When Beatriz Solano became Chief Information Officer of the Arizona Department of Economic Security in March 2023, she inherited 6,800 employees, a $47 million IT budget, and exactly zero data scientists. The agency had just signed a contract for an AI-powered benefits eligibility system going live in eighteen months. She had to build the internal capacity to run it, inside a pay structure where her top technical roles topped out at $98,000 while the same skills commanded $180,000 across the street at Intel. An AI strategy is only as real as the workforce that executes it.

The Gap Before the Plan

Workforce planning for AI starts with an honest skills inventory. Not a survey asking employees whether they are "comfortable with technology," but a structured skills mapping that identifies specific competencies and gaps at the role level. Think of it like a bridge inspection. Before you can reinforce a bridge you need to know which beams are load-bearing, which are cracked, and which are missing. A skills map does the same thing for a workforce, and it produces something a budget officer can act on rather than a mood reading.

Beatriz ran her mapping in two phases. First, she audited 340 position descriptions across IT and program operations, every role that touched data, managed vendors, or carried any analytical component, using two analysts over six weeks. Second, she ran a voluntary self-assessment against a twelve-competency framework adapted from the National Institute of Standards and Technology Workforce Framework for Cybersecurity, covering data literacy, process automation, model governance, and vendor management. Sixty-one percent of eligible employees completed it, and that participation rate was itself a signal about organizational readiness.

What the mapping found:

• Fourteen percent of program operations employees had advanced data skills going completely unused in their current roles.
• Forty-two percent of IT staff had vendor management experience but had never managed an AI vendor contract.
• Zero employees had formal training in algorithmic auditing or AI governance.
• Thirty-eight roles directly affected by the eligibility system had not had their position descriptions updated in more than five years.

Mapping Skills to the Classification System

A skills map that does not connect to your classification system produces insight you cannot hire against. In federal service the inventory is grounded in the Office of Personnel Management job-series architecture, so that position descriptions and grade determinations align with OPM classification standards. Data scientist work sits in the 1560 series. IT management, including cybersecurity specialties, sits in 2210. Operations research and quantitative modeling belongs in 1515. Management and program analysis spans 0343, program management sits in 0340, and contracting officer and contracting officer's representative work sits in 1102.

Misclassifying the series is a common failure mode. It undermines grade authority, shrinks the hiring pool you can draw from, and blocks career progression for the people you do recruit. For services acquisition, the 1102 series carries its own qualification requirements under the Defense Acquisition Workforce Improvement Act or the federal acquisition certification for contracting, so an AI acquisition role is not a matter of assigning a willing analyst. OPM also designates Mission-Critical Occupations, the skills judged essential to mission, which is the hook for arguing that an AI role deserves scarce hiring attention.

Inventory the existing workforce against those series and against the AI role archetypes a public AI enterprise actually needs. The federal curriculum names eleven archetypes: model developer, MLOps engineer, data engineer, AI product manager, AI safety engineer, AI program manager, AI contracting officer's representative, AI red-teamer, AI evaluator, AI policy analyst, and Chief AI Officer designee. Most agencies discover that their coverage of these archetypes is partial and accidental, spread across people who were hired to do something else and picked up the skill on their own time.

Hire, Train, Contract, or Partner

For each capability gap you have four options, and the four-quadrant decision is calibrated to appropriations stability, retention risk, and criticality rather than to preference. HIRE when the skill is specialized and persistent and you have direct-hire authority or enough competitive-examining capacity to use it; settle appropriations stability, special salary rate availability, and the retention plan before you post. TRAIN when the gap is adjacent and internal, when training dollars exist, and when time-to-competency is shorter than time-to-hire.

CONTRACT when the skill is project-specific, when the statement of work can be written under FAR Part 37 with clear deliverables, and when the agency has retained enough in-house capability to manage and inspect what comes back. The failure mode here is service-contract misuse that erodes inherently governmental functions, which OMB Circular A-76 and FAR 7.5 exist to police. PARTNER when you need a capability surge paired with career-staff growth: federally funded research and development centers, university-affiliated research centers, and details from universities or state governments under the Intergovernmental Personnel Act.

In government the constraints are layered in ways private-sector frameworks rarely account for. Arizona's State Personnel Board sets pay bands for classified positions. A data scientist role classified under the existing Information Technology Analyst III job family maxes out at $98,000. Reclassifying it requires a formal compensation study, State Personnel Board approval, and a budget amendment, which runs eight to fourteen months on a good cycle. That timeline does not fit an eighteen-month go-live date, and no amount of executive urgency compresses it.

Four government-specific factors decide the quadrant for you before preference ever enters:

1. Position classification lag. If you cannot reclassify a role fast enough, contracting or training existing staff becomes the default by elimination, not by choice.
2. Budget cycle constraints. Hiring requires an authorized position. Training can often draw from an existing operating budget. Contracting requires a procurement that may take six months. Note also that the Antideficiency Act constrains implied commitments of future-year funding for any of these investments, so a multi-year promise to a candidate is a promise you may not be able to make.
3. Union agreement scope. Assigning AI-related duties to a role may require a meet-and-confer with the relevant bargaining unit if it changes the position's essential functions. In federal service, 5 USC chapter 71 requires union notice and bargaining on impact and implementation when AI changes working conditions.
4. Mission-criticality of the skill. Permanent, core-function skills should be built internally. Narrow, time-limited skills are appropriate to contract.

Beatriz mapped her gaps to three buckets: model governance, which was permanent and mission-critical, so she trained internal staff; data engineering, which was time-limited, so she contracted it; and AI literacy across program staff, which was broad and foundational, so she trained it internally at scale.

Hiring Authorities and Talent Pipelines

Federal hiring timelines through the standard competitive-examining or merit-promotion process frequently exceed 90 to 120 days, and an urgent AI use case can lose its sponsor inside that window. That is why the authorities matter more than the job posting. Schedule A supports targeted non-competitive appointment for defined categories, including the disability appointment authority. The Cyber and IT direct-hire authority allows direct appointment into the 2210 series. Direct-hire authority generally, governed by 5 CFR part 337, lets OPM or an agency appoint non-competitively for shortage occupations.

Other instruments fit shapes that a contract does not. Expert and consultant appointments under 5 USC 3109 bring short-term specialist expertise when a procurement would be the wrong vehicle. Re-employed annuitants under 5 USC 8344 can return with dual-compensation waivers, which is how agencies recover institutional knowledge that walked out the door. Term fellowships and secondments bring high-skill capacity without consuming a classified position at all. Each authority carries distinct rules on ceiling, duration, documentation, and union notice, and using the wrong one is a personnel action that an inspector general will eventually read.

One caution about named programs. Governmentwide talent vehicles are created, renamed, consolidated, and shut down on political timelines, and a workforce plan built on one of them inherits that fragility. The federal Presidential Innovation Fellows program historically placed technologists into agencies on twelve-month terms, and some technology companies offer analogous secondment arrangements. Beatriz placed two fellows in her first year, and they built the data pipeline architecture that permanent staff then maintained. Confirm the current authorization, funding, and intake schedule of any named program before you write it into a plan.

Why You Cannot Compete on Salary, and What You Can Do

A machine learning engineer at a Phoenix technology company earns $140,000 to $200,000. Arizona's top classified pay band for comparable roles is $98,000. That gap does not close. The federal picture is the same shape: a senior data scientist on the General Schedule may be paid roughly twice as much at a peer technology company, and a senior MLOps engineer in the 2210 series is competing against total-compensation packages that federal service cannot match without deliberate use of special rates and retention incentives. Any workforce plan that assumes otherwise is wishful thinking dressed as strategy.

You are not competing with a large technology employer for talent in general. You are competing for a specific person who wants to do work that matters in a way that employer cannot offer. That is a real competition, and one you can win, but only if you know who you are looking for and you make the offer legible. What government can put on the table falls into four categories, and none of them work if they stay implicit.

Mission specificity. The Arizona Department of Economic Security administers SNAP, Medicaid eligibility, and child welfare services. The eligibility system directly affects whether families receive food assistance within the statutory 30-day window. Some technical professionals find that more compelling than optimizing advertising click-through rates. Be explicit about it in every job posting. Do not assume candidates will infer the stakes from an agency name and a series number, because they will not.

Loan repayment and forgiveness. These are two different instruments and it is worth keeping them straight. Public Service Loan Forgiveness cancels the remaining federal student loan balance after ten years of qualifying payments for borrowers in qualifying public service employment; eligibility also depends on loan type and repayment plan, so treat it as a benefit a candidate must confirm rather than one you can promise. Separately, federal student loan repayment authority under 5 USC 5379 provides up to $10,000 per calendar year and $60,000 lifetime, with a three-year service agreement.

Quantify the benefit honestly rather than quoting a headline equivalence. For a data scientist carrying $80,000 in graduate school debt, divide the cancelled balance across the years of qualifying service and let the candidate compare that number to salary themselves. Beatriz put a written loan-benefit estimate on every technical posting she controlled, which almost no government human resources office bothers to do, and it changed her applicant pool more than any recruiting event she ran.

Structural pay tools. Retention is a system, not a single incentive. Special salary rates, approved for critical positions under 5 USC 5305, address structural pay gaps in a federal labor market. Recruitment, relocation, and retention incentives under 5 CFR part 575 provide targeted payments with required service agreements attached. Telework and remote arrangements under the Telework Enhancement Act widen the labor market you can reach. Where these constitute working conditions, they are negotiated with the recognized union rather than announced.

Pensions and visible career paths. Arizona's state employee retirement system provides a defined-benefit pension that most private employers eliminated decades ago; for candidates in their thirties and forties, quantify it rather than mentioning it. Career-path transparency does similar work: a visible technical ladder through the GS-14 and GS-15 pathways, scientific and engineering track positions, and the Senior Executive Service candidate development program tells a candidate that the job is a career rather than a detour.

What Your People Actually Need to Learn

A training needs assessment is not a course catalog. It is a structured answer to one question: for each role affected by AI, what does the person in that role need to know and be able to do that they cannot do today? Answering it role by role also gives you the paper trail. Training is documented through individual development plans, funded under the Government Employees Training Act, and coordinated with the union where duty-hour training changes working conditions. Beatriz organized her needs into three tiers.

Tier 1, foundational AI literacy, all 6,800 employees. Every employee needed to understand what the system does and does not do, how to identify outputs that warrant human review, and how to escalate anomalies. Four hours of vendor-delivered e-learning at $38 per employee. Total cost $258,400, drawn from the agency's existing $2.1 million training budget, so no new appropriation was required. This is the tier agencies most often skip, on the theory that staff who do not touch the model do not need to understand it, which is exactly backwards.

Tier 2, operational AI skills, 340 IT and program analyst roles. Hands-on proficiency with the eligibility system, data quality procedures, and vendor reporting tools. Beatriz ran a twelve-week cohort program using vendor staff covered under the implementation contract's existing training provisions, which meant additional cost to the agency of zero. Training obligations you negotiate into an implementation contract before award are free. The same training bought afterward is a modification.

Tier 3, AI governance and model oversight, twelve identified roles. Algorithmic auditing, bias detection, and model performance monitoring. Beatriz sent eight employees to a six-week University of Arizona continuing education certificate program at $4,200 per person, and secured four slots in a federal Chief Data Officer Council cohort program, covering all twelve roles. Total $33,600, funded from the capital project budget as a direct implementation cost rather than from operating training dollars.

Role Redesign for AI-Enabled Work

When an AI system takes over a task, the human role does not disappear. It shifts. If you do not redesign it explicitly, employees perform redundant manual steps alongside the automated process, or they work in undefined roles while anxiety fills the gap. Role redesign is the coordinated change of the position description, the knowledge and skill requirements, the performance standards, and the training, done together rather than one at a time, and negotiated with the recognized union under 5 USC 7114 where it applies.

The eligibility system automated a data matching step that consumed 40 percent of a benefits specialist's daily time. Beatriz redesigned 280 specialist roles to redirect that time to three things: quality review of AI-generated recommendations, complex case management the system could not resolve, and proactive outreach to eligible households that had not applied. That third category was work the agency had never had capacity to do, and naming it turned a story about displacement into a story about reach.

The pattern generalizes across four role families. The program manager's old content was timelines and budgets; the new content adds understanding what AI can and cannot do, governance-board interaction, risk-register maintenance, and documented communication with oversight bodies. Frontline staff move from following procedures to reviewing AI recommendations, escalating edge cases, and documenting feedback. Compliance and audit staff move from reviewing documentation to auditing systems for bias, fairness, privacy, and security. Contracting officers move from closing acquisitions to writing statements of work with data-rights clauses and acceptance-test language for AI-specific risks.

None of that is a technical exercise alone. Each redesign runs through the Chief Human Capital Officer for performance-management integration, the equal employment opportunity office, and the reasonable-accommodation coordinator, because a changed essential function changes what an accommodation has to accommodate. Arizona's classified system requires supervisor sign-off, human resources review, and sometimes a compensation study for a position description update. Beatriz started nine months before go-live. At go-live, eleven of 280 descriptions were still under review, and those employees worked under temporary duty descriptions for sixty days.

Building Internal Capacity Over Two to Three Years

Multi-year planning follows the same logic as infrastructure construction: foundation first, structure second, systems third. In workforce terms that is literacy first, operational skill second, governance expertise third. Sequence the plan with explicit milestones and gate criteria rather than a list of intentions, because a workforce plan without gates is a wish list that quietly slips a quarter at a time. Synchronize the gates to cycles you do not control: the annual employee survey, the AI use case inventory refresh, and the budget passback.

The federal version of this plan runs over 24 months. Months 1 through 3 complete the skills inventory, the four-quadrant decision for each AI role, the retention diagnostic, and union consultation. Months 4 through 9 execute hiring under whatever authorities are available, launch the training curriculum, stand up a contractor plan aligned to FAR Part 37, and open partnership agreements. Months 10 through 15 integrate the new capacity into live AI use cases, update position descriptions through classification review, refresh union agreements where necessary, and validate retention through survey and pulse data. Months 16 through 24 scale on evidence and brief the governance board.

Beatriz worked on a longer runway because her classification cycle was slower. Her first year closed with skills mapping complete, Tier 1 training delivered to all 6,800 staff, two fellows placed, vendor-delivered Tier 2 training done for 340 roles, and two reclassification requests filed. Her second year closed with the Tier 3 cohorts complete, the first reclassification approved and one role filled at $118,000, a governance function stood up, fellowship staff transferring knowledge before rotating out, and the skills mapping repeated.

The third year was consolidation: a second reclassification approved, internal training converted from vendor-delivered to internally run at a per-cohort cost 60 percent lower, and the governance team completing its first independent model audit. The goal Beatriz set for month 36 was zero mission-critical AI functions dependent on contractors for day-to-day operations. Her budget argument was risk, not cost: every contractor-dependent function is exposed to contract expiration, vendor price increases, and the contractor's own staffing turnover, none of which the agency controls.

Two political realities shape the whole build. Change in administration resets political leadership while career executives persist, so a workforce plan that does not have a career-executive owner risks a reset every four years. And personnel actions in AI programs draw attention: employee survey results, inspector general audits, accountability-office reviews, and congressional oversight all elevate the question of whether an agency is building capacity or hollowing itself out through contractor dependence. Document the plan to an audit standard from the first month, not the month someone asks.

Anti-Patterns

• The comfort survey as skills inventory. Asking staff whether they feel ready for AI produces a mood reading, not a plan. Map competencies at the role level against a named framework and against your own classification series, or you will be hiring against a feeling.
• Treating a completed skills map as capacity. A map tells you what is missing. It does not close a gap, and it goes stale. Beatriz repeated hers annually precisely because the second map is the only evidence that the first year's investment changed anything.
• Assuming an authority guarantees a hire. Direct-hire authority, a special salary rate, and an incentive payment each remove one obstacle. None of them produce a candidate, survive an appropriations lapse, or override a classification decision. Treat every authority as one lever among several, and confirm it is still available before you build a schedule on it.
• Planning around a named talent program without verifying it exists. Governmentwide fellowship and detail programs are created, consolidated, and terminated on political timelines. A plan that assumes intake will be open next year is a plan with an unfunded gap in it.
• Starting role redesign when the system goes live. Classification review, union consultation, and accommodation coordination do not compress to meet an implementation schedule. Late starts leave people working under temporary duty descriptions during the exact window when adoption is decided.
• Contracting the core and calling it a workforce plan. Contracting a permanent, mission-critical capability does not build it. It rents it, at a price that resets at every option year, and it risks eroding inherently governmental functions the agency is required to keep.
• Quoting a benefit as a salary equivalent. Loan forgiveness, pension value, and incentive payments are real, but a headline "worth more than X in salary" claim rarely survives the division. Publish the inputs and let the candidate do the arithmetic, because they will do it anyway.

Practice Prompts

• Run the inventory. Take every role in your organization that touches data, manages a vendor, or carries an analytical component. Map each to its classification series and to the AI role archetypes listed earlier. Where does your coverage stop?
• Apply the four quadrants. For your three largest capability gaps, decide hire, train, contract, or partner. Write down which government-specific constraint made the decision for you, and what you would have chosen with no constraints at all.
• Time the authorities. For one role you would hire tomorrow, document the authority you would use, the ceiling and duration it carries, the union notice it requires, and the realistic elapsed time from decision to start date.
• Build the three-tier training plan. Assign every affected employee to a foundational, operational, or governance tier. Identify the funding source for each tier separately, and check which tier your existing vendor contract already covers.
• Redesign one role end to end. Pick a frontline role your AI system will change. Draft the revised position description, the new performance standards, the training that supports them, and the consultation steps required before any of it takes effect.
• Write the retention diagnostic. List every person whose departure would stall an AI use case. For each, name the specific retention lever available to you and the date by which it must be exercised.

Reflection

• If your most senior AI-capable employee resigned this week, which use case stops? What does that tell you about where you actually built capacity and where you only staffed a project?
• Which of your AI capabilities are you renting from a contractor, and what happens at the end of the current option year if the price changes or the contractor's own staff turn over?
• Your last technical job posting: did it name the mission consequence of the work and quantify the benefits package, or did it list duties and a grade?
• Who owns your workforce plan after the next change in political leadership? If the answer is a political appointee, what is your continuity plan?
• What would an inspector general find if they asked for the documentation behind your last three AI-related personnel actions?

Glossary

• Skills mapping. A structured, role-level inventory of specific competencies and gaps, distinct from a self-reported readiness or comfort survey.
• Mission-Critical Occupation. An Office of Personnel Management designation identifying skills essential to an agency's mission.
• Direct-Hire Authority. Authority for OPM or an agency to appoint non-competitively for shortage occupations, governed by 5 CFR part 337.
• Cyber and IT direct-hire authority. The authority permitting direct appointment into the 2210 IT management series.
• Schedule A. A non-competitive appointment authority covering defined categories, including the disability appointment authority.
• Special Salary Rate. An approved higher pay scale for specific positions under 5 USC 5305, used to address structural pay gaps.
• Position classification. The assignment of a role to an occupational series and grade against published standards; changing it typically requires review, approval, and sometimes a compensation study.
• Meet-and-confer. Consultation with a recognized bargaining unit required when a change alters a position's essential functions or working conditions.
• Role redesign. The coordinated change of position description, skill requirements, performance standards, and training in response to AI-driven workflow change, negotiated with the recognized union where applicable.
• Individual development plan. The documented record of an employee's training and development commitments, and the vehicle through which training investment is tracked.
• Inherently governmental function. Work that must be performed by government employees rather than contractors; the boundary that service-contract planning is required to respect.

Related Lessons

• Developing an Organizational AI Strategy sets the strategic frame this workforce plan executes; build the plan against the strategy, not alongside it.
• AI Maturity Assessment gives you the organizational baseline that tells you whether a literacy tier or a governance tier is your binding constraint.
• Cultural Transformation for AI covers the adoption work that role redesign depends on; a redesigned role in a hostile culture is a form on a shelf.
• Stakeholder Management and Communication is where the union, oversight, and leadership conversations in this lesson are treated in depth.
• AI Talent Development and Retention extends the retention system sketched here into a full program design.
• Workforce Transformation at Scale takes the same problem to the enterprise and multi-agency level.
• AI Strategy for Different Government Contexts explains why the federal, state, and local versions of this plan use different levers.

Closing

Beatriz did not solve the pay gap, and neither will you. What she did was stop treating the pay gap as the whole problem. She measured what she had, decided each gap deliberately instead of by default, used the authorities that were actually available to her, made her non-salary offer legible in writing, and started the slow classification work early enough that it finished. Three years later the agency ran its own eligibility model with its own people, and the contractors who built it had gone home.

The discipline that made it work was documentation. Every quadrant decision, every authority used, every training dollar and every position description change was written down at the time, against a stated milestone, in a form an auditor could follow. That habit costs a few hours a month and it is the difference between a workforce plan and a set of good intentions that nobody can reconstruct after the leadership changes.

Key Takeaways

• Skills mapping must be role-specific and classification-aware. A self-assessment against a named competency framework, mapped to your own occupational series, produces data you can hire against. A general comfort survey does not.
• Decide each gap across four quadrants: hire, train, contract, or partner. Calibrate to appropriations stability, retention risk, and criticality, and keep enough in-house capability to inspect whatever you contract.
• Government-specific constraints usually decide for you. Classification lag, budget cycle timing, union notice and bargaining obligations, and procurement timelines determine which options are genuinely available inside your implementation window.
• Know the authorities before you know the candidate. Direct-hire authority, Schedule A, expert and consultant appointments, re-employed annuitants, and intergovernmental details each carry distinct rules on ceiling, duration, documentation, and union notice.
• Reframe the salary competition rather than trying to win it. Mission specificity, quantified loan and pension benefits, special rates and incentives under their governing authorities, and a visible technical career ladder are real differentiators, but only when stated concretely in every posting.
• Build training in three tiers with three funding sources. Foundational literacy for all staff from operating budget, operational skills for affected roles from vendor contract provisions negotiated before award, and governance expertise for a small designated group funded as a direct implementation cost.
• Start role redesign months before go-live and run it through the full consultation chain. Position description updates, human capital review, union consultation, and accommodation coordination do not compress to meet an implementation schedule.
• Sequence the multi-year build with gates, not intentions. Foundation, then structure, then systems, with milestones tied to the survey, inventory, and budget cycles you do not control.
• Argue the build as risk reduction. Every contractor-dependent function is exposed to contract expiration, price increases, and the contractor's own turnover. That framing survives a budget hearing better than a cost comparison does.
• Repeat the skills mapping and give the plan a career owner. The second map is your only feedback signal, and a plan owned by a political appointee resets with the next administration.

Frequently Asked Questions

We have no data scientists and no ability to create a position. Where do we start?

Start with the inventory, because it is free and it frequently finds capacity you already own. Beatriz found that fourteen percent of her program operations staff had advanced data skills going completely unused in their current roles. After that, work the options that do not require a new classified position: training existing staff, term fellowships and secondments, expert and consultant appointments, and details from a university or another level of government. A new position is the slowest lever, not the first one.

How do we compete when a private employer pays roughly double?

You do not compete on the number. You compete for the subset of candidates for whom mission, benefit structure, and career path change the calculation, and you make that offer explicit in writing. Quantify the loan benefit and the pension rather than mentioning them, name the mission consequence of the work in the posting, and use the structural pay tools your authorities allow. Then accept that some candidates will decline, and plan your pipeline for that rate rather than around it.

Should we contract the AI capability and skip the workforce build?

Contract what is genuinely project-specific and time-limited. Do not contract a permanent, mission-critical capability, for two reasons. First, service-contract arrangements are constrained where the work approaches inherently governmental functions. Second, every contractor-dependent function is exposed to contract expiration, price increases, and the contractor's own staffing turnover. You also need enough in-house capability to manage and inspect the deliverable, which means some internal build is a precondition for contracting well.

When do we have to involve the union?

Earlier than most program managers expect. Where AI changes working conditions, notice and bargaining on impact and implementation are obligations, not courtesies, and role redesign that alters a position's essential functions is exactly that kind of change. Retention arrangements that constitute working conditions are negotiated as well. Building the consultation into the schedule at the design stage costs weeks; discovering it after a redesign has been announced costs the redesign.

How long should the plan run?

Long enough to include a full classification cycle and at least one repeat of the skills mapping. The federal curriculum sequences a 24-month plan with gate criteria at three, nine, fifteen, and twenty-four months. Beatriz worked on a longer runway because her state classification process was slower. The right answer is set by the slowest process you depend on, not by the length of your implementation contract.