someone says yes. That is the trap this lesson is about. Government cannot out-pay the private sector for AI talent, and trying to is a losing game. But pay is rarely the real reason people leave government AI roles. They leave because of the things money does not fix: no growth path, no peers, no tools, no time to do the actual work. Those are things a thoughtful manager can change.

What you are actually competing on

Let us be honest about the pay gap. A senior AI engineer in industry might earn $250,000 or more in total compensation. A comparable federal role might offer $150,000. You will not close that gap, and you should stop trying to win on salary alone. The federal pay structure caps most technical positions at the top of the General Schedule, with the Senior Executive Service and the Senior Level and Scientific and Professional bands above it, while base salaries at leading private AI labs start at multiples of those bands for comparable roles. Government cannot win on base pay. It has to win on a different value proposition.

That value proposition has five parts. Mission: federal AI work shapes the life chances of citizens, and few private roles match that. Scale: federal systems touch millions of people, and few private systems match that either. Stability and benefits: the federal retirement system, Thrift Savings Plan matching of up to 5 percent, federal health insurance and paid parental leave compare favorably to volatile startup equity. Security clearance value: a cleared AI professional holds a portable asset in the defense and intelligence market. And public recognition: Presidential Rank Awards, the Service to America Medals, and genuine public visibility for meaningful work.

Money is a hygiene factor. It has to be tolerable, not best in class. Above the tolerable line, people choose government AI work for the reasons just listed, and your job is to deliver on those reasons and remove the friction that poisons them. Devon did not leave because of the six-figure gap that was known and accepted going in. Devon left because the mission Karen promised was buried under broken tools and isolation. The competition is not really industry's paycheck. It is your own organization's friction, and that is the part you control.

The three leaks that drain AI talent

Retention failures in government AI cluster into three leaks. Karen's office had all three, and each one has a fix that costs far less than a recruitment campaign. Diagnose which leaks you have before you spend money on any of them, because the interventions are not interchangeable. A career ladder does nothing for someone whose real problem is that no one has ever reviewed their code, and a community of practice does nothing for someone whose real problem is that they cannot get a development environment approved.

Leak 1: no growth path

Most agency job classification systems were built before AI roles existed. There is often no defined career ladder for a data scientist or AI engineer, so the only way up is to become a manager and stop doing the technical work that drew the person in. Too many federal AI programs force technical contributors onto a management track to earn promotions, and that drives attrition directly. Devon looked up and saw a ceiling. People do not stay where they cannot grow, and they can usually tell within a year whether growth is available.

Leak 2: no peers, no learning

A lone AI specialist in a non-technical office stagnates. There is no one to review their work, argue an approach with them, or pull them out of a rut. Industry offers dense teams of peers and a culture where negative results get published internally and technical debate is normal. Isolation is one of the most cited and most underestimated reasons technical people leave government. It also carries a quiet quality cost: work that no qualified colleague has ever challenged tends to ship with its assumptions intact.

Leak 3: no tools, no time

Slow procurement, locked down environments, and underpowered hardware mean talented people spend their days fighting the system instead of doing the work. Nothing demoralizes a skilled engineer faster than watching their skills atrophy behind a ticket queue. This leak is also the one that most damages your recruiting story, because departing staff describe it accurately to their networks. Traditional IT organization design rewards predictability, while AI work needs tolerance for failed experiments, and the tooling gap is where those two cultures collide first.

Why this is now a formal obligation

Karen's problem is also a policy obligation, which is useful when you need cover for the effort. OMB memorandum M-24-10, issued in 2024, requires every CFO Act agency to identify, hire, train and retain AI talent, placing workforce alongside governance and risk management rather than beneath them. The 2023 executive order on artificial intelligence, EO 14110, carried workforce directives of its own, and the National AI Advisory Committee has issued workforce recommendations agencies can draw on. Executive orders are revoked and replaced by later administrations, so treat the order as historical context and the durable point as the obligation itself.

The reason the obligation exists is blunt: the federal AI workforce gap is the single largest constraint on federal AI ambition. An agency can publish a risk framework, complete an inventory and stand up a governance board, and none of it produces a working system without people who can build and evaluate one. That is the argument to make when a talent proposal competes with a tooling proposal for the same money. Frameworks without practitioners are documents, and documents do not clear a claims backlog.

Hiring speed is a retention problem too

Hiring authorities are where agencies win or lose the competition for talent, and speed is the decisive variable. The traditional competitive process can take 90 to 120 days from announcement to offer. Private sector offers close in days. Every week your process consumes is a week a candidate spends fielding faster offers, and the candidates most worth having are the ones with the most alternatives. Speed is also a retention signal: a candidate who waits four months to be hired learns something about how long everything else in your agency will take.

Several federal authorities exist specifically to shorten that clock, and the substance matters more than the labels. A Schedule A hiring authority for data scientists, issued by the Office of Personnel Management in 2020 and expanded in 2021, allows direct appointment without a competitive announcement for up to three years. Direct Hire Authority under 5 USC 3304 covers cyber and information technology positions and has been extended to AI roles. The Subject Matter Expert Qualification Assessment, introduced by OPM in 2020 and codified in 2022, replaces generic self-assessment questionnaires with technical panels, and is reported to cut time to hire by roughly 60 percent.

Entry and rotational pathways are the other half. Pathways Program internship and recent graduate tracks are consistently under-used for AI roles. The U.S. Digital Corps showed what a structured two-year cohort for recent graduates can do. The Presidential Innovation Fellows program has brought senior private-sector technologists in for 12-month terms, and the tour of duty model used by the U.S. Digital Service, governed in part by the Intergovernmental Personnel Act, recruited mid-career technologists into two-year postings across agencies. Named federal programs are created, merged, renamed and closed by administration, so confirm what exists today before you design a pipeline around one.

Classification: stop hiring AI talent as generic IT

Federal AI roles are spread across several occupational series, and the one you choose shapes the promotion path, the qualification standards and the applicant pool you can reach. The common families are GS-2210 information technology management, GS-1530 statistician, GS-1550 computer scientist, GS-0854 computer engineer, and the newer GS-1560 data scientist series established under OPM guidance. Related work in operations research and economics sits in other series again. Getting this wrong is the most common structural mistake in federal AI hiring, and it is expensive to unwind once someone is on board.

The specific failure is treating AI talent as generic information technology. A generalist IT track advertises generalist IT work, attracts generalist IT applicants, and offers a promotion ladder built around systems administration rather than modeling, evaluation and data engineering. It also makes retention harder, because the person's official duties and their actual contribution drift apart until nobody can write a defensible performance plan. Decide deliberately which series each role belongs in, and involve your human capital office early enough that the decision is theirs as well as yours.

Building an AI career track inside government

The most powerful retention tool Karen has is a technical career ladder: a defined path that lets someone grow in pay and stature without leaving the technical work. This directly fixes the first leak, and it signals to every recruit that there is a future here. The pattern to copy is a dual ladder, with a technical track where senior practitioners keep doing hands-on work at increasing scope, and a management track for those who prefer leadership. A simple four-rung version, with both branches present at the top, looks like this.

Rung	Title	Focus	Growth signal
1	AI Practitioner	Builds under guidance	Mastering the craft
2	Senior AI Practitioner	Leads projects independently	Trusted to own work
3a	Principal AI Engineer	Sets technical direction, mentors	Top of the technical branch, no need to manage
3b	AI Team Lead	Manages people and portfolio	Top of the management branch

The dual top rung is the key insight. Your best engineer should be able to reach the senior pay band by getting deeper technically, not by being forced into management they do not want and may not be good at. Government already has working precedents for this shape: research scientist ladders in the Forest Service, the investigator ladder at the National Institutes of Health, and the scientific and professional pay plan at the National Institute of Standards and Technology. The Chief Digital and Artificial Intelligence Office at the Department of Defense built clearer technical ladders for senior AI practitioners when it reorganized from the Joint AI Center.

The levers that exist inside the pay cap

You cannot lift the cap, but several authorities let you work inside it, and most agencies use fewer of them than they are entitled to. Special Salary Rates under 5 USC 5305 for cyber and information technology occupations can lift base plus locality by 20 to 40 percent above the standard schedule. Recruitment, relocation and retention incentives, known as the 3Rs, allow lump sum or installment payments of up to 25 percent tied to a service agreement. Critical Position Pay under 5 USC 5377 covers roughly 40 positions government-wide at rates up to Executive Level I.

The Student Loan Repayment Program under 5 USC 5379 provides up to $10,000 per year and $60,000 over a career, which is unusually effective for early-career AI technologists carrying graduate school debt. Telework flexibility under the Telework Enhancement Act of 2010, as reshaped by post-pandemic guidance, remains a real differentiator against employers who have withdrawn it. Non-monetary recognition costs nothing and lasts: award nominations, public credit for a shipped system, and a named role on a visible program all build the career value that a private offer has to overcome later.

These levers have a failure mode worth naming. Because they are discretionary and paperwork-heavy, they tend to be used for the person who has already resigned rather than for the person quietly deciding whether to look. A retention incentive negotiated in the exit interview is a counteroffer, and counteroffers have a poor record everywhere. Decide in advance which roles are hard enough to fill that you would use these authorities, document that judgment, and apply the lever before the resignation rather than after it.

Development pathways that cost little

Growth is not only promotion. Several high-impact development moves cost almost nothing and plug the second and third leaks directly, which matters because they are also the leaks a manager can address without a classification review or a budget amendment. The list below is deliberately ordered from cheapest to most involved, and the first item is the one that most reliably changes how isolated practitioners feel about their week.

• A community of practice. A standing cross-office group where the scattered AI people across the agency meet to share work and review each other's approaches. This manufactures the peers Devon never had, even when no single office has a full team. Agency-wide AI communities of practice run on exactly this logic.
• Protected learning time. A guaranteed share of the week for skill development and experimentation, defended on the calendar and named in the performance plan so it survives a busy month.
• Rotations and detail assignments. Short stints on a high-profile AI project or with a leading agency team, giving exposure and new colleagues without a permanent move.
• Fellowships and structured cohorts. Fellowship and digital service programs exist precisely to move experienced technologists into agencies for a fixed term and to give early-career hires a cohort rather than a desk.
• Conference and certification support. Sending people to learn and bring knowledge back, which signals investment in them and produces a debrief the rest of the team can use.
• Stretch projects. Deliberately handing someone a problem slightly beyond their current level, with support, which is how skilled people most want to grow.

Onboarding and the knowledge that walks out the door

Neglected onboarding is a retention defect disguised as an administrative one. A new AI hire needs a structured 60 to 90 day start that covers information security and the authority to operate process, the agency's AI governance obligations, the specific hiring and pay authorities that apply to them, and the actual program they will support. Without it, a capable person spends their first quarter guessing at how the agency works, concludes that nobody was expecting them, and starts the internal countdown that ends in a resignation nobody saw coming.

The mirror problem is institutional memory. Documentation, mentorship pairings, rotational assignments, an active community of practice and honest succession planning are what stop a departure from erasing a capability. Digital service teams that invested heavily in onboarding documentation and internal wikis did so for this reason, and the national laboratories have decades of scientific succession practice worth borrowing. Relying solely on contractors makes this worse rather than better: contract staff do not retain institutional knowledge on the agency's behalf, and the agency is exposed when the contract ends.

A retention diagnostic you can run this quarter

Before Karen launches programs, she should diagnose her actual leaks. Ask these questions of every AI role on the team, honestly, and write the answers down where a successor can find them. The point is not the score. The point is that the pattern of negative answers tells you which of the three leaks you actually have, and therefore which of the interventions in this lesson is worth your limited political capital this quarter.

1. Growth. Can this person advance in pay and stature without leaving technical work? If no, you have the first leak.
2. Peers. Does this person have at least two technical peers to learn from? If no, the second leak.
3. Tools. Can this person get the compute and software they need within days, not months? If no, the third leak.
4. Time. What share of their week goes to actual AI work versus administrative friction? Below half is a red flag.
5. Mission line of sight. Can this person name the public outcome their work serves? If not, the one advantage you hold over industry is going unused.
6. Recognition. Has this person's technical contribution been visibly recognized in the last six months?
7. Onboarding. If this person started in the last year, did they get a structured start, or did they improvise it?

For Devon, the first three answers were all no, which is why no salary offer could have saved that hire. Notice that most of these questions can be answered by the supervisor without consulting anyone, and the rest need a single honest conversation with the person themselves. This is a diagnostic you can complete in an afternoon, which is the only reason it will actually get done, and the answers are worth writing down so that next quarter's version shows movement rather than mood.

What a workforce program should report

Talent programs die of vagueness. If you cannot show whether the money and the authorities are working, the next budget cycle will decide for you. Track time to offer, offer acceptance rate, first-year attrition, three-year retention, internal mobility, promotion rates at the grades where your ladder is supposed to be working, average training hours per technical employee, clearance processing time where it applies, and the engagement data you already collect through the federal employee survey for the relevant subgroup.

Two cautions on these numbers. First, counting trainings delivered rather than skills acquired is the classic way to mistake activity for progress, so pair every training count with a retention, promotion or delivery measure. Second, pick a reporting cadence and a named recipient before you start collecting, because a metric with no audience quietly stops being collected. Reporting to your chief AI officer on a monthly rhythm and into the agency's broader AI reporting on a quarterly one is a reasonable default, provided you confirm what your agency's guidance actually requires.

What the federal record actually shows

Federal talent programs have produced usable evidence, and it points consistently at structure rather than salary. The tour of duty model run through the U.S. Digital Service placed more than 500 technologists across agencies on two-year appointments after its founding in 2014. Attrition during the tour ran under 10 percent, and afterwards roughly a third converted to career federal service, a third returned to the private sector carrying federal experience, and a third continued in fellowship or consulting work. The model worked because it asked for mission commitment without demanding a lifetime conversion.

Other programs sharpen different lessons. The Department of Homeland Security stood up an AI Corps in early 2024 aiming to hire approximately 50 AI specialists across its components, and its early lessons were that speed to offer is decisive, that specialized technical interviews beat generic written qualification statements, that compensation transparency closes candidates faster, and that a shared onboarding experience across components builds cohort identity. The Presidential Innovation Fellows program, running since 2012, placed approximately 150 senior technologists on 12-month assignments and is small by design, with influence per fellow well out of proportion to its size.

The Veterans Health Administration ran a different play entirely. Its AI Tech Sprint model, launched in 2019, attacked clinical and operational problems in 90-day cycles using mixed teams drawn from the department, the private sector and academia, sustaining the format for years with notable work on suicide risk prediction, radiology workflow and scheduling. The retention effect was a side effect of the design: participating staff built portable AI skills on the job, and several were promoted internally or moved into other agencies rather than out of government entirely.

The manager is the retention strategy

No agency program substitutes for a manager who clears blockers, names a growth path, and protects time for real work. The single most effective retention act Karen can take is to personally fight to get her engineers the tools and the runway to do the mission work they came for. Talent development is not an HR initiative that happens to people. It is a daily management practice, and the agencies that keep their AI talent are not the ones with the biggest budgets. They are the ones where managers treat their people's growth as part of the job rather than as an annual form.

Anti-Patterns

• Recruiting hard and onboarding not at all. Karen's original failure. Nine months of effort to close a candidate, then no structured start, no equipment, no named mentor. The hire spends the first quarter deducing how the agency works, concludes nobody was expecting them, and the recruitment cost is written off fourteen months later. Budget the first 60 to 90 days with the same seriousness you budget the search.
• Treating AI talent as generic information technology. Classifying every AI role into a generalist IT series advertises the wrong work, attracts the wrong applicants, and offers a promotion ladder pointed away from modeling and evaluation. It is also the setting in which an AI specialist ends up owning a ticket queue.
• Promotion only through management. When the sole route to a higher band runs through supervising people, you convert your strongest engineer into a mediocre manager and lose them twice. Build the technical branch of the ladder before you need it, not during someone's resignation conversation.
• The single-specialist office. Hiring one AI person into a non-technical unit and calling it a capability. Nobody reviews the work, nobody argues with the approach, and the person's skills and confidence both decay. If you cannot fund a second technical role, join the person to a cross-office community of practice on day one.
• Retention levers used only as counteroffers. Discretionary incentives are paperwork-heavy, so they get used after a resignation letter arrives. By then the person has already run the comparison and mostly decided. Decide in advance which roles justify these authorities and apply them before the letter.
• Contractors as the workforce plan. Contract staff can extend capacity, but they cannot hold institutional knowledge on the agency's behalf and cannot be the decision-makers. An AI capability that exists only inside a contract disappears when the contract does.
• Counting trainings instead of capability. Reporting courses delivered and hours logged feels like a program and measures nothing. Trainings are inputs. Retention, promotion, internal mobility and delivered systems are the outputs, and only the outputs survive a budget review.
• Assuming a competitive salary is sufficient. Pay has to clear the tolerable line, and above that line it is not what keeps people. Believing that a special rate authority has solved retention is how an agency ends up with a well-paid engineer who cannot get a development environment approved.

Practice Prompts

1. Run the diagnostic on one person. Take a single technical role on your team and answer all seven diagnostic questions in writing. Which leak is real for that person right now, and what is the smallest action this month that would change one answer from no to yes?
2. Draw your ladder. Sketch the current promotion path for a data scientist in your organization. Where does the technical branch stop? What would have to be true, and who would have to agree, for a senior practitioner to reach the next band without supervising anyone?
3. Audit your series decisions. List the occupational series your AI roles are currently classified under. For each, ask whether the qualification standards and promotion path match the work the person actually does. Which one is most misclassified, and what does that cost you?
4. Time the pipeline. Measure how many days pass between an approved AI vacancy and an accepted offer in your organization. Which authority discussed here would compress that clock, and what is stopping you from using it?
5. Find the friction. Ask a technical staff member to log, for one week, the share of their time spent on the work you hired them for. Do not interpret the result for them. Bring the number to whoever controls the friction.
6. Write the counterfactual exit interview. Assume your best AI person resigns next month. Write the exit interview they would honestly give. Then choose the one item on it you can fix before the month is out.

Reflection

Think about the last technical person who left your organization. What did the official record say, and what did the people who worked with them believe? If the two answers differ, which one is currently driving your workforce plan? Ask yourself who on your team could describe the public outcome their model serves, and who could not. Consider what your agency currently spends recruiting compared with what it spends on the first 90 days after arrival, and whether that ratio is a deliberate choice or an accident. And if the honest answer to the growth question for your strongest engineer is no, how long do you believe you have before someone else asks them the same question?

Glossary

• Dual career ladder. A promotion structure with parallel technical and management branches, so a practitioner can reach senior bands by deepening technical scope rather than by supervising staff.
• Occupational series. The federal classification code that defines a position's qualification standards, promotion path and pay structure. AI work commonly falls into information technology, statistics, computer science, computer engineering or data science series.
• Direct hire authority. An authority permitting appointment without the full competitive announcement process for occupations with a severe shortage of candidates or a critical hiring need.
• Schedule A appointment. An excepted-service hiring authority permitting direct appointment outside the competitive process, used for data science roles under OPM guidance for a defined period.
• Subject Matter Expert Qualification Assessment. A hiring method in which technical subject matter experts assess candidates through structured technical review and interviews instead of generic self-rated questionnaires.
• The 3Rs. Recruitment, relocation and retention incentives: discretionary payments tied to a written service agreement, used to close or hold a hard-to-fill position.
• Special salary rate. An authority raising the pay schedule for specified occupations and locations where standard rates cannot recruit or retain qualified staff.
• Tour of duty. A fixed-term appointment, typically a small number of years, designed to bring experienced technologists into government without requiring a permanent career conversion.
• Community of practice. A standing cross-organizational group of practitioners who share work, review each other's approaches, and provide the peer environment a single specialist otherwise lacks.
• Time to offer. The elapsed days between an approved vacancy and an accepted offer. The workforce metric most directly under a hiring manager's influence and the one candidates feel most sharply.

Related Lessons

The organizational container for much of this work is covered in Building AI Centers of Excellence, which treats the staffing model as part of a chartered mandate rather than a hiring plan. Workforce Planning for AI sits upstream, sizing the roles you should be recruiting for in the first place, and AI Talent Pipeline: Education System Alignment looks further upstream still. Cultural Transformation for AI addresses the environment a new hire lands in, and Change Management for AI Adoption covers the union and workflow conversations that shape their daily experience. For the reporting side, AI Metrics and KPIs for Government shows how workforce measures fit an agency dashboard, while Moving from Pilot to Production and Data Infrastructure for Enterprise AI describe the work your retained talent should be spending its time on. Communicating AI Success Stories is the recognition mechanism that costs nothing and holds people longer than it should.

Closing

The salary gap is real, permanent and mostly irrelevant to your retention problem. Every candidate who accepts a federal AI offer has already run that comparison and decided the mission was worth it. What breaks the deal afterwards is the discovery that the mission is not actually available: that the work is queued behind procurement, that no colleague can review it, and that the only promotion in sight requires abandoning it. Those are management failures, and management failures are the kind you can fix without an act of Congress.

Start with the diagnostic rather than the program. Seven questions per role, answered honestly, will tell you in an afternoon whether your problem is growth, peers or friction, and the three problems need different money and different allies. Then use the authorities you already have before the resignation rather than after it, classify roles for the work people actually do, and make the technical branch of the ladder visible to everyone you are trying to recruit. Karen could not have paid Devon enough to stay. She could have made the fourteen months worth staying for.

Key Takeaways

• Stop competing on salary. You cannot win the pay war. Keep compensation tolerable, then compete on mission, scale, stability, clearance value and recognition, which is where the federal value proposition is genuinely strong.
• People leave over the three leaks, not the paycheck. No growth path, no peers, and no tools or time drain talent far more than the salary gap they already accepted going in.
• Build a technical career ladder with two top branches. A path that lets people advance in pay and stature without abandoning hands-on work is the strongest retention tool available, and federal science ladders already prove the shape works.
• Classify for the work, not the department. Choosing the occupational series deliberately determines the promotion path, the qualification standards and the applicants you can reach. Generic IT classification is the most common structural mistake.
• Speed is a retention signal. A process that takes 90 to 120 days loses candidates to offers that close in days, and it teaches the ones you do hire how long everything else will take.
• Use the levers before the resignation. Special rates, the 3Rs, loan repayment and telework are discretionary, so they default to counteroffers. Decide in advance which roles justify them.
• Manufacture peers and protect time. A cross-office community of practice and defended learning time fix isolation and stagnation at almost no cost, and they work even when no single office can staff a full team.
• Diagnose before you program. Ask each role about growth, peers, tools, time, mission line of sight, recognition and onboarding. The pattern of negative answers tells you where to spend your limited energy.
• The manager is the strategy. Clearing blockers and defending the mission work people came for does more than any HR initiative, and it is the one intervention that requires nobody's approval but yours.

Frequently Asked Questions

We genuinely cannot match private offers. Is it worth recruiting senior AI people at all? Yes, but recruit on the things you can actually deliver. Candidates who take a federal AI role have almost always accepted the pay difference before the first interview, and they are buying mission, scale, stability, clearance value and public recognition. What loses them is discovering after arrival that the mission work is not reachable. Your recruiting promise and your working conditions have to be the same document, because the gap between them is what shows up in the exit interview.

Which hiring authority should we use? That depends on the role and on what your human capital office can support today, so treat the authorities in this lesson as a list to ask about rather than a menu to choose from. The general pattern is that direct appointment authorities and technical panel assessments compress the clock most, entry cohorts and fellowship programs fill different gaps than mid-career recruiting does, and named programs change over time. Confirm current availability with your agency before you build a pipeline that assumes one.

Is a retention bonus worth it? Sometimes, but almost never as a response to a resignation letter. By the time someone has an offer in hand they have already tested the market, told a few colleagues, and decided how the comparison came out. Incentives work better as a planned instrument: identify in advance the roles that are genuinely hard to fill, document why, and pair the incentive with the structural fix. Money without a growth path buys months, not years.

We have one AI person in a non-technical office. What do we do? Connect them to peers outside the office immediately, before you do anything else. A standing cross-office community of practice, a detail assignment onto a larger technical team, or a formal mentoring relationship with a senior practitioner elsewhere in the agency all supply what a single-specialist posting structurally cannot. Then be honest with yourself about whether the role is a capability or a headcount, because a lone specialist with no reviewer is also a quality risk.

How do we prove the talent program is working? Report outputs rather than activity. Time to offer, acceptance rate, first-year attrition, three-year retention, internal mobility, promotion rates and delivered systems are measures that survive scrutiny. Training hours and course counts are not, because they measure what you spent rather than what changed. Pick the cadence and the recipient before you collect anything, since an unread metric stops being collected within two quarters.

Should we just use contractors? Contract staff can extend capacity and cover surges, but they cannot be your workforce plan. Institutional knowledge accumulated by contract staff belongs to the contract, and federal decisions about an AI system need federal decision-makers who will still be there next year. The practical test is whether your agency could keep operating and improving a deployed system if the contract ended tomorrow. If the answer is no, the staffing model is the risk, not the technology.