On July 16, the U.S. Bureau of Labor Statistics put a quiet table inside a familiar government product. The table did not forecast one sudden AI job crash. It did something more useful for a workforce meeting: it split the labor market into occupations that gain demand as AI spreads and occupations that lose demand as software handles more repeatable work.

The numbers were precise. In the BLS 2024-34 employment projections, data scientist employment is projected to rise 33.5%, adding 82,500 jobs. Information security analysts are projected to rise 28.5%, adding 52,100 jobs. Software developers are projected to rise 15.8%, adding 267,700 jobs.

The other side of the same table is harder for HR leaders to hide inside a generic “future of work” slide. Customer service representatives are projected to decline 5.5%, losing 153,700 jobs. Procurement clerks are projected to decline 8.7%, losing 5,400 jobs. Credit authorizers, checkers and clerks are projected to fall 6.2%. Legal secretaries and administrative assistants are projected to fall 5.8%.

That is the workforce split.

It does not say AI eliminates work. It says AI changes where headcount has to sit, what type of work gets priced upward, which service desks become automation targets and which employees will ask for protection before the budget moves.

For a CHRO or CFO, the BLS table creates a practical problem. A company can hire data scientists, security analysts and AI-enabled product roles while cutting support work in the same planning cycle. A board deck can call that productivity. Employees will call it a replacement plan unless the company can show who gets moved, who gets trained, who gets contracted out, which work still needs human judgment and which roles are being removed rather than redesigned.

Picture the July planning room. Finance has an AI savings target. IT has an agent rollout calendar. HR has an open requisition list for analytics and security roles. The support desk has a ticket backlog that no longer looks like last year’s work. Everyone can be right and still leave the worker with no path across the split.

The conversation has already started outside government data. Upwork’s 2026 Future Workforce Index says skilled freelancing among U.S. knowledge workers rose from 28% in 2025 to 38% in 2026, while 58% of full-time employees are considering freelancing to reach better opportunities. Autodesk’s 2026 AI Jobs Report says AI jobs across architecture, engineering, construction, product design, manufacturing, media and entertainment rose 147% over two years. PwC’s 2026 AI Jobs Barometer says jobs requiring specific AI skills grew 69% while the overall jobs market grew 9%, and that the average AI skill wage premium reached 62%.

Those numbers do not cancel the BLS decline lines. They explain why the same company can be hiring and cutting at once.

AI moves work out of one column and into another. The problem is that the people do not move automatically.

July 16 put AI into the occupation table

BLS employment projections are slow by design. They are not a social-media layoff tracker, a vendor survey or a CEO quote about replacing people with agents. The Monthly Labor Review overview says total U.S. employment is projected to grow from 170.0 million jobs in 2024 to 175.2 million in 2034, an increase of 5.2 million jobs, or 3.1%.

Inside that slower labor market, the occupational split matters more than the headline total.

Computer and mathematical occupations are projected to grow 10.1%, adding 545,600 jobs. Management occupations are projected to add 833,400 jobs. Business and financial operations occupations are projected to add 586,900 jobs. Installation, maintenance and repair occupations are projected to add 301,400 jobs.

Office and administrative support occupations sit on the other side. BLS projects a 3.9% decline for that occupational group, a loss of 761,900 jobs. Sales and related occupations are projected to decline 2.0%, a loss of 297,800 jobs.

BLS names AI as one of the forces behind the split. The agency says growing adoption of AI technologies, including generative AI tools, is expected to dampen labor demand in fields such as sales, design and administrative support. It also says demand for AI technologies boosts employment demand for computer occupations involved in development and implementation.

That language is careful. It should be. Employment projections are not a verdict on one tool or one company. They describe a structural pressure.

But careful language can still change a budget meeting.

A CHRO cannot read the BLS table and treat AI workforce planning as one number. The same plan has at least four ledgers: the roles the company needs to build AI, the roles it needs to secure AI, the roles it needs to apply AI inside business functions and the roles whose routine workload is being absorbed by software.

Those ledgers do not reconcile themselves. If finance counts a support-role reduction as savings but HR has not built the redeployment path, the plan creates hidden cost. If IT buys agents for service workflows but legal has not decided how decisions will be reviewed, the plan creates risk. If managers expect AI to remove routine tasks but still need employees to explain, repair and own edge cases, the plan creates a training gap.

The BLS table makes the split visible enough to manage.

Growth sits where judgment meets systems

The fastest projected AI-adjacent growth in the BLS table is not in vague “AI jobs.” It sits in occupations that combine systems knowledge, statistical skill, security, research and operational responsibility.

Data scientists sit at the top: 33.5% projected growth from 2024 to 2034. Information security analysts follow at 28.5%. Operations research analysts are projected to rise 21.5%. Computer and information research scientists rise 19.7%. Software developers add the largest absolute number in the selected AI and IT table, with 267,700 additional jobs.

That mix tells a workforce planner something specific. AI adoption does not only ask companies to hire model builders. It asks them to hire people who can make messy operations measurable, defend systems from new risk, convert workflow constraints into models and turn software into production work.

PwC’s 2026 Global AI Jobs Barometer points in the same direction from a different dataset. PwC analyzed more than one billion job ads across 27 countries and territories. It describes a two-track labor market: roles where AI amplifies expert judgment grow faster than roles where AI makes the work easier for non-experts to perform. PwC reports that professionalized roles see twice the growth in available jobs and 42% faster salary growth than democratized roles.

The wage signal is even sharper. PwC says the average wage premium for AI skills reached 62%, up from 57% last year. Jobs requiring specific AI skills grew 69%, compared with 9% for the overall jobs market. Companies most able to use AI showed faster headcount growth than the least AI-exposed companies, 52% versus 36% on a 2018 baseline.

This is why “AI will reduce headcount” is too blunt as a planning sentence. Some work gets compressed. Some work gets repriced. Some work gets moved from routine execution into design, review, security, operations and judgment.

Microsoft’s 2026 Work Trend Index adds the management layer. Microsoft surveyed 20,000 AI users across 10 countries and analyzed Microsoft 365 Copilot use. It found that organizational factors such as culture, manager support and talent practices account for more than twice the reported AI impact of individual mindset and behavior. It also found that AI users name quality control of AI output and critical thinking as the human skills that become more important as AI takes on more work.

Those are not soft add-ons. They are the work surface the company has to fund.

If a company hires data scientists but does not fund evaluation, data governance, security review and manager judgment, it is only buying the visible part of the workforce shift. If it cuts administrative roles before deciding where exceptions, corrections, employee questions and customer escalations will go, it has reduced payroll before it has redesigned work.

BLS forecasts the occupation. PwC prices the skill. Microsoft names the organizational condition. Together, they point toward the same answer: AI workforce planning is no longer a headcount spreadsheet. It is a map of which tasks move to software, which decisions stay with people and which jobs change because the interface between the two is now the business process.

Office support carries the official decline

The sharpest BLS losses are not in glamorous white-collar strategy roles. They are in the support layer that has always absorbed routine information work.

Customer service representatives lose the largest absolute number in the selected BLS AI and IT table: 153,700 jobs, a projected 5.5% decline. Procurement clerks decline 8.7%. Credit authorizers, checkers and clerks decline 6.2%. Legal secretaries and administrative assistants decline 5.8%. Medical transcriptionists decline 4.9%. Executive secretaries and executive administrative assistants decline 1.6%.

The Monthly Labor Review overview broadens the point. Office and administrative support occupations, as a group, are projected to lose 761,900 jobs from 2024 to 2034.

The AI labor debate becomes operational in this layer. Many support roles contain two kinds of work that look similar in a budget system but behave differently in practice.

One kind is repeatable routing. A worker receives a request, checks a policy, enters data, schedules a follow-up, updates a record, sends a status message or routes the issue to the right owner. AI agents and workflow software are aimed directly at this layer.

The other kind is exception handling. A worker notices that a request is politically sensitive, that a customer is angry for a reason the form does not capture, that a benefits answer could affect protected leave, that a procurement request hides a security risk or that a manager is trying to push a messy people problem into a ticket. This work is less visible and more expensive to automate safely.

BLS does not say which tasks vanish inside each occupation. That is the company’s job.

For HR and finance, the mistake is to treat the decline line as permission to remove the role before separating routing from exception work. If 40% of the work can be automated, that does not mean 40% of the role can disappear cleanly. The remaining work may be harder, more sensitive and more dependent on institutional memory than the tasks that left.

That pattern has appeared in several recent AI adoption stories. Employees save time on draft production and spend more time reviewing output. Recruiters get more applications and need stronger signals. Engineering teams generate more code and still need senior review. Managers receive AI-written work and have to decide whether it is useful, copied, wrong or risky.

Support jobs can follow the same path. The routine ticket count falls. The remaining ticket is more complex.

If a company misses that second-order effect, it will think AI lowered the support cost while moving the hard part to managers, legal, HR business partners, customer success or the employees themselves. The BLS decline then becomes a false saving.

Agents make the service desk a workforce plan

The product market is already trying to capture the support layer BLS flags as exposed.

ServiceNow announced in May that its Autonomous Workforce would expand AI specialists across IT, CRM, employee service teams, security and risk. Its release says the company is introducing AI specialists across HR, workplace services, legal, finance, procurement, supplier management and health and safety. ServiceNow estimates that 23 million employees use its employee portal every month, generating more than 40 million cases annually. Across its customer base, ServiceNow says AI specialists resolve 91% of cases without reassignment.

Workday’s HR agent framing points to similar work surfaces: applicant screening, interview scheduling, PTO, benefits, policy requests, onboarding, engagement, retention risk and workforce planning. The company describes agents as a way to handle administrative load while HR professionals focus on coaching, retention and culture.

The product promise is plausible because the support desk is structured data attached to repeatable requests. That is also why the employment impact can be direct.

If a service desk agent can resolve a benefits question, update a record, route a case and generate a response, then the company will ask why the old staffing model still holds. If an HR agent can model attrition, flag a pipeline gap and suggest sourcing actions, the workforce-planning team will change. If a procurement specialist agent can triage supplier requests and draft routine answers, procurement support roles will shrink or shift toward supplier strategy and risk.

Buyers should not stop at whether the agent works in a demo. They need to know whether the company has mapped the people around it.

Who reviews the cases the agent closes? Which requests must stay with a human because they touch legal rights, pay, protected leave, medical accommodation or employment status? Which support workers are retrained into exception analysts, policy owners, workflow designers or quality reviewers? Which managers inherit the work that used to be hidden in the support desk? Which cost center receives the savings, and which one pays for the review hours?

Those questions belong in the purchase order.

Without them, AI service agents can become a budget transfer mechanism. HR cuts administrative roles. IT owns the agent. Legal carries risk when the agent mishandles a sensitive case. Managers spend more time explaining policies. Employees wait longer for the human exception queue because the easy cases disappeared and the remaining cases are harder.

That is workforce design failure, not automation failure.

BLS gives employers the warning early. Customer service and administrative support roles are projected to decline because software can absorb more of the routine work. ServiceNow and Workday show how that absorption is being packaged. The missing artifact is a workforce map that shows where the human work goes after the service desk changes shape.

Freelancers price the split first

Freelance markets often price a workforce shift before enterprise job architecture catches up. Upwork’s 2026 Future Workforce Index is useful for that reason.

The report says skilled freelancers now represent 38% of U.S. knowledge workers, up from 28% in 2025. It also says 58% of full-time employees are considering freelancing, up from 36% the prior year. Workers are not only responding to layoff anxiety. They are looking for markets where AI-enabled skill, domain knowledge and judgment can be sold more directly.

Upwork’s pricing data shows the split. Freelancers performing AI work on the marketplace earn 34% more per hour than those not incorporating AI. But low-complexity AI execution is already getting cheaper. Generative AI and creative production work saw 90% year-over-year growth in contract starts while per-contract earnings declined 13%.

Complex AI-augmented work moved the other way. Upwork reports that AI-augmented professional services grew 72% year over year and that earnings in that area rose 22%. Freelancers doing more complex work with AI saw earnings increase 45%. Upwork uses the label “AI orchestrator” for the worker who connects tools, domain expertise, judgment, workflow design and business outcomes.

That phrase can sound like marketing. The pricing pattern underneath it matters more.

Execution tasks scale and lose margin. Judgment tasks scale less easily and gain pricing power.

Autodesk’s 2026 AI Jobs Report finds the same pattern in industries that design and make the physical world. AI jobs across those industries rose 147% over two years and 33% in the past year. The fastest-growing titles in its report are not only AI engineers. They include AI UX Designer, AI Creative Technologist, AI Consultant, AI Strategist, AI Content Designer and AI Systems Designer. Autodesk also reports that operation skills, communication skills, leadership skills, collaboration skills, people skills and training skills rank among the most in-demand skills for AI roles in Design and Make industries.

That does not mean everyone becomes a strategist. It means companies need more people who can put AI into the work of a field.

An architect using AI for design review still needs to understand building constraints. A manufacturing team using AI for process optimization still needs operators who know where the floor actually breaks. A media team using generative AI still needs people who can judge originality, rights, quality and audience fit. A support team using agents still needs people who can read the exception that the system failed to understand.

The BLS occupation table meets the market here. The jobs at risk are often defined around repeatable process execution. The jobs gaining pricing power are defined around judgment applied through systems.

The difference should change pay bands. It should also change training budgets.

If the company only trains employees to prompt tools, it is training for the cheaper side of the split. If it trains employees to redesign workflows, review output, make domain tradeoffs, handle exceptions, preserve customer trust and explain decisions, it is training for the side of the labor market that still prices upward.

A workforce map for AI adoption

The workforce split needs a planning artifact. A simple version can sit inside the annual headcount plan.

Work surfaceCurrent signalAI pressureHuman work that remainsBudget ownerProtection trigger
Data science and analyticsBLS projects data scientists +33.5%, adding 82,500 jobsMore business processes need measurement, forecasting and model evaluationProblem framing, data quality, interpretation, decision supportAnalytics, product, financeOverhiring model builders without business owners
Security and riskBLS projects information security analysts +28.5%, adding 52,100 jobsAgents, data access and automated actions widen the attack surfaceIdentity, permissions, incident response, audit reviewCIO, CISO, legalAgent access grows faster than controls
Software and implementationBLS projects software developers +15.8%, adding 267,700 jobsAI creates more software paths and more generated codeArchitecture, review, integration, release ownershipCTO, engineering, productCode volume rises without review capacity
Customer serviceBLS projects customer service representatives -5.5%, losing 153,700 jobsAI agents answer, route and resolve routine requestsEscalations, empathy, policy exceptions, trust repairCOO, CX, HR, legalAutomation handles employment, health, pay or legal-sensitive cases
Office and administrative supportBLS projects the group -3.9%, losing 761,900 jobsScheduling, records, billing, procurement and forms are easier to automateException handling, data correction, workflow ownershipHR, finance, operationsSupport cuts move hidden work to managers
AI-applied domain rolesAutodesk reports Design and Make AI jobs +147% over two yearsAI moves into industry workflows beyond softwareField knowledge, design judgment, customer context, tool evaluationBusiness units, L&D, productGeneral AI training misses field-specific tools
AI-augmented independent workUpwork reports AI work earns a 34% hourly premium, but low-complexity AI creative earnings fellFreelancers test pricing faster than job laddersClient diagnosis, domain advice, workflow delivery, business impactProcurement, HR, business leadersContractor mix grows without career path or worker-protection plan
Early-career workWEF says more than one in three young workers are in occupations with medium to high AI task exposureRoutine first-pass work moves to softwareApprenticeship tasks, review practice, customer context, judgmentManagers, L&D, schoolsEntry roles require senior judgment without training time

This table is not a forecast model. It is a meeting tool.

Its purpose is to force a company to state what the AI plan is doing to work across software spend, headcount, training time and worker protection. For each work surface, a leader should be able to answer five plain questions.

Which tasks are being automated? Which roles grow because AI needs implementation, security, review or domain judgment? Which employees are close enough to the growing work to be retrained? Which jobs are being reduced because routine demand fell? Which protection or notice process starts before the role is removed?

That last question is the one many companies avoid until too late.

The BLS table is not a layoff list. It does not tell any employer which employee to remove. But it does make certain reductions foreseeable. Once a company knows customer service, procurement support, administrative support or transcription work is exposed, it cannot pretend displacement is a surprise when the next software rollout arrives.

Foreseeable displacement should have a file.

That file should include affected tasks, roles at risk, adjacent roles, internal openings, training time, severance standards, voluntary exit options, manager-review duties, employee notices and evidence that protected leave or disability accommodation does not get mistaken for low activity. The same company that can measure AI adoption should be able to measure the transition it creates.

Workers ask for protection while budgets move

Workers can see the split before a company names it.

The Alphabet Workers Union’s job-security campaign says more than 4,500 Googlers have signed its petition. Its demands include companywide voluntary exit packages, guaranteed severance, no GRAD quotas and severance as leave. The campaign started before the latest BLS table, but it belongs in the same discussion because employees are reading AI spend, role reductions and performance pressure as one system.

This is the employee side of workforce planning. When a company says it is investing in AI while also running layoffs, workers ask whether the AI budget is being funded with their jobs. When a company changes performance systems while discussing productivity gains, workers ask whether the new metrics will become a layoff filter. When a company offers voluntary exits to some groups but not others, workers ask why protection depends on org chart timing.

Those questions are not separate from adoption. Trust shapes the quality of AI rollout.

Employees who think AI telemetry can be used against them will behave differently. They may avoid tools that create usage data. They may stop sharing failed prompts. They may hide workarounds. They may treat every productivity metric as a future employment record. They may look for freelance work, internal transfers or external job options before the company has a chance to redeploy them.

Upwork’s freelancing shift is one signal of that behavior. The Google job-security petition is another. Neither source says every worker is leaving the full-time labor market because of AI. The sharper point is that workers are looking for control as employers move budgets around them.

A company can respond in two ways.

It can treat worker anxiety as resistance and keep the AI plan inside finance, IT and executive meetings. That may move faster in the short term. It also makes every role change look like an announcement from above.

Or it can publish a transition standard before the cuts arrive. That standard does not have to promise no layoffs. It has to say what happens when AI changes a role: how early employees are told, which roles are offered first, how internal mobility works, how managers review role changes, which training counts as paid work, how severance is calculated and which metrics are prohibited in employment decisions.

That is not generosity. It is operating discipline.

AI changes the cost curve for some tasks. It does not remove the employer’s need to manage trust, legal exposure, knowledge transfer and future hiring reputation. A company that burns employee trust during one AI restructuring will pay for it later in retention, recruiting, adoption quality and litigation risk.

The BLS data makes that future easier to see. The office support line is moving down. The data science, security and software lines are moving up. Inside the company, the hard decision is whether an employee gets a path across the split before the plan becomes a termination.

The plan starts before a role disappears

The wrong time to build an AI workforce plan is after a manager receives a list of roles to cut.

By then, the company has already made several decisions. It has chosen the work to automate. It has bought or built the system. It has decided the savings target. It has defined which cost center gets relief. It may have allowed new metrics to shape performance ratings or staffing models. The employee arrives at the end of the process, when the company has the least flexibility and the most risk.

BLS offers a slower clock. Its projections run from 2024 to 2034. That gives employers time to build a workforce plan that is more honest than “AI will make everyone more productive” and more useful than “some jobs will go away.”

The plan can start with a simple split.

First, name the growth roles. Data scientists, security analysts, software developers, operations research analysts, domain AI specialists, workflow designers and AI implementation roles are not interchangeable. Each needs a hiring market, pay band, manager model and review standard.

Second, name the exposed support work. Customer service, procurement support, administrative support, transcription, billing, scheduling and clerical coordination should be mapped by task, not by job title alone. A title can hide work that is automatable and work that should never be left to a system without review.

Third, build the bridge. The bridge is not a learning portal. It is paid time, manager capacity, internal job priority, skills evidence, human review, severance clarity and a rule for when AI adoption data cannot be used as employment data.

Fourth, publish the stop signals. If an agent closes sensitive cases without review, stop. If a support cut moves hidden work to managers, stop. If protected leave depresses an activity metric, stop. If AI output quality falls while headcount savings are already booked, stop. If employees cannot understand which data affected a role decision, stop.

The BLS table does not decide these questions. It gives them dates and numbers.

That is enough to make evasion harder.

A company that plans to hire AI talent while reducing support labor now has an official public forecast it can cite. It also has an official public forecast employees can cite back. The same data that helps finance justify a workforce redesign can help workers ask for redeployment, training, severance and notice.

This is where the workforce split becomes real. It is not a debate about whether AI creates or destroys jobs in the aggregate. It is a decision about whether a particular company can move work without treating people as an afterthought.

The July 16 BLS table is not dramatic. That is its strength. It gives managers a sober version of the AI labor market: growth in the roles that build, secure and apply systems; decline in the roles that carried repeatable support work; pressure on employees who need a path from one side to the other.

The next planning meeting should start there: with a table, a task map and a list of people whose work is about to change before their title does.

Published July 22, 2026.