Priscilla Osei stood in front of 47 employees at the Oregon Department of Revenue in January 2024 and asked who had used any AI tool in the past month. Three hands went up. She had just signed a $1.2 million contract for an AI-assisted tax document processing system, and the go-live date was nine months away. Nine months to turn three hands into forty-seven, in an organization where the people whose cooperation she needed had watched technology projects arrive and leave before, and had good reasons to wait and see.

Why civil service culture resists differently

Private sector AI resistance is mostly about job security and speed. Public sector resistance runs deeper, braiding together accountability culture, institutional memory and the specific logic that governs civil service careers. A private employee who resists new tools risks being let go. A civil service employee protected by collective bargaining agreements and tenure rules has no such pressure. That protection is legitimate and valuable, because it shields employees from political interference, but it also means you cannot rely on the levers that technology executives use.

Three patterns appear repeatedly in government agencies, and none of them is irrational. They are worth naming precisely, because each one implies a different response and a leader who treats all three as generalized fear will address none of them.

1. Exposure to public records. Freedom of information requirements mean every email, chat log and decision record can become public. Employees worry that AI-generated outputs, especially wrong ones, will surface in a records request and attach to their name. They are not wrong to worry.
2. The audit mindset. Government workers are trained to document decisions, follow procedures and never deviate without written authorization. AI tools often feel like a deviation, and there is no procedure number for asking the chatbot.
3. Equity obligation awareness. Public agencies serve everyone, including the most vulnerable residents. Employees who have sat through bias training and accessibility reviews know that AI systems can fail specific populations, and their skepticism is often more sophisticated than leadership realizes.

These are the trained instincts of people who have watched agencies get audited, sued and featured on the front page. Risk aversion is rational in an environment where mistakes produce oversight findings, hearings and news coverage. Process orientation is not a personality trait; it is encoded in administrative procedure law, paperwork reduction requirements, privacy statute, the acquisition regulation and information security law. Employment concern is concrete because position management rules, reduction in force procedures under 5 CFR part 351 and union agreements all attach to changes in duties. Treating resistance as obstruction misses the intelligence inside it.

Culture decides whether any of the rest works

Culture is the invisible variable that determines whether an AI initiative delivers value or becomes the next cautionary case study. You can hold an authorized cloud environment at the right impact level, a published privacy impact assessment, a completed authority to operate package, a finished risk assessment under current AI guidance and a green-lit cost benefit analysis, and still fail entirely if the component that owns the work does not accept the tool. None of those artifacts is wasted, and none of them is a substitute for the thing this lesson is about.

Culture shapes five outcomes that are observable and, with effort, measurable. Whether people trust the system's output enough to use it as a first-pass check. Whether they see the tool as a threat to their job or as an amplifier of their mission. Whether they accept workflow redesign that has been negotiated through the union. Whether they supply the feedback the system needs in order to improve, which is the measurement function of any risk framework you have adopted. And whether the organization can absorb a failed pilot, learn from it, and continue rather than quietly abandoning the whole programme.

Culture is also not uniform across an agency, which is why an agency-wide change plan usually lands unevenly. An examination culture, where the job is to find the discrepancy, responds differently to an assistive tool than a claims processing culture organized around throughput and accuracy targets. A field operations culture judges a system by whether it holds up at three in the morning with no connectivity. A medical professional culture asks about evidence, liability and clinical judgment. An oversight-driven culture asks first what the finding would say. A grant stewardship culture asks about the recipients. Diagnose at the component level or your plan will be right on average and wrong everywhere.

The agencies that have succeeded with public-facing AI and digital services generally treated cultural transformation as a first-class deliverable rather than as communications support attached to an engineering programme. A national direct tax filing service launched in 2024 and a shared government login service both required workforce and public trust work at the same scale as the technical build. That framing has a budget consequence worth arguing for early: if culture work is a deliverable, it has an owner, a schedule and a line in the plan, and it survives the first quarter when the engineering runs late.

The memory of failed modernizations

The median federal employee tenure exceeds a decade, which means the audience for your AI announcement carries the memory of every technology programme that came before it. Technology skepticism is rational for staff who have watched six consecutive modernization programmes fail across their careers. The cases are not obscure: cost growth on a major law enforcement case management system, the termination of a defense human resources programme, the cancellation of an agriculture farm services modernization, the termination of an air force enterprise resource planning effort, and the long struggle to modernize core tax processing.

Each of those left cultural scars that affect every subsequent initiative, and your programme is received against that memory whether or not anyone mentions it in the room. The useful response is not to argue that this time is different, which is exactly what the last five programmes said. It is to name the history explicitly, state what you are not doing, and let a short sequence of completed, measured steps make the argument that a speech cannot. Trust here is rebuilt by finishing things, and the first finished thing matters more than its size.

The legal frame you are working inside

Federal cultural transformation is harder than the private sector version for structural reasons rather than accidental ones. Civil service protections mean performance management is deliberately slow, and that is a feature protecting merit system principles rather than a bug. Collective bargaining under 5 USC chapter 71 means that changes in working conditions, including AI-driven workflow changes, require statutory notice and bargaining with recognized unions. Engagement with union representatives under 5 USC 7114 is a legal obligation with a defined process, not a courtesy extended when convenient.

Public scrutiny closes the frame. Records requests, inspector general audits, oversight reviews and legislative hearings mean that every internal artifact of a cultural change effort, from survey instruments to all-hands slides to consultation notes, can become public. The workforce survey your agency already runs is administered annually and its results are publicly ranked, which creates both an accountability mechanism and a reputational stake. Change also happens across administration cycles that reset political leadership while career executives remain, so a transformation designed to survive transition has to be owned by people who will still be there.

The renovation analogy

Think of AI adoption in a public agency like a building renovation carried out while the building stays open. You cannot close the motor vehicles office for six months. Constituents need services every day, and employees have to keep processing cases, answering phones and meeting statutory deadlines at the same time as they learn new tools. Nothing about the transition happens in a protected environment, and any plan that assumes otherwise is describing a different organization than the one you work in.

The analogy holds all the way through. Renovations require permits, which are your procurement approvals. They disturb people working on floors that are not being touched yet. They generate dust and noise, which is confusion, extra steps and two systems running in parallel. And they succeed only when the people who work in the building trust that the contractor has thought about their safety. Priscilla used the framing explicitly, calling the rollout the renovation in every staff meeting, and when the new system ran alongside the old paper process for 60 days she called that construction scaffolding. It gave employees a mental model matching their actual experience: temporary disruption, a clear endpoint, and a building that would work better afterwards.

Diagnose before you try to change anything

Assess the culture you have before attempting to change it, and use the data your agency already collects. Annual workforce survey results, published workplace rankings, inclusion indices and exit interview data held by your human capital office give you a baseline without commissioning anything. Supplement them with listening sessions scheduled in coordination with recognized unions, focus groups segmented by component, tenure and role, and structured interviews with career executives whose continuity across administrations defines the cultural substrate you are working in.

Then rate each barrier on two axes: how intense it is and how broadly it is held. The three combinations call for different responses. An intense but narrow barrier, such as resistance concentrated in a single unit, is addressed by focused champion development inside that unit. A broad but weak barrier is addressed by communication. A barrier that is both broad and intense requires executive sponsorship, bargaining and phased implementation, and attempting to communicate your way out of one is the most common way a change plan fails while looking busy.

Four barriers and the response each one needs

The barriers are specific, and so are the remedies. Generic change management fails here because it addresses an average employee who does not exist, while the four patterns below each have a different owner, a different evidence requirement and a different timeline. Work out which of them you actually have before choosing an intervention, because the wrong remedy applied confidently does more damage than no remedy at all.

• Job security. Rational given reduction in force procedures and the real workforce implications of automation. The response is a written position management commitment, a retraining and redeployment plan designed with your human capital office, and transparent metrics on filled positions, converted positions and vacancies.
• Trust in technology. Rational given the modernization history described above. The response is phased rollout with explicit statements of what you are not doing, measured pilots with baseline comparisons, and publication of incident reports rather than quiet handling.
• Workflow change. Real, because statutory and procedural workflows encode hard-won institutional knowledge that is not written down anywhere else. The response is co-design with frontline staff, union representatives and line supervisors, rather than external consultants working around them.
• Change fatigue. Real for staff who have lived through repeated modernization programmes without seeing one finish. The response is thoughtful sequencing, protected time for training, and completing one phase before initiating the next.

Building champions from within

External consultants cannot build internal culture, because they leave. Champions have to come from the ranks, and in a government setting they have to come from three populations at once. Career senior executives matter most for continuity, because their tenure across administrations is what sustains a multi-year programme when political leadership turns over. Recruit them onto the governance body, have them co-author the charter, and make them the face of internal communication. Career ownership improves the odds of surviving a transition considerably; it does not guarantee it, and a programme with no written record of its decisions will not survive one regardless of who sponsors it.

Recognized union stewards are the second population. Engaging them as partners under the statutory consultation process, early and with real information, turns many of the hardest conversations into negotiations with a clear record. That is a substantially better position than an unrecorded disagreement, though it is not the same as agreement: partnership changes the forum and the quality of the conversation rather than guaranteeing the outcome. The third population is early-career technical staff brought in through fellowship, digital service and direct-hire routes. Pair them with career leaders, or you get the standard failure in which technical brilliance collides with institutional reality and both sides conclude the other is the problem.

Priscilla built hers from all three. She identified twelve employees across four divisions who showed curiosity about the new system during the first two weeks of demonstrations and pulled them into a working group she called the Renovation Crew, giving them four hours a week of protected time formalized in their performance plans to learn the system, document use cases and report problems back to the vendor. The crew cost nothing extra, because the time was already budgeted. What changed was how that time was authorized and recognized.

Her selection criteria are worth copying. Peer credibility matters more than technical skill, and the best champions are often mid-career employees with ten or more years at the agency whose colleagues trust their judgment about what actually works. Willingness to say that something does not work yet is essential, because champions who only celebrate the tool lose credibility fast. Representation across bargaining units matters, since champions drawn only from management send an adversarial signal. And at least one champion should have started as a skeptic, because converted skeptics persuade other skeptics better than believers do.

By month four the Renovation Crew had grown to nineteen employees. They had documented 34 specific use cases, flagged 8 system errors that the vendor fixed before go-live, and trained 180 colleagues in informal 30-minute lunch sessions. The formal training budget covered external vendor sessions. The internal sessions cost lunch, about $2,400 in total, which was the least expensive line in the entire project and arguably the one that moved the adoption numbers most.

Communication that works inside government

Government communication culture is formal and hierarchical. Memos get filed and announcements go through official channels. This is not a flaw, because it creates the paper trail that accountability requires, and AI change communication has to work inside that culture rather than around it. The practical test is that every message you send should survive a records request, an inspector general review and an oversight hearing without embarrassment. If a slide would look bad quoted in a hearing, rewrite it before you present it, not after.

Anchor every announcement to a statutory obligation. Telling staff that a system will help the agency meet its statutory response requirement is more persuasive than telling them it will make their job easier. Employees know the laws they work under, and connecting AI to legal obligations frames it as mission-critical rather than discretionary. Priscilla anchored hers to the response deadline her division works under, citing the specific statutory provision in the announcement.

Address union leadership before addressing staff. In Oregon, the American Federation of State, County and Municipal Employees represents most of the Department of Revenue's clerical and processing staff. Priscilla met with local union leadership three months before the all-staff announcement, sharing the contract, the timeline and the workforce impact analysis. The union did not endorse the project, but it did not oppose it either, and that neutrality was essential. Sequencing matters here for legal as well as practical reasons, since changes to working conditions carry statutory notice obligations.

Put failure scenarios in writing first. Employees trust leaders who have thought about what goes wrong. Priscilla distributed a two-page error protocol before go-live specifying exactly what an employee should do if the system produced a wrong result: which form to complete, which supervisor to notify, and how the incident would be reviewed. That document did more for adoption than any promotional material, because it answered the question every skeptic was actually asking.

Use the budget cycle as a timeline anchor. Fiscal year transitions are natural change moments in government. Priscilla aligned full go-live with the start of the new fiscal year in July, and employees understood that framing immediately because the new budget year was already a reset point in their mental calendar.

Tailor the register to the audience. Frontline staff need operational clarity: how the work changes, from what date, what training is provided, how exceptions are escalated and where feedback goes. Line supervisors need leadership tools, including talking points aligned to the union agreement. Middle managers need to know how AI outputs are reviewed and how their quality assurance metrics change during transition. Senior executives need strategic alignment and the decisions only they can make. Unions need scope, timeline, training investment and the bargaining schedule. And the public needs plain language: what the system does, what it does not do, how appeals work and where to file a complaint.

Measuring cultural readiness

You cannot manage what you do not measure, and cultural readiness is harder to quantify than system uptime without being impossible. Priscilla tracked four indicators monthly. None required a consultant or a new platform. They required someone willing to count things and report them honestly, including the months when the numbers went the wrong way, which is the part that makes the exercise credible to the people being measured.

1. Voluntary usage rate. The system logged which employees chose the AI-assisted workflow over the legacy workflow for identical task types. In month one, 8 percent of eligible employees chose the AI workflow. By month seven that figure was 61 percent, with no mandate in place.
2. Error report submission rate. If employees are afraid to report AI errors, the error log goes quiet while the errors continue. A healthy culture produces a noisy error log in the early months, so Priscilla tracked whether reports rose, which indicates trust in the reporting channel, or fell, which is worth investigating rather than celebrating.
3. Champion network reach. Each month the Renovation Crew reported how many colleagues they had directly assisted, which measured whether the peer learning network was actually spreading or had stalled at its original members.
4. Supervisor escalation comfort. A quarterly survey asked how confident staff were that they could escalate an uncertain AI result without negative consequences. The target was 80 percent expressing confidence. The agency reached 74 percent by go-live and 83 percent by month three after launch.

A fuller federal treatment tracks five dimensions quarterly, with thresholds the agency sets for itself and uses to gate progression decisions rather than as external standards. Awareness measures the share of affected staff who can explain what the system does, what it does not do and how it changes their work, with a first-year target of 80 percent and a second-year target of 95 percent. Trust measures net sentiment in a short pulse survey, with a first-year target of 60 percent net positive and a commitment to track union member sentiment separately.

Participation measures completion of required training alongside voluntary involvement in feedback, design sessions and champion networks, with first-year targets of 70 percent required completion and 30 percent voluntary participation. Feedback volume counts substantive issue reports, feature requests and incident reports per month per 100 users, on the principle that a healthy system generates substantive feedback and silence is a warning sign rather than a success signal. Retention measures separation rates in AI-relevant roles, with a target of 95 percent and separate tracking for executive, senior general schedule and early-career technical populations.

The equity obligation as an asset

Public agencies have something private companies mostly lack: a legal obligation to serve all residents equitably. Most agency leaders treat that obligation as a constraint on AI adoption, something that slows approvals and adds review steps. Priscilla treated it as a feature. Her department serves filers who speak 47 languages, filers with disabilities, and filers with no internet access who submit paper returns, and she put equity review of the system on the Renovation Crew's agenda from the first week rather than as a gate before launch.

When the system showed a higher error rate on returns from filers with non-English-language names, the crew caught it during the parallel run, before go-live. The vendor fixed the training data, and the fix took six weeks. That catch became the most important story Priscilla told about the project, because it showed employees doing something the system could not do for itself. It reframed the project as an expression of the agency's equity commitment rather than a threat to it, and it gave skeptics a reason to engage rather than withdraw.

Be careful about the lesson drawn from it. One catch by attentive reviewers is evidence that the review layer works, not proof that the system is safe for everyone. Human review finds what reviewers happen to look at, and a parallel run surfaces disparities visible in the sample and time window you ran. The honest framing is that equity review caught a real problem this time and needs to keep running afterwards, with its own sampling and its own reporting, because a review process praised for one success is exactly the kind that quietly becomes a formality.

A twelve month plan that synchronizes both streams

The cultural and technical streams have to move together, because a system that arrives before the culture is ready gets used badly and a culture programme with no system to point at loses momentum by the third quarter. A twelve month sequence used in federal programmes runs in four phases. Months one and two diagnose the culture through survey analysis, listening sessions, union pre-consultation and leader interviews. Months three and four stand up the change management plan with the human capital office, establish the champion network across executives, unions and early-career technical staff, and begin communication.

Months five through eight run a bounded pilot with full human review in place, train staff and convene governance body reviews at defined decision points. Months nine through twelve scale on the evidence, expand training, complete bargaining on implementation for the next phase, and publish results into the agency's public AI use case inventory and its annual performance reporting. The phases add to twelve months by design, and the sequencing is the point: nothing scales before there is evidence, and no bargaining obligation is discovered after the fact.

Anti-Patterns

• The top-down mandate. Announcing required usage without consultation, procedures or bargaining. Employees comply minimally and resentfully, the usage statistics look fine, and nothing about the work actually improves. The pattern contributed to trouble in more than one major federal modernization. Executive sponsorship paired with bottom-up champion networks and genuine union partnership is the alternative.
• Training treated as the whole answer. Training solves skill gaps. It does not solve trust gaps, and a completion rate is not an adoption rate. An agency can train every employee to the last percentage point and change nothing about whether people choose the tool when they have a choice, which is why voluntary usage is the number worth watching.
• The values statement as a substitute for change. A published commitment to innovation, a signed executive message and a launch event are cheap and satisfying, and none of them alters what happens at a desk on a Tuesday. Treat them as the announcement of work rather than as the work, and be suspicious of any plan whose first milestone is a statement.
• Equity review as a one-time gate. Running a fairness check before launch, finding something, fixing it, and treating the question as closed. A single catch is evidence that the review works, not proof the system is safe for everyone. Give the review its own ongoing sampling, cadence and reporting line, or it will become a formality that produces the same clean result every quarter.
• Human review as a rubber stamp. Requiring that a person approve every AI output, then measuring approvals rather than disagreements. If the reviewer approves nearly everything, you have added a step and a false assurance rather than a control. Track the rate at which reviewers actually change or reject an output, and investigate when it approaches zero.
• Silo adoption. One bureau deploys AI without coordinating with peer bureaus or shared services, producing data lock-in and integration costs that a budget examiner will eventually notice. Cross-component governance and alignment to the agency's data strategy is the fix, and it is far cheaper before the second system than after it.
• Pilot perpetuity. Running pilots indefinitely with no cutover to production, which feels safe and consumes years. It has been the proximate cause of cost growth in several major programmes. Pre-commit the cutover criteria and the evidence that would satisfy them before the pilot begins.
• Abandonment at transition. An incoming administration pauses all AI activity pending review and the programme never restarts. This is predictable, so plan for it: career executive ownership, governance documentation ready for a transition briefing, and explicit alignment to a mission that outlasts any administration.
• Leading with cost savings in front of the people whose jobs might go. Even where headcount reduction is a genuine goal, opening with it ends the conversation and poisons everything that follows. Be honest when asked, but do not make efficiency the headline to an audience that hears it as a threat.
• Skipping procurement transparency. Employees who learn about a major contract from a rumour rather than from leadership conclude that leadership is hiding something, and they are then correct to discount everything else they are told. Share the contract, the timeline and the workforce impact analysis before you need cooperation.
• Ignoring the calendar. Setting go-live during budget hearings or a legislative session puts a major change in direct competition with the thing everyone's attention is already committed to. Fiscal year transitions are free momentum. Use them.

Practice Prompts

1. Count the hands. Ask Priscilla's question in your own organization: who has used any AI tool for work in the past month? Then ask the people who did not what would have to be true for them to try. Do not defend the tool while they answer.
2. Name the scars. List the technology programmes your organization has lived through in the last decade and what people say about each one. Which of those memories will your AI initiative be measured against, and what would you have to finish first to change the comparison?
3. Map two axes. Take the three strongest objections you have heard and place each on the intensity and breadth grid. Which is narrow and intense, which is broad and weak, and does your current plan apply the matching response to each?
4. Draft the error protocol. Write the two-page document that tells an employee exactly what to do when the AI system produces a wrong result. Which step in it does not yet have an owner? That gap is what your staff are actually worried about.
5. Sequence the union conversation. Identify the working condition changes your AI project will produce and the notice or bargaining obligations they trigger. What is the latest date you could start that conversation and still be early rather than late?
6. Pick four indicators. Choose the cultural readiness measures you could track monthly with tools you already have. For each, decide in advance what direction of movement would be bad news, and who would hear about it first.

Reflection

Think about the most credible skeptic in your organization, the person whose objection carries weight because their judgment has been right before. What is their actual concern, and could you state it in their words rather than yours? Consider whether your last major change was communicated to unions before staff or after, and what that sequence signalled to everyone who noticed. Ask what your organization currently counts as evidence of adoption, and whether that number would move if nobody voluntarily used the tool at all. And if the honest answer is that your plan's first milestone is an announcement rather than a finished, measured piece of work, what would you have to reorder to make the first thing people see be something that shipped?

Glossary

• Organizational culture. The beliefs, attitudes and practices that shape how an agency actually works, as distinct from how its documented procedures say it works.
• Cultural barrier. A characteristic of the organization that impedes AI adoption. In a government context these are frequently rational responses to real accountability structures rather than irrational obstruction.
• Champion. An individual with peer credibility who influences colleagues and sustains a programme. In government the three populations that matter are career senior executives, recognized union stewards, and early-career technical staff.
• Cultural readiness. A measured assessment of awareness, trust, participation, feedback volume and retention, tracked on a fixed cadence against thresholds the agency sets for itself.
• Change fatigue. The rational response of staff who have experienced repeated reorganizations and modernization programmes that were never completed, and who therefore withhold effort from the next one.
• Voluntary usage rate. The share of eligible staff who choose a new workflow when an alternative remains available. The most honest available adoption measure, because it cannot be produced by a mandate.
• Error protocol. A short written procedure specifying what a staff member does when an AI system produces a wrong result: what to record, whom to notify, and how the incident is reviewed.
• Parallel run. A period during which a new system operates alongside the process it will replace, so that outputs can be compared and disparities caught before the old process is retired.
• Notice and bargaining obligation. The statutory requirement to inform and negotiate with recognized unions over changes to working conditions, which AI-driven workflow changes commonly trigger.
• Position management commitment. A written statement of what will happen to positions and duties as automation is introduced, including retraining and redeployment plans and the metrics that will be reported.

Related Lessons

Change Management for AI Adoption is the closest companion to this lesson and goes further on sequencing and stakeholder mechanics, while Stakeholder Management and Communication develops the audience-specific communication work in more depth. AI Talent Development and Retention covers what happens to the technical staff your culture has to hold, and Workforce Planning for AI sizes the roles the transformation will need. Developing an Organizational AI Strategy and AI Roadmap Development provide the plan the twelve month sequence executes, with AI Maturity Assessment supplying the honest baseline underneath it. Building AI Centers of Excellence describes the organizational form that usually carries this work, Building an AI Quality Culture extends the trust and error-reporting themes, and AI and Equity: Reaching All Communities develops the equity obligation this lesson treats as an asset.

Closing

Cultural transformation in government is not a softer version of the technical work. It is the part that determines whether the technical work is used, and it runs on the same evidence discipline: a diagnosis before an intervention, a named response for each specific barrier, measures chosen in advance, and honest reporting when the numbers disappoint. Nothing in this lesson guarantees adoption. Training does not, executive sponsorship does not, a values statement certainly does not, and a completed checklist has never once made a wrong output right.

What does move the numbers is unglamorous. Diagnose with the data you already have. Engage the unions before you engage the staff. Write down what happens when the system is wrong, before it is wrong. Give a handful of credible colleagues protected time and permission to say that something does not work yet. Then count the things that would embarrass you and report them anyway. Priscilla had nine months and three hands, and what changed the count was not the contract she had signed. It was a working group with four protected hours, a two-page error protocol, and a bias catch that her own employees found before anyone outside the building did.

Key Takeaways

• Civil service resistance has specific, rational causes. Records exposure, audit culture, equity obligations, modernization scars and real employment rules produce resistance patterns that differ from private sector skepticism and deserve specific responses.
• Diagnose on two axes before intervening. Intense and narrow barriers need focused champion work, broad and weak barriers need communication, and broad and intense barriers need sponsorship, bargaining and phasing. Communicating at the third type is the most common way a plan fails while looking busy.
• Champions come from three populations. Career executives supply continuity across administrations, union stewards convert conflict into recorded negotiation, and early-career technical staff supply capability. Pair them, give them protected time in writing, and include at least one converted skeptic.
• Engage unions before staff. Changes to working conditions carry statutory notice and bargaining obligations, and neutrality obtained early is a win. Opposition that forms because people learned about a contract from a rumour is very hard to reverse.
• Error protocols outperform promotional material. A short document naming exactly what to do when the system is wrong answers the question skeptics are actually asking, and it is the cheapest trust-building artifact available.
• Voluntary usage is the honest adoption measure. Training completion rates measure what you spent. Which workflow people choose when they have a choice measures whether anything changed.
• A noisy error log is a good sign. Silence usually means people are afraid to report rather than that nothing is going wrong, so treat a falling report rate as something to investigate rather than to celebrate.
• Equity obligations are a strategic asset and not a one-time gate. Framing AI adoption as a way to better meet equity duties gives skeptics a reason to engage, but one catch by human reviewers is evidence the review works rather than proof the system is safe for everyone.
• Plan for transition and for the calendar. Career ownership plus documentation ready for a transition briefing is what survives a change of administration, and aligning go-live with the fiscal year rather than a legislative session is free momentum.

Frequently Asked Questions

Our staff are unionized. Does that make AI adoption impossible? No, but it makes the sequence non-negotiable. Changes to working conditions carry statutory notice and bargaining obligations, so union engagement is a legal step rather than a courtesy, and it has to come before the all-staff announcement rather than after it. Handled early and with real information, including the contract, timeline and workforce impact analysis, it produces a recorded negotiation. Handled late, it produces an opposition position that is far harder to move than the original concern.

We trained everyone and adoption is still low. What went wrong? Probably nothing about the training. Training addresses skill gaps, and low voluntary adoption after full training usually indicates a trust gap instead: staff are unsure what happens when the system is wrong, whether an error attaches to their name, or whether using it is formally sanctioned. Look at whether a written error protocol exists, whether anyone has escalated an AI result without consequence, and whether supervisors have talking points. Those are trust artifacts, and they move numbers that another course will not.

How do we handle the employee who is loudly and persistently skeptical? Find out whether they are right. The most credible skeptics in government agencies are frequently people whose judgment has been vindicated before, and their objection often contains information your plan needs. Ask them to state the failure mode they expect, then put it in the error protocol or the pilot evaluation. Converted skeptics are also the most persuasive champions you can have, and a skeptic who was listened to is much more likely to become one.

Should we mandate usage? Not before the procedures exist. Mandating usage without documented procedures produces minimal, resentful compliance and destroys the one measure that would have told you the truth, since voluntary usage rates stop existing the moment a mandate does. Build the procedures, the error protocol and the champion network first, watch voluntary adoption, and consider a mandate only where a genuine consistency or compliance requirement makes uniform use necessary rather than merely convenient.

What do we do when the pilot finds a fairness problem? Exactly what Priscilla's team did, and then one thing more. Fix it before go-live, document what was found and what was changed, and tell the story internally, because it demonstrates that the review layer is doing real work. The additional step is to keep the review running afterwards with its own sampling and reporting cadence. A fairness check that runs once before launch tells you about the population and time window you sampled, not about the system's behaviour next year.

How do we keep this alive through a change of leadership? Assume the pause is coming and prepare for it. Career executive ownership rather than political sponsorship, decisions documented well enough that a new leader can read the reasoning rather than reconstruct it, alignment to a mission that outlasts any administration, and a transition briefing package prepared before it is requested are the practical defenses. None of them guarantees continuation. All of them shorten the pause and make restarting a decision rather than a rebuild.