AI Adaptive People Function/Roles & Skills

Companion paper · Roles, skills, sequencing

Roles, skills, and the reskilling limit.

What the HR org chart looks like when agents do the work, why most of the middle layer cannot be reskilled into it, and where the new shape is hired from.

Catalyst question

Fiona Cicconi

Chief People Officer, Alphabet

Author

Rahul Jindal

Drawing on 20 years driving transformations

Read the eight skill primitives Skip to role evolutions

What this paper is

The AI Adaptive People Function paper named the operating instruments the function needs in the agentic era. This paper names who runs them. The two questions are co-equal. A function that can describe what it should be doing but cannot describe who is in the seats has not finished the work.

The dominant narrative inside HR is that agents handle the tasks, ICs become managers of agents, and the function reskills its way into the new shape. The narrative is correct at the slogan level and wrong at the seat level. The seat-level shifts are sharper, more uneven, and more political than "manage the agent" captures. Some roles compress 5x. Some roles split into two new roles. Some roles expand because a long-standing bottleneck disappears. Some incumbent populations cross. Most do not. The function that pretends otherwise loses three years to reskilling theater while AI-native competitors hire the new shape directly.

This paper has three jobs. First, name the cognitive-shape shift that is actually happening underneath the role-level changes (it is not "more technical"). Second, name the role-by-role evolution at a level of specificity a CHRO can take to a workforce-planning meeting. Third, name the limit of reskilling. The strongest available evidence puts the cross-rate for the middle layer of disrupted functions at roughly 20-30%. Pretending otherwise is the dominant failure mode of HR transformations.

The paper assumes the parent paper's framing (EMI, Purpose, Same-Breath, capability supply chain). It does not re-derive them. Read AI Adaptive People Function first if you have not.

The reframe

The cognitive-shape shift, not the soft-vs-hard skill shift

The shift in HR is not soft skills to hard skills. The middle layer that survives in HR will keep most of the soft-skill stack. The shift is a different cognitive shape underneath the work.

The old shape: deterministic, case-by-case, sequential. A case lands; you read it; you apply policy; you close it. A cycle runs; you administer it; you calibrate; you close. A program ships; you measure; you adjust. The mental model is execution of a defined process on discrete cases.

The new shape: probabilistic, system-by-system, continuous. An agent handles 10,000 cases a week with a 3% error rate. You design what counts as an error, set the error budget, sample the worst 1% for review, route the genuinely-ambiguous to humans, retrain on the edge cases. The mental model is design and operation of a system under uncertainty.

This is not "more technical." Most engineers do not think this way either; SREs and Trust & Safety leads and platform PMs do. Most analysts do not; ML engineers and ad-auction designers and marketplace pricing leads do. The cognitive shape lives in specific functions, not at a generic "tech" altitude. It is not a step on the same staircase the middle layer of HR is currently climbing. It is a different staircase.

The implication for hiring: the people who think this way are findable. They sit in concentrated talent pools — Trust & Safety, platform engineering, ad-tech, regulatory product, mechanism-design researchers, internal mobility platforms. Hiring is a pool-by-pool operation, not a generic upskill.

The middle-layer claim

Three layers, three different stories

Three layers of HR move differently:

Senior (CPO direct reports, function GMs, VP-band CoE leads): adapts. The senior layer was already doing system-shaped work, often manually. Their old job was approximately the new job, with worse tools. The reskilling investment for the senior layer is real but contained. Modest curriculum, plus coaching, plus repeated exposure to operational decisions in the new shape. Most senior HR leaders cross with deliberate investment.

Junior (early-career, new hires, L3-L4 equivalent): replaces cleanly. The function does not reskill them; it changes the JD and the candidate pool. Net-new HR hires from 2027 onward come in shaped for the new world. Junior HR talent is in fact the easiest of the three layers to refit, because the function is rebuilding the entry point.

Middle (L5-L7, the bulk of the function — HRBPs, comp analysts, L&D PMs, scaled ops leads, immigration case managers, benefits analysts): structurally hard to bridge. This is where the political and economic pain concentrates. The middle-layer claim is the load-bearing argument of this paper. It is also the most uncomfortable. Naming it, planning around it, and being honest with the function about it is the difference between a transformation that works and a reskilling program that becomes theater.

The reskilling limit

What the evidence actually says

The claim is sharp because the evidence is sharper than HR generally treats it.

Cognitive ability predicts training success more strongly than any other measured factor. The Schmidt-Hunter meta-analyses, replicated across eight studies, put the correlation at 0.56 to 0.67. Cognitive aptitude tests are roughly 1.6x more predictive of training success than interviews and 4x more than years of experience. This is the strongest finding in 100 years of organizational psychology and the one most consistently understated by HR programs.

The historical record on cohort transitions is unkind to the optimistic narrative. Telephone operators, the largest occupation for young women in the US, were absorbed not by reskilling but by cohort replacement: hiring stopped, existing operators kept working until retirement, the next generation entered different occupations (Feigenbaum and Gross, NBER w28061). Bank tellers, the canonical "reskilling success story," produce only roughly 4% transitions to loan-officer roles per the Burning Glass Institute; the popular Bessen narrative captures gradual role mutation over decades, not a reskilling program. Floor traders did not become quants; the few who survived monetized network and tacit market knowledge into adjacent roles.

The corporate reskilling record is structurally honest about its silence. AT&T's Future Ready program has invested over $1B since 2013 with 180,000 of 203,000 employees participating; AT&T does not publicly disclose the absorption rate, which is to say what percentage of badge-earners actually moved into targeted future-skill roles. Amazon's Upskilling 2025, scaled to $2.5B, similarly reports inputs, not absorption. Walmart's Live Better U is the rare program with rigorous third-party evaluation; the population (associates) and goal (retention plus general education) is not the L6-incumbent-into-new-technical-role transition HR is contemplating.

BCG's analysis of the cost-benefit of reskilling, modeled across disrupted workforces, finds that reskilling is economically favorable for roughly 25% of disrupted workers when the company bears all costs. Below that line, the math does not work. The 20-30% hypothesis this paper centers on is the more rigorous reading of the literature, not the contrarian one.

The aptitude predictor question matters because the alternative is spending three years and tens of millions on programs that compress a 25% population into the same 25% the function could have identified ex ante in four months. Three signals predict crossing; three more do not.

The aptitude predictor

Identify the bridge cohort before any program runs

What predicts:

Signal 1: Cognitive aptitude. The single strongest predictor. Off-the-shelf assessments exist (CCAT, Wonderlic, Raven's Progressive Matrices). The validity is documented; the political resistance to running them on existing employees is the binding constraint, not the evidence.

Signal 2: Learning agility. Distinct from cognitive aptitude (low correlation, incremental validity), and the strongest predictor specifically for novel and ambiguous transitions. The TALENTx7 and Lombardo-Eichinger work documents this. Structured behavioral assessment, not unstructured interview.

Signal 3: Early agent-tooling adoption rate. The most observable signal, and the one no other employer has yet. People who, today, are actively using AI tools in their current work and getting steadily better at it have already demonstrated the receptive cognitive shape. People who have used AI tools twice and given up have not. Pull this from existing telemetry; do not survey for it.

What does not predict, even though every HR program treats them as if they do:

Tenure. Predicts weakly. Some 25-year HRBPs cross gracefully; some 5-year HRBPs cannot.

Title or band. Predicts weakly. Some L5s cross; some L7s do not.

Self-reported interest in AI. Self-report predicts almost nothing because everyone says they are interested.

Manager nominations alone. Susceptible to halo, recency, and similarity bias. They are an input, not an output.

Prior performance ratings in the old role. Weak predictor of performance in the new role when the new role is fundamentally different. This is the central HR mistake: performance in the to-be-disrupted job is not transferable evidence.

The function can identify the bridge cohort with three observation methods that take less than four months and cost less than $200K: cognitive aptitude assessment, structured learning-agility assessment, and agent-tooling adoption telemetry. Triangulating across the three produces a tiered list:

Tier A · 10-15%

High signal across all three methods. Invest heavily; these are the future leaders. Six to twelve month structured pathway with paired mentorship in the new role and explicit role-absorption commitment from the receiving manager.

Tier B · 15-20%

High signal on at least two methods. Invest moderately; coach the third dimension. Many will cross with the right scaffolding.

Tier C · 30-40%

Mixed signal. Invest selectively in adjacent skills; honest career conversations about residual or sideways roles. Do not pretend full reskilling will work for this cohort.

Tier D · 30-40%

Low signal across methods. Plan dignified transition out, not transformation up. Generous severance, outplacement, internal mobility to less-disrupted functions.

The function that runs this triage in Q1 has a different transformation arc than one that runs reskilling for everyone in Q1. The first compresses transformation timeline by 18 months. The second loses 18 months and ends up at the same place anyway.

Tier 1 · Skill primitives

The eight primitives, unbundled

The phrase "system skills" hides eight distinct primitives that hire from different talent pools and require different curriculum if developed internally. The single most important move in this paper is unbundling them. Functions that frame the shift as "the team needs to be more technical" miss the texture and end up hiring generic data analysts when they need T&S policy engineers and behavioral economists. The hire is wrong because the diagnosis was generic.

The eight primitives split into three pools. The technical-leaning four (decision-tree authoring, error-budget, eval design, probabilistic reasoning) are hired from tech and quantitative pools. The product-leaning two (agent PM, system design under uncertainty) are hired from product orgs. The strategy-leaning two (mechanism design, Trust & Safety thinking) are hired from specialized academic and policy pools. A function that hires "AI talent" generically gets the wrong distribution. The right hire for an agent PM role is not the right hire for a mechanism-design role.

Skill primitive

Decision-tree and policy-as-code authoring

Surface the implicit context that human discretion was filling in.

What it is

Converting fuzzy human policy into explicit branching logic with explicit inputs and outputs. The hardest part is surfacing the implicit context that human discretion was filling in.

Concrete example

Writing the Leave-of-Absence agent's policy. "Approve if eligibility met" is not a policy; it is a placeholder. The actual policy says: approve if (a) tenure > 90 days AND (b) prior LOA in last 12 months not denied for the same reason AND (c) supervisor not flagged for review AND (d) jurisdiction not in the manual-only list. The author has to discover (c) and (d) by interviewing humans who never wrote them down. The discovery process is the work.

Hire from

→Trust & Safety policy teams (Meta, YouTube, TikTok, OpenAI; published Community Standards and Model Spec are operator-grade artifacts)
→Legal Operations (decision-tree authoring is a daily craft in modern legal ops)
→Regulatory product (financial services compliance product, healthcare compliance product)
→Claims-handling design teams (insurance, third-party administration)

Roles that need it

·Global Process Owners
·HR Policy Engineers (new)
·Comp and Performance Management strategy

Skill primitive

Error-budget reasoning

Choose the failure rate; design accountability around the rate, not the case.

What it is

Accepting that systems will fail, choosing acceptable failure rates, and designing accountability around the rate rather than the individual case. SRE-shaped thinking applied to HR.

Concrete example

The LOA agent will incorrectly route 1.5% of cases. The function decides: 0.5% gets flagged for human review (cost: 8 FTEs), 1% sits below the human-review threshold and is recovered through the appeal flow (cost: 3 FTEs plus reviewer well-being investment). The right answer is the one that minimizes total harm, not the one that drives error to zero. Driving error to zero forces the agent to escalate so much that humans become the bottleneck and the system breaks.

Hire from

→Site Reliability Engineering and platform engineering
→Trust & Safety engineering (already runs error budgets on classifiers and enforcement)
→Reliability product roles in infrastructure orgs

Roles that need it

·HR Agent PMs (new)
·Global Process Owners
·Scaled HR Ops Tier 3 leadership

Skill primitive

Agent product management

Treat the agent the way a tech PM treats a feature.

What it is

Owning a specific HR agent surface as a product. Roadmap, metrics, escalation rate, error budget, redteam, deprecation milestones. Treats the agent the way a tech PM treats a feature, not the way a process owner treats a workflow.

Concrete example

The Promo Cycle Agent PM owns: time-to-decision, manager-override rate, calibration-error rate, time-to-cycle-close, reviewer well-being score for escalation cases, deprecation milestones for the human-only fallback. Metrics are public inside the function. Quarterly business review with the CPO. Roadmap published with quarterly OKRs.

Hire from

→HR-tech product PMs (Workday, Lattice, Gusto, Eightfold, Gloat)
→Trust & Safety product PMs (the closest direct analog)
→Internal-platform PMs at large tech companies

Roles that need it

·HR Agent PMs (new role; one per agent surface)

Skill primitive

Eval design

Figure out what "matter" means for a probabilistic system.

What it is

Designing tests that measure agent quality on the dimensions that matter. Distinct from product analytics; closer to ML evaluation. The hardest part is figuring out what "matter" means for a probabilistic system that affects humans.

Concrete example

The Comp Recommendation Agent ships. The naive eval is "accuracy vs. human comp recommendation." The right eval is: equity by demographic, drift over time, alignment with the published comp philosophy, manager override patterns, employee perception scores, retention impact at 6/12/24 months. Each is a separate eval suite. Each is run quarterly. The Responsible AI playbook from the Trust & Safety world (Meta's Casual Conversations v2 dataset, OpenAI's external red-teaming methodology in arXiv 2503.16431) is the operational template.

Hire from

→Machine learning engineering
→Trust & Safety evaluation teams
→AI-native product orgs (the eval-design discipline is most mature here)

Roles that need it

·HR Agent PMs
·HR Trust & Safety (new)
·Responsible AI partners embedded in CoEs

Skill primitive

Mechanism design

Design the system; the agent operates inside it.

What it is

Designing systems where individuals' rational responses produce the aggregate outcome the org wants. Theory: Hurwicz, Maskin, Myerson. Practical analogs: ad auctions, marketplace pricing, two-sided matching. The discipline that turns vague comp philosophy into operating instruments.

Concrete example

Dynamic comp. The naive design (continuous AI-recommended adjustments based on market data) creates an internal arms race and erodes retention of high-judgment talent. The mechanism-designed alternative: floor and ceiling bands by role, transparent triggers for adjustments, capped frequency, predictable decay. The agent operates inside the bands; the bands are the mechanism. Same logic for performance feedback rhythm, internal mobility, succession.

Hire from

→Behavioral economics academics (especially marketplace and matching researchers)
→Marketplace pricing leads (consumer marketplaces, freight, two-sided platforms)
→Ad-auction product
→Compensation strategy roles at companies running internal market structures

Roles that need it

·Comp strategy (new shape)
·Performance Management strategy (new shape)
·Talent Intelligence

Skill primitive

Trust & Safety thinking

Bias audit, fairness review, redteam, escalation routing, reviewer well-being. The discipline of running an agentic system that affects humans without producing systemic harm.

What it is

Bias audit, fairness review, redteam, escalation routing, reviewer well-being. The discipline of running an agentic system at scale without producing systemic harm. The DTSP Safe Framework, now ISO/IEC 25389, is the institutionalized version.

Concrete example

The HR Trust & Safety team runs a quarterly bias audit on every agent that touches comp, perf, or hiring. The audit measures outcomes by demographic, identifies drift, surfaces failure modes, forces remediation before the next quarter. The audit is a standing operational instrument with veto authority on agent launches, not a one-time launch check. The HR analog of Meta's Responsible AI team plus OpenAI's external red-teaming program.

Hire from

→Trust & Safety functions in large tech (the cleanest direct analog; the field has been institutionalizing for a decade)
→Fairness research teams (academic, AI policy, civil-rights ML)
→Regulatory affairs (especially anti-discrimination law)

Roles that need it

·HR Trust & Safety (new; one VP-band leader required from day one)

Skill primitive

Probabilistic reasoning

The disposition to think in distributions and base rates.

What it is

Comfort with distributions, error rates, base rates, and reasoning under uncertainty. Not statistics per se; the disposition to think probabilistically. A meta-skill that enables most of the other seven.

Concrete example

The workforce planner asks "how many headcount do we need for the new agent?" The probabilistic version asks "given the agent fails 3% of the time and 70% of failures need human escalation, with cases averaging 25 minutes of human review and arriving Poisson-distributed at peak rate of 200/hour, what is the staffing for 95th-percentile coverage?" The second question gets answered correctly; the first does not.

Hire from

→Actuarial science
→Machine learning engineering
→Ad-tech and marketplace pricing
→Operations research and industrial engineering

Roles that need it

·Strategy & Planning
·Workforce Architects (new)
·Global Process Owners

Skill primitive

System design under uncertainty

The most senior of the eight. Hardest to teach; most easily recognized.

What it is

Designing organizational and technical systems with feedback loops, anticipating failure modes, accounting for second-order effects. The most senior of the eight primitives. Hardest to teach; most easily recognized in the wild.

Concrete example

The function decides to launch the LOA agent in three regions. The system-design question is not "how do we deploy?" It is: how do we know the agent is failing differently in each region? How do we surface failures to the right humans within the right SLA? How do we prevent the success in region 1 from blinding us to failure modes in region 3? What is the deprecation plan if the agent fails to clear the bar in 6 months? Without this primitive at the senior leadership of the function, every other investment leaks.

Hire from

→Senior product leaders from infrastructure orgs
→Complex-systems researchers and consultancies
→Ex-COOs of agentic-product companies

Roles that need it

·Workforce Architect (new)
·CPO direct reports
·Senior Global Process Owners

Tier 2 · Role evolutions

Twelve roles, role-by-role

For each role: what dies, what grows, the reskill viability, where to hire from, and the 5-year arc. The 2-3 year horizon dominates; the 5-year arc is the direction the role settles into.

Role evolution

People Partners (HRBPs)

The role does not disappear. It goes more senior. Compression ratio is at least 2x.

What dies

Information retrieval, drafting, meeting prep, basic pattern detection on org data, policy explanation, talent calibration ops. These tasks evaporate over 24-36 months as agents handle them.

What grows

Exec coaching, political navigation, sensitive investigation, judgment under high information, integration of agent insights into leader conversations. The relational layer concentrates and gets harder, not easier.

Reskill viability

High for top quintile; low for L4-L5 layer

The top quintile was already doing the senior work. The L4-L5 layer's role is gone; the senior version is a different job, not a promotion. The middle is the painful conversation.

Hire from

→Senior internal coaches at scale-up product cos
→Ex-strategy consultants who pivoted to people work (the McKinsey People & Org alumni cohort)
→Exec advisory firms (the partner-band of org-development practices)
→Behavioral org-design consultancies

5-year arc

Collapses into something closer to "Org Health Architect": fewer in the seat, more senior, more diagnostic, less ceremonial. Reports to CPO at smaller orgs; reports to a function COO at larger orgs.

Role evolution

Scaled HR Operations (Case Management)

Hardest-hit function. 70-90% of current case volume is policy lookup dressed up as casework.

What dies

Case throughput, FAQ handling, ticket triage, routine policy explanation. The L1 case-handler role disappears within 18-24 months.

What grows

Ambiguity adjudication (Tier 3), agent quality assurance, escalation specialism, edge-case translation back to GPOs as policy refinement. The work concentrates and becomes harder per case.

Reskill viability

Very low

The function self-selected for high-empathy, structured, repetitive work. Meta-cognitive review of agent decisions is a different mental shape, and most case handlers cannot bridge.

Hire from

→Trust & Safety reviewers from large tech (the closest direct analog; over 17,000 T&S jobs publicly listed in the US alone)
→Legal Operations
→Claims adjusters with technology-mediated review experience
→Ex-call-center QA leads who moved into agent QA
→Regulated-industry policy reviewers

5-year arc

Approximately 80% reduction in headcount. Survivors are mostly net-new hires from agent-ops backgrounds. The function is renamed (e.g., "Employee Trust & Safety," "Agent Quality") to signal the change. Reviewer well-being infrastructure (counselor access, exposure caps, rotation policies) is built in from the start, not bolted on after a Sama-style incident.

Role evolution

Global Process Owners (GPOs)

Role grows in importance and shifts shape. GPOs become product managers for agent-executed processes.

What dies

Linear process mapping, manual SOP authoring, change-management-as-comms, vendor-managed process design.

What grows

Agent surface ownership, decision-tree authoring, error-budget design, escalation routing, policy-as-code translation. Job shape moves from process designer to product manager for agent-executed processes.

Reskill viability

Medium

GPOs are already systems-thinkers; approximately 40% can cross with deliberate investment. The Lean and Six Sigma-trained GPOs are the most likely crossers because the cognitive shape transfers.

Hire from

→Technical PMs (especially platform PMs)
→Trust & Safety engineering managers
→Ops leaders from product orgs
→Regulatory product managers

5-year arc

GPO role merges with HR Agent PM. Same job. The function might rename it (e.g., "Process & Agent Owner," "People Operations PM") to signal the change.

Role evolution

Learning & Development (L&D)

Largest expansion of any HR function. The bottleneck that constrained L&D — slow content production, low personalization — disappears.

What dies

Traditional curriculum design, LMS administration, instructor-led delivery design, vendor catalog management, classroom training as default.

What grows

Skill-graph architecture, in-flow learning UX (delivered through the agents employees already use), AI-tutor design, manager-development as system, continuous skill-currency telemetry.

Reskill viability

Low

Most L&D came in as instructional designers or program managers, not learning-systems designers. The new role needs adaptive-systems design plus UX product thinking plus manager-development theory. Different parents.

Hire from

→Edtech product orgs (consumer learning apps, adaptive learning platforms)
→Learning science PhDs
→Instructional designers already working with adaptive systems
→Manager-development consultancies

5-year arc

Largest CoE; co-owns the talent-intelligence layer with Talent Management. The L&D leader becomes one of the most strategic roles in HR. The function is sometimes renamed "Learning & Talent Intelligence."

Role evolution

Compensation

Bifurcates hard. Comp ops disappears. Comp strategy becomes more important.

What dies

Market data analysis, equity modeling, comp ops, ad-hoc analyst work, manual cycle execution.

What grows

Comp philosophy plus mechanism design plus dynamic comp logic authoring. The senior comp role becomes principle-and-system design rather than analyst-and-cycle execution.

Reskill viability

High for senior; low for analyst layer

Senior comp leaders were already principle-driven; many cross. The analyst layer's role evaporates and the new role is mechanism design, which is a different parent.

Hire from

→Behavioral economics academics
→Marketplace pricing leads
→Mechanism design researchers
→Total-rewards strategists from product cos
→Ad-auction product

5-year arc

Comp ops gone. Comp strategy bigger and more strategic. The senior comp role is sometimes renamed "Total Rewards Architect."

Role evolution

Benefits

Most automatable. Role compresses fastest.

What dies

Over 90% of benefits administration and inquiry handling.

What grows

Vendor strategy, plan design, regulatory positioning. A small senior layer that was always strategic.

Reskill viability

Not the right framing

The role collapses faster than reskill timelines can run. Attrition manages the reduction. The strategic layer that survives was already strategic.

Hire from

→Existing senior benefits strategists. No new external pool needed.

5-year arc

Approximately 70% smaller. Strategic layer remains; ops layer evaporates.

Role evolution

Performance Management

Two paths. The boring one (cycle automation) is additive. The interesting one (continuous signals) requires a different function.

What dies

Cycle administration, calibration prep, manual evidence assembly, manager nudging.

What grows

Continuous-signal architecture, feedback-system design, calibration agent QA, real-time performance-comp coupling. The function shifts from cycle ops to feedback systems engineering.

Reskill viability

Low

PM teams are deeply trained in the cycle paradigm. The new role is feedback-system engineering, which has different parents (behavioral science, mechanism design, real-time product).

Hire from

→Behavioral science
→Product analytics
→Mechanism design
→Real-time feedback product orgs
→Conversational-AI product

5-year arc

Rebuilt around continuous signals. Annual cycles dissolve into rolling cadences. The function is sometimes renamed "Feedback & Recognition Systems."

Role evolution

Talent Management

Expands. The bottleneck disappears. TM finally does what it always wanted to do.

What dies

Manual succession decks, nominator-driven hi-po identification, static skill taxonomies, biennial talent-review-as-only-instrument.

What grows

Talent intelligence layer, real-time skill matching, internal mobility marketplaces, AI-augmented executive development, dynamic succession bench.

Reskill viability

Medium for strategic; low for operational

TM has more crossers than HRBP because TM was historically more analytical. The strategic layer adapts. The operational layer (succession deck preparation, hi-po nomination management) does not.

Hire from

→Talent marketplace product orgs (Gloat, Eightfold, Workday Skills Cloud)
→Org strategy consultancies
→Talent-intelligence vendors

5-year arc

TM and L&D blur. Both become layers of one talent intelligence function: skill graph plus opportunity matching plus development paths plus succession, run as one continuously-running system rather than four programs.

Role evolution

Immigration & Mobility

Most rules-based area in HR. Compresses fastest of all CoEs.

What dies

Approximately 95% of case handling, status tracking, deadline management, FAQ.

What grows

Government relations, litigation, regulatory edge cases, complex-jurisdiction strategy.

Reskill viability

Low

The compression rate (12-18 months for the bulk of cases) outruns reskill timelines. The function survives only at the senior strategic layer.

Hire from

→Immigration law (already happens)
→Regulatory affairs
→Government relations

5-year arc

Approximately 90% reduction. Survivors are senior legal and regulatory.

Role evolution

Program Managers

Splits into three different roles, only one of which transitions cleanly from existing PgM ranks.

What dies

Manual coordination, status tracking, scheduling, ceremony hosting.

What grows

HR Agent PM (new), cross-functional PgM (transitions from existing), strategy PgM (transitions from existing).

Reskill viability

Mixed

HR Agent PM is low for existing PgMs; hire from product. Cross-functional and strategy PgM are medium; existing PgMs can transition with deliberate investment.

Hire from

→Product PMs from HR-tech for HR Agent PMs
→Trust & Safety product
→Internal-platform PMs

5-year arc

Agent PMs become the dominant PgM type. Traditional PgM shrinks.

Role evolution

Strategy & Planning

Becomes the analytical brain of HR. The function that was always under-resourced gets the data work cheap.

What dies

Manual workforce planning, basic scenario modeling, headcount tracking as core artifact, role-based planning.

What grows

Probabilistic workforce design, agent-and-human ratio optimization, real-time scenario simulation, future-of-work portfolio bets, capability-supply-chain modeling.

Reskill viability

High for senior; low for operational

Senior layer was already strategic. Operational layer (headcount tracking, role-architecture maintenance) does not transfer.

Hire from

→Management consulting strategy practices
→Corporate strategy from product cos
→Operations research and industrial engineering
→Future-of-work researchers (Brookings, NBER affiliates, MIT Shaping Work)

5-year arc

Becomes the analytical brain of HR. Absorbs much of what is currently called "Workforce Architect" or "Future of Work."

Role evolution

Employee Engagement

Bifurcates. Survey ops disappears. Experience design grows.

What dies

Survey administration, pulse-as-process, basic comms drafting, listening session ops, engagement-survey-only listening.

What grows

Experience design, in-the-moment intervention design, AI-mediated feedback systems, internal NPS-equivalent design.

Reskill viability

Low

EE is heavily comms-shaped; the new role is design plus system, with different parents (CX, service design, behavioral design).

Hire from

→CX leaders from product orgs
→Product UX research leads
→Behavioral design firms
→Service design

5-year arc

Becomes "Internal CX." Operates more like a product function than an HR program.

Tier 3 · The Trust & Safety analog

T&S is roughly 8-10 years ahead of HR

The cleanest existing analog for the agentic HR function is Trust & Safety at large tech companies. T&S has already gone through the transition HR is about to start. The deeper point: T&S is roughly 8-10 years ahead of HR on this. Everything HR is about to discover, T&S has documented, settled in court, ISO-standardized, and published transparency reports about. The thought-leadership lift is to read it forward, not to invent it.

Published, versioned policy spec is the operational artifact

OpenAI's Model Spec, dated and changelogged, is the most mature public example of policy-as-code in the AI era. Meta and YouTube publish quarterly Community Standards Enforcement and Integrity Reports with prevalence, removal volume, proactive detection rate, appeals overturned per policy area. "Our values" is not enforceable. "Here is the current policy on terminations / leaves / harassment / promotion calibration, version 4.7, dated, with a changelog" is. The policy team writes it; the agents enforce it; both reference the same artifact.

T1 / T2 / T3 split is the org structure

T1 (obvious cases) goes to agents. T2 (contextual, ambiguous) is agent-pre-sorted, human-decided. T3 (worst cases, novel harms, geo-political, appeals) stays human, with wellness scaffolding. TikTok runs the cleanest version of this: hashing for known harmful content (auto-removal), ML for novel, humans for contextual or nuanced. HR's analog: T1 is policy questions, simple approvals, status checks. T2 is calibration disputes, perf disagreements, leave edge cases. T3 is mental health crises, harassment investigations, terminations involving litigation risk.

Wellness infrastructure for T3 is contractual, not optional

Meta's Scola settlement: $52M for moderator PTSD (2020). Sama Kenya: 81% of moderators classified as severe PTSD; mediation broke down 2024. Cognizant exited the content moderation business in 2020 directly because of the fallout. The institutionalized response: mandatory wellness staffing (24/7 EAP, on-site psychologists), content interventions (grayscale, blur-by-default, audio-off), tenure caps on the worst queues, post-Sama vendor contracts that specify mental health support. HR's T3 specialists handling termination-grade cases need scheduled rotation, on-staff counselor access, exposure caps, and SLAs in vendor contracts if outsourced. Build it on day one.

Compression and growth happen at the same time

Meta's proactive detection of removed hate speech went from 24% in 2017 to over 95% today. Meta's reviewer count stayed in the 15,000 range across that period. YouTube went from a few thousand to 10,000+ in 2017-18 because automation could nominate-for-review at scale but not make final calls. The pattern: T1 compresses, T2 stays roughly flat with harder cases, T3 grows. Aggregate human T&S headcount across the industry has grown, not shrunk, because volume keeps rising and edge cases multiply. HR should expect the same shape: the work does not disappear, it migrates up the stack.

Stand up calibration and QA between policy and enforcement

YouTube's leadership meets weekly with QA leads worldwide to discuss enforcement quality. Without it, policy interpretation drifts across geos, managers, and agents, and the org loses fairness without noticing. For HR, this is the team that audits whether the agent is enforcing the policy the policy team thinks it wrote. Standalone, with operating cadence, with published drift metrics.

Pre-launch and continuous bias audits, with veto authority

Meta's Responsible AI team owns dataset balance (Casual Conversations v2: 45,000+ videos balanced across age, gender, skin tone, lighting), pre-deployment fairness evaluations, and bias monitoring. OpenAI's external red-teaming methodology (arXiv 2503.16431) specifies access tier, time budget, expertise composition, reporting structure. The DTSP Safe Framework, now ISO/IEC 25389, is the international standard for T&S assessment: five commitments, 35 specific best practices, third-party assessable. The HR agentic stack will hit disparate impact issues immediately (performance scoring, promotion recommendations, layoff selection). The fairness team needs veto authority on launch and visibility into ongoing drift, owned outside the team that ships the agent.

Hire from outside HR for senior agentic-HR roles

T&S's strongest leaders came from product, intel, civil society, and platform-internal product trust roles. Yoel Roth started as a Twitter intern in 2014, mentored by Del Harvey (whose own path: lifeguard to reality TV psychological screener to Twitter T&S). Pure HR-generalist backgrounds will struggle to operate the policy-as-code stack. The Yoel Roth / Del Harvey archetype — technical, policy-fluent, builds tooling instincts, comfortable with the worst cases — is what the senior agentic-HR role requires.

A 24-month sequence

For a CHRO starting now

A pragmatic sequence that respects the dependencies: foundation roles before agent surfaces; warehouse work before strategic CoEs; CoE rebuild before HRBP rebuild; bridge-cohort triage before any reskilling investment.

Phase 0 · Months 0-3

Foundation

Hire HR Agent PM 0 (the first one; builds the playbook). Hire HR Trust & Safety lead (sets up evaluation, bias, redteam capability). Hire HR Policy Engineer 0 (starts surfacing policy debt). Run aptitude triage on the middle layer. Pick the first agent surface as a pilot (lowest political cost: leave management or routine case routing). Stand up Strategy & Planning rebuild (the brain; needs to be functional from Phase 1).

Phase 1 · Months 3-9

Highest-volume, lowest-political-cost rebuilds

Scaled HR Ops: policy-as-code for top 30% of cases. Immigration & Mobility: 95% rules-based, fastest ROI. Benefits admin: same logic. These three are the warehouse work of HR. Compression proves the model. Cross-functional muscle builds. Bridge cohort (Tier A from triage) gets first cohort training cycle. Reviewer well-being infrastructure stands up before any T3 escalation reaches a human.

Phase 2 · Months 9-15

Strategic CoE rebuild

Comp: ops automation first, then mechanism-design capability for dynamic comp. Performance Management: cycle automation first, then continuous-signal architecture. Talent Management: skill-graph plus mobility marketplace; evolves into talent intelligence. The strategic-gain layer. Politics rises because comp and perf touch every leader. L&D rebuild kicks into high gear.

Phase 3 · Months 15-24

The hardest. HRBP and Employee Engagement

HRBP rebuild: most political. Compresses to roughly 50% headcount with a 2x more senior bar. Employee Engagement: rebuild as Internal CX. These are last because (a) you need agent and system foundation in place to give HRBPs leverage, (b) compression has been proven elsewhere, (c) the bridge cohort is identified and ready.

Continuous · All phases

The standing rebuilds

L&D rebuild (continuous; scales in Phase 2). Strategy & Planning rebuild (continuous; load-bearing from Phase 0). Capability supply-chain ledger maintenance. Aptitude triage refresh (annual). T&S audits (quarterly). Policy spec versioning.

Hard dependencies

→HR Trust & Safety must be in place BEFORE you scale agent decisions in any high-stakes area. Skipping this is how bias lawsuits enter the picture.
→HR Policy Engineering must precede every CoE rebuild. You cannot build a Comp agent if Comp policy is implicit.
→Strategy & Planning rebuild precedes everything else because someone has to model the headcount-and-cost trajectories and design the org-of-the-future.
→L&D rebuild starts immediately because reskill cohort selection takes 6-12 months minimum.
→HRBP rebuild cannot precede CoE rebuild because HRBP work depends on CoE outputs.

What the function should not do

×Run a multi-year reskilling program for the whole function. Approximately 70% of investment will not return.
×Lead with HRBP rebuild. Politics will sink it before agent foundations exist.
×Hire generic "AI talent." Match the eight skill primitives to specific talent pools.
×Outsource scaled HR ops to commodity BPOs. Repeats every Trust & Safety mistake of the last decade.
×Skip HR Trust & Safety. The bias incident will land within 18 months without it.
×Spread reskill investment evenly across the middle layer. Tier A through Tier D require different investments and different conversations.

Open questions

What this paper does not yet answer

Five questions where the framework runs out of confident ground. Naming them so they are visible, and so future versions earn the right to claim coverage.

What is the right triage cadence for the aptitude assessment?

Annual is too slow if the function is moving fast. Quarterly is operationally heavy. Possibly semi-annual with continuous tooling-adoption telemetry as the real-time signal layer underneath. The pattern is open.

How does the function handle the political reality of the Tier D conversation?

The framework names the cohort that should transition out. It does not yet name the conversation, the support, the brand-protection logic of doing it well, or the boundary with severance and outplacement design. This is its own paper.

How does the works-council and union overlay change the math?

In Europe and other co-determination jurisdictions, the function cannot make middle-layer transitions unilaterally. The 24-month sequence likely extends to 36-42 months, and Phase 3 in particular gets longer. The framework's spine should hold; the sequencing changes.

What is the right hiring ratio for senior HR Trust & Safety leaders, internal vs. external?

T&S itself does not have a settled answer; HR should not pretend to have one yet. The pattern from T&S is: 60-70% external (from product, intel, civil society, ex-tech-T&S), 30-40% internal mobility from product-adjacent roles. Whether that ratio holds for HR is open.

How does the framework adapt for organizations where HR is decentralized into business-unit HR?

The CoE-and-HRBP model is a Fortune-500 abstraction; many real orgs are messier. The framework's spine (skill primitives, cognitive shape shift, reskilling limit) holds. The specific role names, the layering, and the rebuild sequence will differ. Worked examples needed.

Building the new HR org chart?

The framework is operator-grade today. If you are designing the rebuild and the framing here lands, the next step is a conversation.

Start a conversation Read the parent paper Read the C-level brief

Catalyst question: Fiona Cicconi · Chief People Officer, Alphabet. Author: Rahul Jindal. Companion to AI Adaptive People Function (Layer 5 of the Adaptive Org transformation stack).