When the State Department of Revenue announced its AI-assisted audit selection pilot, the union representing 3,400 tax examiners filed a formal grievance before the pilot processed its first return. Lorena Quintero, the agency's Chief Human Capital Officer, had expected skepticism. She had not expected a grievance in week one. The union's concern was narrow but entirely representative: the announcement said AI would "identify" audit candidates, and the union read "identify" as "decide." No labor-management agreement authorized an algorithm to decide which taxpayers would face an audit without examiner judgment as a meaningful part of the process.

Lorena had not chosen that word carelessly. She had chosen it without consulting the people it would affect most. What follows is what government workforce transformation at scale actually requires: not only the technology deployment, but the organizational, contractual, and cultural work that decides whether the deployment survives its own first quarter.

Why Government Transformation Is Different

Private sector transformation accounts can treat employee resistance as a change management problem to be overcome. Government transformation leaders do not have that option. Government workforces are protected by civil service rules, collective bargaining agreements, and public employment law in ways that private sector workforces are not. Layoffs require legislative authorization in many jurisdictions. Role changes require bargaining in unionized agencies. New performance metrics require negotiation where the existing agreement already specifies what may be measured and how.

None of those constraints makes transformation impossible. They make it sequential. Consultation and agreement come before deployment, not after. Agencies that reverse the sequence, deploying tools and then managing the labor response, consistently face delays, grievances, and morale problems that cost more time and money than early consultation would have. The sequence is not a courtesy extended to the workforce; it is the cheapest available path to a working system. Lorena's first lesson from the grievance was blunt enough to write on a card: next time, talk to the union before you name the pilot.

The naming point generalizes further than it looks. In a labor-management context, the vocabulary of an announcement is not marketing copy; it is a description of authority. "Identify," "recommend," "flag," "select," and "decide" each imply a different allocation of judgment between a system and a person, and the agreement in force may already define some of those words. A communications team optimizing an announcement for clarity can, in a single verb, describe a change in working conditions that nobody has bargained. Route the language through the same people who would review the change itself.

The Reskilling Imperative

AI does not eliminate government jobs at the rate early predictions suggested. It does change what those jobs require. A tax examiner who previously spent 60 percent of their time identifying statistical anomalies in returns now spends that time evaluating AI-flagged anomalies and applying judgment about which ones warrant investigation. The work has not disappeared. The skill required to do it well has shifted, from finding the pattern to interrogating a pattern someone else's system has already found.

The resulting gap is predictable, which means it can be planned for rather than discovered. Examiners trained to identify patterns manually may struggle to critically evaluate model-generated flags, and specifically to distinguish a genuine anomaly from a systematic bias in the training data. New data literacy is required: understanding what a flag actually means, knowing when to override it, and documenting the reasoning when human judgment diverges from the model's recommendation. That last skill is the one agencies most often omit from training and most often need in front of an oversight body.

Lorena's agency built a 40-hour reskilling curriculum for the 3,400 examiners, delivered in four-hour modules over ten weeks, which is exactly the 40 hours the design promised. The curriculum covered three areas: how the audit selection model works at a conceptual level, enough to understand its inputs and outputs without pretending to make examiners into engineers; how to evaluate a flagged case against the existing examiner judgment criteria; and how to document a human override in a way that supports both quality review and model improvement.

Delivery ran on internal trainers, which is worth noticing because it changes more than the cost line. Internal delivery puts the training in the vocabulary the workforce already uses, lets trainers answer questions about this agency's cases rather than generic examples, and leaves the capability inside the organization when the rollout ends. It also means the trainers are themselves staff whose time has to come from somewhere, which is a real cost rather than a saving, and one that shows up in the delivery figure rather than disappearing into goodwill.

The economics deserve to be stated with the inputs visible rather than as a headline. Curriculum development cost $280,000, including external design support. Delivery, using internal trainers, cost $190,000 in staff time. That is $470,000 across 3,400 examiners, or roughly $140 each for 40 hours of structured skill development, which the agency assessed as less than a day of fully loaded examiner pay. Whether that ratio holds in your jurisdiction depends on your own pay structure, and it is worth calculating rather than assuming, because the comparison is what wins the budget argument.

New Roles Created, Not Just Modified

Transformation at scale creates genuinely new roles rather than only modifying old ones, and agencies that fail to name the new roles end up assigning the work informally to whoever is willing. Three roles recur across agencies deploying AI at scale. Each has a distinct failure mode it exists to catch, and each needs a position description, a time allocation, and a place in the organizational chart before the tool goes live.

AI output reviewer. A senior staff member responsible for reviewing model recommendations before they generate agency actions affecting constituents. This is not ordinary case review. It requires understanding the model's typical failure modes, watching for patterns in error rates rather than judging cases one at a time, and escalating systematic issues to the AI governance function. In Lorena's agency, six senior examiners were designated as output reviewers, each with a 20 percent time allocation and a salary supplement of $4,200 annually reflecting the added responsibility.

Algorithmic equity monitor. A staff member or team tracking demographic patterns in AI-assisted decisions and flagging disparate impact concerns before they become complaints or litigation. In tax administration that means tracking audit selection rates by geography, business type, and industry sector to surface systematic patterns that might reflect training data bias. The role often sits inside an existing civil rights or equity office, which is sensible, but it requires technical knowledge about bias detection that such offices do not automatically hold.

Human-AI workflow designer. A role focused on continuously improving the interface between model outputs and human decisions. As staff gain experience, some flagged categories prove highly reliable and warrant streamlined review, while others stay high-variance and warrant sustained intensive review. The workflow designer tracks that evolution and proposes process changes to leadership and to the union, which matters because a change in review intensity is a change in how people work and therefore a change somebody has to agree to.

Redesigning the Work, Not Just the Roles

Naming three new roles is the visible half of organizational redesign. The invisible half is that the division of labour between people and the system does not stay where it was set on the first day. As staff accumulate experience, some categories of flagged case prove consistently reliable and some stay high-variance, and a review process that treats every flag identically will over-invest attention in the easy cases and under-invest it in exactly the ones that need judgment. That drift is not a defect in the tool. It is what happens when a process designed under uncertainty meets a year of evidence.

Handling it deliberately is why the workflow designer role exists. The pattern is to let review intensity follow demonstrated reliability by category, streamlining where the record supports it and sustaining intensive review where variance stays high. Two constraints keep that from becoming a quiet loosening of standards. The evidence for reclassifying a category has to be visible to the people who will work under the new process, and any change in how a category is reviewed is a change in how people work, which means it goes to leadership and to the union rather than into a procedure note.

The same logic applies to the reverse direction, which agencies find harder. If a category that was streamlined starts producing errors, review intensity has to go back up, and the process for restoring it should be agreed before it is needed. Designing only the loosening path and not the tightening path is a common asymmetry, and it usually surfaces at the least convenient moment, when someone has already been harmed by a case that a streamlined path waved through.

Union Considerations and Labor Relations

The labor-management dimension of AI transformation is not a communications problem. It is a governance problem. Unions have legal bargaining rights over wages, hours, and working conditions, and tools that change the conditions workers experience, how their performance is measured, or what tasks constitute their job are subject to those rights in most jurisdictions. Treating consultation as an announcement exercise misreads what the other party is entitled to, which is why the grievance arrived in week one rather than in month six.

What is bargainable and what is a management right varies by jurisdiction and by the agreement in front of you, so the useful move is not to assert the answer but to ask the questions early and in writing. Does the deployment change what is measured about an individual's performance, or how? Does it change the tasks that constitute the job description? Does it change working conditions in a way the agreement already speaks to? Does anything in it touch discipline or promotion? Put those to agency counsel and to labor relations before the schedule is published, because each yes converts a project decision into a negotiation with its own timeline.

Lorena's agency resolved the initial grievance by agreeing to three things. Audit flags are advisory, not directive, and examiners retain full authority to decline to recommend an audit for a flagged return. Model recommendations will not be used to evaluate individual examiner performance. And the union receives quarterly data on audit selection rates and override rates, disaggregated by examiner cohort, for the duration of the pilot. Those were not concessions that weakened the program. They were the governance commitments that made it legitimate enough to proceed.

It is worth being precise about what each commitment does and does not deliver. Advisory status is real only if the override is genuinely available and genuinely unpunished; if override rates quietly become a performance signal, the commitment has been reversed without anyone amending the agreement. Regular data sharing gives the union the information to be an informed partner in quality improvement, which is the necessary condition for that relationship rather than a guarantee of it. What the commitments buy is a defensible process, not a defence in itself, and the difference matters the first time an oversight body asks how the agency knows the safeguards held.

The side letter documenting the agreements took three weeks to negotiate. The pilot it enabled has been running for 22 months. In that time the AI-assisted process has produced a 31 percent improvement in audit-to-assessment conversion rates, meaning selected audits are more likely to result in an assessment, with no measurable increase in examiner overtime and no formal complaints from affected taxpayers about disparate selection. Three weeks against 22 months is the ratio to quote when somebody proposes skipping the consultation to save time. It is also the reason the results are quotable at all: an agreed pilot produces figures both parties accept, while a contested one produces figures each side reads differently and neither side can use.

Managing Workforce Anxiety at Scale

Anxiety about displacement in government is real and not irrational. Direct job elimination is slower in government than in the private sector, but role changes, reclassification, and eventual headcount reductions in positions that AI makes more efficient are genuine possibilities over a 5 to 10-year horizon. Leaders who wave this away with "AI is a tool, not a replacement" lose credibility with frontline staff who have watched enough technology cycles to discount that sentence entirely.

The more credible posture is honest and specific: no plans to eliminate examiner positions as part of this pilot; if future budget decisions affect staffing levels, those decisions go through the normal legislative and bargaining processes; quarterly workforce data goes to the union, and any AI-related change in staffing gets reported to the legislature. That statement is not a guarantee, and it should not be dressed up as one. It is a commitment to transparency and process, which is what a credible government employer can honestly offer, and staff can tell the difference between that and reassurance.

What to Measure from Day One

Transformation programs are usually measured by whether the technology works, which is the least contested question in the room. The measures that decide whether the program lasts are the ones that speak to the agency, the staff, and the public at once. Lorena's agency tracked audit-to-assessment conversion, examiner overtime, and formal complaints about selection from the affected population, and reported override rates to the union alongside them. Each covers a different constituency's version of failure, and each was in place before the first return was processed rather than reconstructed afterwards.

Baselines are the part that cannot be added later. A conversion rate improvement means nothing without the prior rate, an overtime claim means nothing without the prior hours, and a complaint count means nothing without knowing what a normal quarter looked like. If a program is already live without baselines, say so plainly rather than presenting post-deployment figures as evidence of change. Oversight bodies are considerably more forgiving of a missing baseline than of a comparison that turns out to have been assembled after the fact.

Reporting cadence is part of the measurement design rather than an administrative detail. In Lorena's agency the same quarterly rhythm carried selection rates and override rates to the union and workforce data to the same audience, which means the numbers arrive on a schedule nobody has to request and no single quarter's figures land as an event. A metric produced only when someone asks for it is a metric produced under pressure, usually to answer a question that has already been framed by whoever asked.

Anti-Patterns

• Naming the pilot before consulting the workforce. The verb in an announcement describes an allocation of authority. "Identify," "recommend," and "decide" are not synonyms in a bargaining context, and choosing one without the union in the room is how a communications decision becomes a grievance.
• Deploying first and bargaining afterwards. Reversing the sequence does not save time. It relocates the negotiation to a point where the agency has less room to move, the workforce has more reason to distrust the effort, and the delay is visible to everyone above you.
• Letting override rates become a performance measure. An advisory-only commitment collapses the moment overriding the model starts to look costly to the individual examiner. If override data is shared, be explicit about what it will and will not be used for, and hold that line.
• Treating data sharing as the relationship. Quarterly reporting gives a union the information required to be a partner in quality improvement. It does not make anyone a partner by itself, and presenting it as though it does tends to insult the people receiving it.
• Funding the tool and not the training. A transformation budget with no reskilling line assumes the skill shift happens for free on staff time. It does not; it happens badly, on staff time, and shows up later as inconsistent judgment nobody can explain to an auditor.
• Leaving the new roles unnamed. Output review, equity monitoring, and workflow design are jobs whether or not anyone is assigned to them. Unnamed, they land on the most conscientious person available, without time, recognition, or authority to escalate.
• Offering reassurance instead of process. "Nobody will lose their job" is a promise most agency leaders cannot keep and staff know it. A described process, with the bargaining and legislative steps named, is weaker as a sentence and far stronger as a commitment.

Practice Prompts

• Audit your own announcement language. Take a planned or recent AI announcement and underline every verb describing what the system does. For each, write what authority it implies and whether your agreements already define that word.
• Map the sequence. List the consultation, bargaining, and approval steps required before your next AI deployment can change how anyone works, with a realistic duration for each. Then compare that against the deployment date already in the project plan.
• Cost the reskilling honestly. Estimate curriculum development, delivery, and staff time for your affected population, then divide by headcount and compare the result to a day of fully loaded pay for that role. Bring both the total and the per-person figure to the budget conversation.
• Write the three position descriptions. Draft position descriptions for an output reviewer, an equity monitor, and a workflow designer in your context, including time allocation and who each escalates to. Note which of the three your organization is currently doing informally.
• Draft the honest staffing statement. Write what you can truthfully commit to about staffing over the next two years, naming the processes that govern any change, and read it aloud to a colleague who works on the frontline.

Reflection

• Where in your last deployment did consultation actually sit in the sequence, and what did the placement cost you in time you did not attribute to it?
• If your workforce were asked today whether AI recommendations affect how their performance is judged, what would they say, and would the agreement in force support their answer?
• Which of the three new roles is already being performed by someone in your organization without a title, time allocation, or recognition?
• What is the most specific staffing commitment you could make and keep, and what stops you from making it in those words?
• If an oversight body asked for your pre-deployment baseline on the metrics you now report, could you produce it, or would you be reconstructing it?

Glossary

• Advisory, not directive. A governance commitment that model output informs a human decision without determining it, with the deciding staff member retaining full authority to disregard the recommendation.
• Override documentation. The record a staff member creates when human judgment diverges from a model recommendation, capturing the reasoning so it can support both quality review and model improvement.
• AI output reviewer. A senior staff role reviewing model recommendations before they produce agency actions, focused on failure modes and error patterns rather than individual case outcomes.
• Algorithmic equity monitor. A role tracking demographic and geographic patterns in AI-assisted decisions to surface disparate impact before it becomes a complaint or a lawsuit.
• Human-AI workflow designer. A role that adjusts the division of labour between system and staff as experience accumulates, and takes the resulting process changes to leadership and the union.
• Side letter. A negotiated document recording commitments agreed between an agency and a union alongside the main agreement, used here to record the terms under which the pilot could proceed.

Related Lessons

• Workforce Planning for AI covers the planning work that should precede a transformation of this size.
• Change Management for AI Adoption handles the adoption mechanics that sit alongside the labour relations track.
• AI Talent Development and Retention extends the reskilling question into keeping the people you have retrained.
• AI and National Workforce Transformation takes the same question to national scale, beyond a single agency's workforce.
• Measuring AI Impact covers the measurement discipline behind conversion rates, baselines, and what a claimed improvement has to survive.

Closing

The grievance in week one was not a setback in Lorena's programme. It was the programme finding out, early and cheaply, what it had skipped. Three weeks of negotiation produced a side letter, the side letter produced a pilot, and the pilot has now run long enough to show results that nobody is contesting. The alternative history is not a faster deployment; it is the same deployment fought over for a year with less trust available at every step.

Workforce transformation in government is not about managing resistance. It is about earning the right to change how public work gets done, through consultation, specific commitments, and demonstrated respect for the people doing the work. That is slower than a private sector account would suggest, and it is the only version that holds up when the people who authorized it have moved on.

Key Takeaways

• Consult before you name the pilot. Word choices such as "identify" against "recommend" carry legal weight in labor-management contexts. Early consultation prevents grievances that cost far more time than the conversation would have.
• Sequence, do not skip. Civil service rules, bargaining agreements, and public employment law make transformation sequential rather than impossible. Deploying first and negotiating afterwards relocates the negotiation to worse ground.
• Budget the reskilling with the inputs visible. Development at $280,000 and delivery at $190,000 across 3,400 examiners is roughly $140 each for 40 hours of structured training. Bring the components, not just the per-person number, to the budget conversation.
• Create new roles, not just modified ones. Output reviewer, equity monitor, and workflow designer are real positions with real time costs. Name them, fund them, and place them in the workforce plan before deployment.
• Advisory status has to be defended, not just declared. Human authority over recommendations is the architecture that makes deployment defensible to oversight bodies, and it fails quietly if overriding the model becomes costly to the individual.
• Share performance data on a defined schedule. Quarterly selection rates, override rates, and cohort-level metrics without identifying individuals give a union what it needs to act as an informed partner in quality improvement.
• Treat workforce anxiety as governance, not communications. Specific commitments about process are more credible than generic reassurance, and credibility with frontline staff is what carries a programme past its first pilot.
• Set baselines before day one. Conversion rates, overtime hours, and formal complaints only mean something against a prior measurement. A baseline cannot be reconstructed after deployment without inviting the obvious question.

Frequently Asked Questions

Does consulting the union early actually save time?

In this case the arithmetic was stark. The side letter took three weeks to negotiate and enabled a pilot that has since run for 22 months. The grievance that prompted it arrived before the first return was processed, when the agency had the least leverage and the least trust available. Early consultation does not remove the negotiation; it moves it to a point where the design can still absorb the outcome.

What exactly does "advisory, not directive" commit the agency to?

That staff retain full authority to disregard a model recommendation, and that the recommendation does not by itself produce an agency action. Making it real requires two further things: not using model agreement or disagreement to evaluate individual performance, and making sure the override path is quick enough that people actually use it. A commitment that survives only while nobody exercises it was never a commitment.

How much reskilling is enough?

The lesson's example is 40 hours delivered in four-hour modules over ten weeks, covering conceptual understanding of the model, evaluation of flagged cases against existing judgment criteria, and documentation of overrides. Whether that fits your setting depends on the work, and the more transferable point is the third topic: agencies routinely train people to use the tool and forget to train them to record why they disagreed with it.

Can we promise staff their jobs are safe?

Most agency leaders cannot keep that promise and staff generally know it. What can be committed to is process: no eliminations as part of this pilot, any future staffing change routed through normal legislative and bargaining channels, workforce data shared on a schedule, and AI-related staffing changes reported to the legislature. That is weaker as a sentence and far more durable as a commitment.

Where do the new AI roles sit organizationally?

Output reviewers usually sit with the operational staff whose work they review, because the role depends on subject matter judgment. Equity monitoring often sits in an existing civil rights or equity function, which brings the mandate but not automatically the technical knowledge for bias detection. Workflow design needs proximity to the operational unit and to whoever takes process changes to leadership.