HR Cannot Redesign Work From the Old Org Chart
On June 8, 2026, McKinsey gave HR leaders a job description that no org chart can handle cleanly.
The firm called it HR’s dual mandate. HR has to guide the whole company through AI transformation - redesigning roles, workflows, skills, career paths, and human-agent collaboration - while also rebuilding the HR function itself. The same report said 88% of companies now use AI in at least one business function, but only 39% report a material contribution to EBIT from those deployments.
That gap is where the old HR operating model starts to break.
For the past two weeks, this site has followed one narrow but expensive edge of that break: AI hiring budgets, candidate trust, proof-of-person workflows, fraud desks, appeal queues, agent passports, and post-hire recovery rooms. Those pieces matter because recruiting is the first place many companies see AI touch real people at scale. A bad AI screen, fake candidate, weak identity check, or missing appeal path quickly turns into a buyer problem.
The wider problem now sits above recruiting.
If agents can answer employee questions, draft manager packets, update HR and finance records, screen candidates, summarize performance inputs, recommend learning, approve routine cases, route payroll issues, and help redesign jobs, HR is no longer managing a set of people processes around a stable workforce. It is being asked to help define a workforce that includes people, agents, managers, reviewers, service queues, skill graphs, identity controls, and evidence records.
That work cannot be done from a headcount spreadsheet alone.
It also cannot be delegated to IT, Legal, or a vendor dashboard. IT can govern identity and systems. Legal can define risk thresholds. Finance can demand proof of value. Vendors can ship agents into Workday, Gemini Enterprise, ServiceNow, Microsoft 365, payroll, ATS, and learning workflows. None of those owners can decide, by themselves, which tasks should move to agents, where human judgment must stay, how managers should supervise the new work, which employees need reskilling, and which metrics prove the change helped the business.
That is HR’s uncomfortable opening.
The function is expected to be the architect of human-agent work while its own house still contains short-term workforce planning, process-centric HR operations, fragmented people data, cautious AI measurement, uneven HR-IT collaboration, and managers who already carry too much review work. The old org chart can assign a leader to recruiting, learning, total rewards, HR operations, employee relations, analytics, and business partnering. It cannot show how a performance review moves across an AI summary, a manager judgment, a compensation recommendation, an employee appeal, a payroll correction, a vendor evidence file, and a culture question.
The new work crosses the boxes.
By June 10, the practical test has narrowed. Many HR teams already use AI tools. Far fewer have a work model that keeps those tools from turning into another layer of automation sitting on top of old processes, old metrics, and old accountabilities.
June 8 Gave HR Two Jobs
McKinsey’s dual mandate essay is useful because it refuses to let HR choose between enterprise transformation and HR transformation.
The first job is outward-facing. HR must help the enterprise redesign work itself. That means defining how tasks are distributed across humans and AI agents, how jobs change when routine execution moves elsewhere, what skills become more valuable, which career paths still make sense, and where human judgment remains non-negotiable.
The second job is inward-facing. HR has to prove the model inside its own function. A people team that still runs annual planning, case management, recruiting coordination, learning operations, and employee communications with the same structure it had before agents arrived will struggle to tell sales, finance, engineering, support, or operations how to rebuild their own work.
Credibility becomes operational.
This is not a communications problem. A CHRO can speak fluently about responsible AI, employee trust, and new skills while the HR function still routes cases through slow queues, measures success by activity, and treats workforce planning as a headcount reconciliation exercise. Employees will notice. Business leaders will notice faster.
McKinsey put numbers around the pressure. AI could generate $150 billion to $200 billion in annual global value in HR alone, and two-thirds of HR processes can be partially or fully automated. The same article argues that current technologies could automate activities accounting for about 57% of US work hours, while 76% of jobs sit in a messy middle that is neither fully automatable nor untouched.
That middle is the CHRO’s problem.
If a job is untouched, HR can keep the old model. If a job disappears entirely, HR can run a reduction, redeployment, or reskilling plan. Most work will not be that clean. A recruiter may stop doing first-pass screening but spend more time judging structured evidence, resolving candidate appeals, and advising hiring managers. A payroll specialist may stop answering routine policy questions but handle more exceptions created by agent-routed cases. A manager may gain drafting help but lose time reviewing AI-generated recommendations and explaining them to employees.
Those are not minor productivity adjustments. They are job architecture changes.
The old HR playbook treats job architecture as a compensation and leveling discipline. It defines roles, levels, job families, pay ranges, and career ladders. Human-agent work redesign asks a different set of questions before the pay range arrives:
| Redesign question | Old HR owner | New operating conflict |
|---|---|---|
| Which tasks move to agents? | HR transformation or process owner | Business value, legal risk, employee trust, manager capacity |
| Who supervises the output? | Manager or HRBP | Reviewer skill, workload, override authority, evidence trail |
| Which skills gain value? | Learning or talent management | Strategic capability planning, redeployment, hiring mix |
| What should finance fund? | HR operations or procurement | License, workflow meter, change management, governance, rework |
| Who handles disputes? | Employee relations or legal | Manager workflow, HR case queue, vendor evidence, payroll correction |
The table shows why the dual mandate is hard. It does not fit into a single HR center of excellence. It forces HR to redesign its own internal handoffs before it can credibly redesign work elsewhere.
The first useful move is to stop treating “AI in HR” as a software adoption category.
AI in HR used to mean a recruiting chatbot, a resume-ranking tool, a learning recommendation engine, or an employee service assistant. In the new model, AI in HR means a change to how work is designed, measured, staffed, supervised, disputed, and funded. That is closer to operating model design than HR tech procurement.
The CHRO cannot win that argument with a pilot slide.
The CHRO needs a work map.
Workforce Planning Still Counts Heads
The June 2026 McKinsey HR Monitor surveyed about 1,300 HR professionals and 5,500 employees across ten countries. Its workforce-planning finding should make every AI transformation deck uncomfortable: only 11% of organizations take a long-term perspective, while workforce planning remains focused on short-term headcount planning.
This is the wrong planning unit for human-agent work.
Headcount tells finance how many people sit in the budget. It does not tell a manager how much agent output they can trust, how much review time a team needs, how many exceptions will reach HR, which tasks should be redesigned before automation, or how much reskilling budget must be tied to a specific workflow. It also does not tell employees what part of their job will change and which skills will matter after that change.
The old workforce plan usually starts with demand, capacity, attrition, open roles, and budget. That worked when the central question was whether the company had enough people in enough roles. It breaks when one role becomes a bundle of human judgment, agent execution, manager review, policy escalation, data stewardship, and customer or employee communication.
Consider an HR shared-services team that handles employee questions about leave, payroll, benefits, internal transfers, and policy interpretation.
A traditional plan asks how many cases the team handles, how many HR agents or service reps it needs, and how many cases each person can close. An AI plan that only replaces Tier 1 answers with a chatbot still asks a similar question: how many cases can be deflected?
A human-agent workforce plan asks more.
Which cases should agents answer directly? Which answers can trigger a transaction in Workday, payroll, or a ticketing system? Which cases require a manager or HRBP? Which require legal review? Which require employee relations because the answer affects performance, pay, schedule, leave, discipline, or accommodation? How many AI answers will be disputed? How many will require correction? Which policies need rewriting because an agent cannot safely infer exceptions from old language? How much review time will managers inherit?
The capacity model changes from one column to five:
| Planning column | Meaning | Budget owner |
|---|---|---|
| Human headcount | People assigned to the work | HR and business function |
| Agent capacity | Workflows or tasks agents can execute | HR, IT, procurement |
| Oversight hours | Manager, reviewer, HRBP, legal, or specialist review time | Business function and HR |
| Exception volume | Cases routed out of automation because risk, ambiguity, or dispute appears | HR operations and employee relations |
| Reskilling budget | Training and redeployment tied to changed tasks | Learning, business leaders, finance |
That model is not elegant. It is closer to reality.
Microsoft’s 2026 Work Trend Index makes the same point from the employee side. Microsoft analyzed more than 100,000 Microsoft 365 Copilot chats and surveyed 20,000 AI users across ten markets between February 18 and April 7, 2026. It found that 49% of Copilot conversations supported cognitive work such as analysis, problem solving, evaluation, and creative thinking. The report argues that the constraint is not individual effort alone. Organizational factors - culture, manager support, and talent practices - account for twice the reported AI impact of individual effort.
That puts workforce planning inside management practice.
If employees can already use AI to perform more complex work, the company has to decide what it will recognize, reward, govern, and redesign. A worker who uses AI to analyze a dataset, draft a policy memo, prepare a client call, or test a workflow may increase their own output. The organization may still fail to capture value if the job description, approval path, manager expectation, skill model, performance review, and compliance record all assume the old work pattern.
This is where HR has more authority than it often uses.
HR owns or influences job architecture, learning, manager enablement, performance systems, internal mobility, employee communications, workforce analytics, HR policy, and culture. Those are the levers that turn individual AI use into organizational redesign. If HR leaves the conversation to tool owners, the result will be a productivity patch: more output, unclear accountability, uneven adoption, and little proof that the company changed how value is created.
Finance will then ask why the AI bill keeps rising.
The answer cannot be “people are using it.”
The answer has to be a redesigned work model with evidence.
Agents Move HR Work Into Daily Tools
The work map is urgent now because HR and finance agents are no longer confined to HR portals.
On May 28, 2026, Workday and Google Cloud announced that Workday’s Sana Self-Service Agent would be available inside Gemini Enterprise. Gemini also became the default AI model inside Sana for Workday. The companies described employees asking questions in Gemini and receiving personal answers pulled from Workday with policies and permissions applied. Managers could review team goals, approve timesheets in bulk, start performance reviews, or submit payroll input from the AI experience.
That is not a simple front-end change.
It changes where HR work starts.
In the old model, HR owned the portal, the case intake, the policy page, the workflow button, and much of the employee service journey. In the emerging model, the employee may start inside Gemini, Copilot, Teams, Slack, ServiceNow, a browser sidebar, or a company-specific agent interface. Workday, Google Cloud, Microsoft, ServiceNow, and other platforms want the same outcome: work happens where the employee already is, while governed systems execute behind the scenes.
This creates a product advantage. It also creates an operating conflict.
If a manager approves timesheets through an AI surface, who confirms that the manager understood the exception? If an employee asks about leave and the agent answers from policy and personal data, who handles an incorrect answer? If a performance review starts through an agent, where does the draft live, who can inspect the prompt trail, and how does an employee challenge a summary? If a payroll input is submitted without leaving the AI experience, which system owns the audit record?
The front door moves, but accountability stays with the employer.
ServiceNow is making a similar claim from the control layer. Its May 2026 AI Control Tower expansion emphasized discover, observe, govern, secure, and measure across AI deployed in any enterprise system. The company highlighted cost tracking, ROI dashboards, least-privilege enforcement, real-time shutdown when an agent goes off script, and a control layer grounded in workflow data, the CMDB, and business context.
The language matters. ServiceNow is selling more than a better HR assistant. It is selling a way to see, govern, secure, and measure AI work across systems.
HR should read that as a warning.
If HR does not define the human side of the work, platforms will define the machine side first. The control tower will know agent identity, cost, workflow, model route, access, and risk. Finance will see dashboards. IT will see integrations. Security will see access graphs. HR may still be left trying to explain whether the work redesign helped people, improved decisions, reduced manager burden, protected employee trust, or created a new appeal queue.
The daily tool shift also changes adoption politics.
Employees do not experience “HR transformation” as an operating model. They experience a leave answer that is right or wrong, a performance summary that feels fair or unfair, a manager packet that saves time or creates confusion, a payroll case that resolves faster or loops through more systems, and a career recommendation that opens a path or locks them into a profile.
That means HR cannot measure the shift only through usage.
Usage says whether people touched the agent. It does not say whether work improved.
The useful metric might be time to correct a wrong payroll answer, percentage of AI-generated manager recommendations changed after review, employee trust after an agent-assisted case, case deflection without recontact, reduction in manager prep time, appeal volume by workflow, or cost per resolved issue after counting agent, integration, review, and rework costs.
Those metrics require HR, IT, finance, legal, and managers to agree on the work design.
The interface will not do that for them.
Finance Will Not Fund Vibes
SHRM’s State of AI in HR 2026 gives the finance room its opening line.
In a sample of 1,908 HR professionals, 39% reported AI adoption in their HR functions, 7% intended to launch AI in HR this year, and another 23% had AI launched elsewhere in the organization. That means 62% were using AI somewhere in the organization. Yet HR use remained concentrated: recruiting at 27%, HR technology at 21%, learning and development at 17%, and employee experience at 14%.
The measurement gap was larger. SHRM found that only 16% of HR professionals use their own ROI metric to assess AI success, while 56% do not formally measure AI investment success at all.
That is not sustainable.
The early wave of HR AI could survive on experimentation language. Teams could say they were improving recruiter productivity, speeding employee service, personalizing learning, or reducing administrative load. In 2026, that language has to survive budget review. Workday, Microsoft, Google, ServiceNow, Salesforce, Oracle, SAP, ADP, and a long list of HR tech vendors are pushing agents, credits, workflows, AI models, marketplace listings, and control tools into the same buyer account.
Finance will not fund every promise.
The CFO will ask which work changed, which costs moved, which risks fell, which outcomes improved, and which new costs appeared. The CHRO should expect that question because HR has already seen the same pattern in recruiting. AI interviews claimed speed. Buyers then had to count candidate exits, disclosure duties, appeal queues, signal quality, human review, proof-of-person cost, and vendor evidence obligations. The headline efficiency number did not survive contact with operations.
Enterprise AI transformation will follow the same path.
An employee service agent may reduce routine tickets but increase review work for sensitive cases. A performance-summary assistant may save manager drafting time but create evidence, explanation, and appeal obligations. A learning agent may personalize content but still fail to close a strategic skill gap. A workforce-planning agent may create scenarios faster while hiding weak assumptions about agent capacity or redeployment. A recruiting agent may process more candidates while forcing new fraud, identity, and trust controls.
The work redesign ledger has to separate four kinds of value:
| Value claim | Proof HR needs | Common failure |
|---|---|---|
| Productivity | Time saved after review, rework, and exceptions | Counting draft speed as finished work |
| Cost reduction | Net cost after licenses, integrations, manager review, governance, and vendor support | Ignoring consumption and change-management costs |
| Decision quality | Better outcomes, fewer corrections, clearer evidence, lower dispute rates | Treating automation rate as quality |
| Employee trust | Adoption with confidence, fair appeals, transparent use, lower friction | Measuring clicks instead of lived experience |
This is why HR AI measurement should be designed with finance before procurement.
If measurement begins after deployment, the vendor dashboard wins by default. Dashboards are useful, but they often measure system activity: active users, tasks completed, cases deflected, answers generated, credits consumed, or workflows automated. A business case needs a different unit. It needs to connect an AI-assisted workflow to a business result, a human workload change, a risk record, and a cost line.
This is uncomfortable because HR often owns soft outcomes that are hard to tie to operating results.
Trust, culture, employee experience, manager effectiveness, internal mobility, and learning readiness matter. They also need operational proxies. Did employee-service recontact fall? Did payroll corrections shrink? Did manager calibration time improve without more disputes? Did internal candidates move faster into scarce roles? Did AI-assisted learning close a real skill shortage? Did offer acceptance improve? Did attrition change in teams where work redesign was done carefully?
Those questions belong in the funding model.
Deloitte’s 2026 Global Human Capital Trends argues that many organizations still are not intentionally designing how humans and machines interact, and that those which redesign roles, workflows, and decision-making around human-AI collaboration are more likely to exceed return expectations. Deloitte also frames advantage as a shift from allocating talent in static structures to orchestrating people, skills, data, and technology in real time.
That is finance language, even when it appears in an HR report.
Static allocation is a budget problem. Real-time orchestration is a control problem. Human-machine decision design is a risk problem. Learning and redeployment are investment problems. HR cannot keep those in separate decks.
The CHRO needs one ledger.
Managers Inherit the Redesign
The cleanest AI transformation story says agents absorb routine work and humans move to judgment.
Managers hear a different story: more things to check.
Microsoft’s Work Trend Index says AI puts a premium on judgment, clarity of intent, and work design. McKinsey describes new role archetypes such as builders, orchestrators, and strategists. ADP’s 2026 HR trends release says organizations are using skills-based approaches and role redesign to align talent with business needs. Those ideas are right. They also create a manager-capacity problem that HR cannot hide.
A human-agent workflow usually adds at least seven manager tasks:
- Framing the work clearly enough for an agent or AI-enabled workflow to act
- Reviewing output quality before it affects an employee, customer, or candidate
- Explaining AI-assisted decisions to people who are affected
- Handling exceptions when a policy, person, or case does not fit the default
- Recording why they accepted, changed, or rejected an AI recommendation
- Identifying skill gaps created by redesigned work
- Protecting team trust when employees suspect automation is changing opportunity, pay, or evaluation
That is real work.
It does not appear automatically in a staffing model. If HR tells the business that AI will remove administrative burden but fails to budget review time, managers become the shock absorber. They absorb ambiguity from the tool, anxiety from employees, pressure from finance, and questions from legal or compliance. They also become the person employees blame when an AI-assisted recommendation feels unfair.
This is already visible in recruiting. Candidate appeal queues require recruiters and hiring managers to review transcripts, scores, structured notes, identity signals, and vendor evidence. Fraud desks require managers to recheck work done by suspicious hires. Procurement stop-loss reviews require TA Ops and hiring leaders to decide whether a vendor’s promised outcome held. Every “automation” story creates human review work somewhere.
Post-hire work will do the same.
Performance management is the obvious case. An AI tool can summarize feedback, draft review language, flag goal progress, compare competencies, or suggest calibration topics. The manager still owns the judgment. If the summary includes inaccurate context, stale project data, biased language, or missing employee contributions, the manager must correct it. If the employee challenges the review, HR needs evidence of what the agent generated, what the manager changed, and why the final decision was made.
Scheduling and frontline operations create another version. An AI scheduling system may optimize coverage, labor cost, skills, availability, and legal constraints. The manager still has to handle employee hardship, last-minute illness, local norms, fairness perceptions, and override requests. If the schedule is efficient but trusted less, the system has not improved work.
Learning and internal mobility create a quieter version. An AI career assistant may recommend courses or roles based on skills data. The manager still has to decide whether the employee can spend time on learning, whether the skill is relevant to future work, and whether internal movement helps or hurts current team capacity. A recommendation without manager capacity can become another ignored notification.
HR’s job is to make this manager work visible.
That starts with capacity math. If a workflow generates 10,000 AI-assisted outputs per month and 8% need human review, who reviews 800 items? How long does each review take? What authority does the reviewer have? What evidence do they need? What happens if the queue backs up? Which decisions can wait, and which affect pay, work authorization, leave, performance, or candidate rights?
The answer should shape the redesign.
Some workflows should stay narrow because review capacity is scarce. Some should automate only low-risk parts. Some should route to HR operations rather than line managers. Some should require employees to see when AI was used and how to challenge it. Some should wait until the policy language, data quality, and manager training improve.
HR has to defend those limits.
The pressure will come from vendors and executives who want the faster story. Agents will promise to move work through the system. The CHRO has to ask whether the organization has enough human capacity to own the outputs. That is not anti-AI. It is the difference between automation and operating design.
Managers will not ask for a human-agent operating model in those words.
They will ask why their team has more review work after buying a tool that was supposed to save time.
HR Has to Become Its Own Pilot
The dual mandate only works if HR can show its own redesign.
That does not mean HR should automate every process first. It means HR should choose a few high-volume, high-friction workflows where the organization can see the new model end to end: task redesign, agent role, human review, exception path, measurement, employee communication, evidence record, and cost attribution.
Employee service is a strong candidate.
It has volume. It touches real employees. It includes low-risk questions and high-risk exceptions. It spans policy, HRIS, payroll, benefits, leave, manager approvals, and case management. It has visible metrics: response time, first-contact resolution, recontact, correction rate, escalation rate, employee satisfaction, cost per case, and trust after resolution.
But the pilot should not be framed as “deflect more tickets.”
It should be framed as “redesign employee service work.” That means deciding which answers agents can give directly, which transactions they can initiate, which cases require human review, which policies need structured logic, which employee populations need special handling, which records must be retained, and which costs count when finance asks for value.
Recruiting can also serve as a pilot, but it carries fatigue. Many HR teams have already spent months on AI interviews, candidate communications, scheduling automation, sourcing assistants, fraud checks, and appeal paths. The value of another recruiting pilot depends on whether it connects to the broader work model. A proof-of-person workflow is useful. A fraud desk is useful. A recruiter signal recovery model is useful. Yet the dual mandate requires HR to show that it can move beyond talent acquisition and redesign work across the employee lifecycle.
Performance, learning, and workforce planning may be more powerful test beds.
Performance workflows reveal judgment and evidence. Learning workflows reveal skills and redeployment. Workforce planning reveals finance and capacity. Together they force HR to answer the real human-agent questions:
- Which tasks will agents perform?
- Which decisions stay with people?
- Which managers can review output responsibly?
- Which employees need new skills because their task mix changed?
- Which metrics prove work improved?
- Which evidence records survive an employee dispute or audit?
- Which costs belong to HR, IT, finance, or the business unit?
ADP’s 2026 HR trends release points toward the same center of gravity. It emphasizes skills-based job design, agentic AI, responsible AI governance, transparency, HR-IT collaboration, and role redesign. It also cites ADP Market Pulse Study data from April 2025: 84% of large organizations agreed AI can streamline processes without replacing employees, compared with 76% of midsized organizations and 73% of small organizations.
That belief is useful only if work is redesigned around it.
“AI will not replace employees” is a slogan until the company names the work that changes, the skills that matter, the review capacity required, and the employee path from old tasks to new value. Without that detail, employees hear a promise and managers receive a workload.
HR’s pilot should therefore produce artifacts that other functions can copy.
It should produce a task map, a human-agent responsibility chart, a reviewer-capacity model, an exception policy, a measurement ledger, a communication template, an appeal path, a vendor evidence checklist, and a finance view. Those artifacts are more valuable than another adoption slide because they turn the dual mandate into reusable operating discipline.
The CHRO can then walk into a sales, support, finance, or engineering redesign meeting with proof.
Here is how HR mapped tasks. Here is where the agent helped. Here is where people stayed in control. Here is what managers had to review. Here is how employees were told. Here is how disputes were handled. Here is what finance counted. Here is what changed after 90 days.
That proof makes HR credible.
It also protects HR from becoming the help desk for AI anxiety without authority over the design choices that created it.
A New Workforce Plan Needs Five Columns
The useful output of the dual mandate is not a manifesto. It is a new planning file.
The old file has roles, headcount, vacancies, attrition, hiring plan, labor cost, and maybe skills. The new file still needs those fields, but it also needs columns that describe work itself. Human-agent redesign without a planning file will stay at the workshop level.
A practical version has five columns.
First, human headcount. This remains necessary. People still do the work, own judgment, build relationships, handle exceptions, manage teams, and carry legal and ethical accountability. Cutting this column would be naive. It simply cannot be the only column.
Second, agent capacity. This should not mean a vague count of bots. It should describe the workflows or task bundles agents can execute, the systems they can touch, the volume they can handle, the risk tier, the cost meter, and the evidence they produce. An employee service agent, recruiting agent, payroll agent, learning agent, and workforce-planning agent do different work. Treating all of them as “AI capacity” hides the operating details.
Third, oversight hours. Every high-value AI workflow needs someone to frame, review, override, explain, or improve it. Oversight may sit with managers, HR operations, HRBPs, legal, compliance, IT, security, finance, or specialist reviewers. The planning file should estimate the load, not assume it away.
Fourth, exception and appeal volume. AI value falls apart when the exception path is invisible. Employee relations, payroll, leave, performance, compensation, scheduling, recruiting, and internal mobility all contain cases where a default answer can be wrong or harmful. Planning should estimate how many cases will leave automation, how fast they need response, what evidence they require, and who owns the backlog.
Fifth, reskilling and redeployment budget. If AI changes task mix, HR has to fund movement. That may mean AI literacy, but it also means judgment training, policy reasoning, data interpretation, vendor management, employee communication, prompt framing, quality review, and business-specific capability building. The budget should connect to the changed work, not a generic course catalog.
Those columns also force the CHRO and CFO into the same conversation.
If finance asks whether AI reduces headcount, HR can answer with a richer model: this workflow reduces routine case work by a specific amount, increases review hours by a specific amount, shifts manager time, requires a reskilling investment, reduces rework, changes vendor costs, and creates measurable quality or trust outcomes. Some cases will justify headcount reduction. Some will justify redeployment. Some will justify no automation until data, policy, or review capacity improves.
That is a more honest budget discussion.
It also changes vendor buying. A vendor demo that shows an agent completing a task is only the first row. The buyer should ask:
| Buyer question | Reason |
|---|---|
| Which human task is removed, changed, or added? | Productivity depends on net work, not feature output |
| Which manager or specialist reviews the result? | Oversight is capacity |
| Which exception cases leave the workflow? | Risk hides in the edge cases |
| Which evidence record is retained? | Disputes and audits need a trace |
| Which cost meter fires? | Usage-based AI can move cost into hidden lines |
| Which employee skill must change? | Work redesign without capability building stalls |
This is where skills inventory and agent inventory will eventually meet.
Companies have spent years trying to build skills graphs. They want to know what employees can do, which skills are scarce, where to hire, where to train, and how to move people internally. At the same time, agent registries and control towers are starting to map what agents can do, which systems they can access, which workflows they execute, and which risks they carry.
Those maps cannot stay separate forever.
If an agent can complete parts of a role, the skills graph needs to know which human skills remain valuable. If a worker can supervise or improve agent output, the agent registry needs to know who is qualified to review it. If a workflow combines employee skill and agent capacity, workforce planning needs both maps. If an employee wants internal mobility, the company needs to know whether the target role requires execution skill, agent orchestration skill, judgment skill, or relationship skill.
That is a control map for work.
It belongs partly to HR, partly to IT, partly to security, partly to finance, and partly to the business. HR should not pretend to own the whole map. It should make sure the human side is not missing.
Without that, the agent inventory will be precise and the workforce plan will be vague.
That would be the worst possible split.
Ninety Days Later, the Org Chart Has Evidence
The first test of HR’s dual mandate should be visible after 90 days, not three years.
Pick one workflow. Redesign it. Name the human tasks, agent tasks, review points, exception paths, evidence records, cost meters, and skill changes. Tell employees what changed. Train managers on the new judgment work. Let finance see the ledger. Let legal see the evidence file. Let IT see the access model. Let employees see how to challenge a bad result.
Then look at the work again.
Did the workflow move faster after review and rework? Did managers actually save time? Did employee trust hold? Did exceptions get handled within the promised window? Did the agent create new hidden costs? Did the work require skills HR had not planned for? Did a vendor dashboard conflict with the company’s own measurement? Did the redesigned role still make sense when a real person had to live inside it?
Those are better questions than whether the AI pilot launched.
They also put HR in the right room. The CHRO is not there to celebrate adoption. The CHRO is there to decide how people, agents, managers, data, evidence, cost, and trust fit into a working system.
That is why the old org chart is such a weak tool for this moment.
The old chart tells the company who reports to whom. It does not show where judgment moves after an agent drafts the work. It does not show who pays for review time. It does not show which employee can challenge an AI-assisted decision. It does not show which skills become scarce. It does not show which system holds the proof. It does not show whether the work got better.
HR needs a different artifact on the table.
It needs a work map that shows task ownership, agent role, human judgment, exception handling, evidence, cost, and capability building. It needs a planning file that counts people and agents together without pretending they are interchangeable. It needs a measurement ledger that finance can understand. It needs a manager model that makes review work visible. It needs an employee path that treats trust as operating infrastructure, not sentiment.
The dual mandate sounds abstract until the first redesigned workflow goes live.
Then it becomes concrete. A manager approves a timesheet through Gemini. An employee asks a payroll question through an agent. A recruiter reviews an AI interview appeal. A learning system recommends a new role. A ServiceNow control view shows AI cost and risk. A finance partner asks whether the work changed enough to justify the bill.
At that moment, HR either has the map or it has a collection of tools.
The difference will decide whether AI redesigns work, or whether old work simply gets a faster interface.
This article provides a deep analysis of HR’s dual mandate in human-agent work redesign. Published June 10, 2026.