TLDR
- Observe the whole task population before changing a title, grade, pay band or headcount plan; AI removes, enriches, creates and shifts work at the same time.
- Describe progression through six visible contribution dimensions, never one employee score or a count of prompts, licences, certificates or drafts.
- Replace lost entry work with supervised consequential practice, and redesign concentrated expert exception queues before the talent pipeline hollows out.
- Build evidence-bearing bridges among practitioner, product, technical, assurance and capability paths, with protected time, receiving managers and fair appeal.
- When demand genuinely declines, make notice, consultation, learning, income treatment and fallback explicit before a new career standard affects people.
01 · The apparently successful programmeThe career ladder breaks before the organisation notices
A customer-service operation deploys an assistant that drafts routine responses. Handle time falls. The transformation dashboard is green. Yet junior advisers no longer learn policy by constructing ordinary cases, experienced advisers inherit a queue dominated by angry customers and ambiguous remedies, and the people improving the knowledge base do so between calls. The old career framework still rewards volume, so the work of judgement, recovery and system improvement is simultaneously more important and less visible.
This is not an unusual training problem. It is a career-architecture failure. AI changes the bundle of tasks inside a role before the job title, grade or organisation chart catches up. Some steps disappear; some become verification; some shift to a customer, reviewer or exception team; new work appears in evaluation, data stewardship, explanation, incident response and model change. Counting what remains by volume confuses the residue of automation with the value of human contribution.
The central proposition is simple: career progression in human-agent work should follow expanding responsibility for decisions, consequences, evidence, systems, boundaries and other people's capability. Tool fluency matters, but it is enabling capability rather than a career level. A person may produce fewer artefacts while handling a wider consequence span. Another may use an assistant constantly while leaving no reusable improvement, reliable evidence or accountable decision. Their usage curves say almost nothing about progression.
Thought experiment: the perfect output machine
Imagine an assistant that produces every standard report perfectly. Analyst output volume becomes infinite and therefore useless as a differentiator. What remains scarce? Framing the right question, recognising when the standard report is inappropriate, tracing evidence, managing consequence, changing the reporting system and teaching others to use it well. A career framework anchored to those scarce responsibilities survives the thought experiment. A framework anchored to production volume does not.
Task-level exposure is not a forecast that whole jobs vanish. Current ILO work emphasises that occupations contain varied tasks and that transformation is more likely than automatic replacement in many settings. OECD evidence also shows that skills demanded in AI-exposed work extend beyond specialist AI knowledge. The career decision therefore follows observation of changed work, not a vendor claim about capability or a projection of technical exposure.
02 · Six dimensions, never one scoreThe Human-Agent Career Architecture
The Human-Agent Career Architecture describes contribution along six dimensions. They are deliberately irreducible. A composite would allow strong throughput to cancel weak evidence, or visible coaching to obscure excessive authority. Each role level states the pattern it requires, the contexts in which that pattern must be demonstrated and the conditions under which evidence expires.
- Decision judgementFrames problems, tests uncertainty, handles exceptions, weighs trade-offs, abstains and makes an accountable choice.
- Consequence spanHandles larger populations, rights, safety, money, reputation or irreversibility with proportionate authority and remedy.
- System leverageImproves reusable workflows, evaluation, data, controls, interfaces or platforms rather than only personal output.
- Evidence responsibilityProduces, tests, interprets, preserves and exposes evidence that supports decision, challenge and recovery.
- Boundary leadershipTranslates among domain, product, engineering, operations, risk, vendors, workers and affected groups without hiding disagreement.
- Capability multiplicationCoaches, codifies practice, repairs apprenticeship, develops communities and enables others to perform safely.
| Dimension | Weak proxy to reject | Evidence of expanding contribution | Boundary |
|---|---|---|---|
| Decision judgement | Number of AI-assisted decisions | Representative decisions, abstentions, uncertainty and reviewed exceptions | Authority remains bounded by role and consequence |
| Consequence span | Seniority or customer volume alone | Population, reversibility, remedy and accountable ownership | Exposure without authority is not progression |
| System leverage | Personal speed | Reusable improvements with verified service or control effects | A polished local workaround is not yet a system change |
| Evidence responsibility | Certificate completion | Traceable claims, counterevidence, versions and recovery records | Evidence must travel only within its tested context |
| Boundary leadership | Meeting attendance | Resolved interfaces, explicit conflicts and accepted handoffs | Independent challenge cannot approve its own design |
| Capability multiplication | Popularity or content volume | Observed transfer, capable successors and reduced dependency | Coaching cannot become unpaid invisible labour |
The architecture does not say that every dimension must rise at the same rate. An expert practitioner may show exceptional judgement and consequence with moderate platform leverage. A product shaper may demonstrate high leverage and boundary leadership while retaining lower authority over regulated decisions. The pattern matters. The role design must also supply the authority, access, time and support needed to exercise the contribution it claims to value.
A level is a compact social contract: the decisions, evidence, consequences, reusable change and capability transfer expected. Grades and pay bands should be reconciled through local employment and reward arrangements, not inferred automatically from the six dimensions.
03 · Observe the bundle before the titleThe Task-to-Career Translation Map
Career redesign starts with a recurring population of work. A demonstration proves that a tool can perform a sample. It does not reveal who verifies the output at scale, which exceptions accumulate, whether customer explanation grows, how error recovery changes or whether demand for the service persists. The translation map forces those hidden movements into view:
Current task bundle → AI contribution → removed, enriched, created or shifted work → new decision and evidence load → authority and control change → capability requirement → role level or transition → grade, pay and support consequence.
| Task state | Question | Career implication | Typical hidden work |
|---|---|---|---|
| Removed | Has the work truly disappeared across the population? | May reduce demand, but only after shifted work and durability are tested | Sampling, correction, customer self-service burden |
| Enriched | Has routine production become judgement or exception work? | May widen contribution and justify retaining or increasing level | Uncertainty, recourse, emotional labour, policy interpretation |
| Created | What new work sustains the human-agent system? | May establish product, evaluation, data, assurance or practice paths | Evaluation, monitoring, model change, incident learning |
| Shifted | Who now performs the task and carries its consequence? | Requires workload, authority and fairness reconciliation | Reviewer queues, customer effort, manager oversight, vendor dependence |
If a claims assistant drafts routine explanations, advisers may inherit contested interpretations, distress and remedy exposure: lower volume, but greater contribution. An engineer accepting thousands of generated suggestions gains no wider contribution if defects, review burden and system quality do not improve. Evaluate changed work, not visible AI use.
A role may decline after durable redesign or a stable fall in demand. The map permits that conclusion, but not from a product demonstration, short pilot or unreconciled time estimate.
04 · Paths, not fashionable titlesSix Career Contribution Modes
Human-agent work needs several routes for contribution. Turning every ambitious employee into a data scientist would discard domain craft and create artificial scarcity. The six modes below can appear at several grades and may be combined when duties are compatible. They describe how value is contributed, not what title happens to be fashionable this year.
- AI-enabled practitionerDelivers a domain service with bounded AI, verifies evidence, decides or escalates and owns the professional standard.
- Expert judgement practitionerHandles ambiguity, rare cases, recourse and policy interpretation while repairing the routine path.
- Product and workflow shaperRedesigns requirements, work, adoption, evaluation and outcome capture across a service.
- Technical integration specialistBuilds data, context, model, tool, platform, security, reliability and production capability.
- Assurance and challenge specialistIndependently evaluates performance, risk, fairness, privacy, security, control and evidence.
- Capability and practice leaderDevelops people, transfers knowledge, changes standards and builds organisational capacity.
Automation can concentrate experts' queues with edge cases, distress and recourse. Treating them as volume producers leaves the hardest work without grade recognition, recovery time or authority to repair its source: both a retention risk and a control failure.
Small organisations can combine compatible practitioner, product and capability hats, using external independent assurance for high-consequence decisions. Explicit evidence and conflict separation matter, not organisational bulk.
05 · Demonstrate contribution without surveillanceThe Career Evidence Portfolio
A career decision needs evidence broad enough to represent the work and narrow enough to respect privacy. The portfolio assembles selected work samples, decisions, outcomes, failures, recovery, system improvements, coaching and independent observation. It records the model, tool, policy and context versions that made the evidence true. It does not become a continuous behavioural exhaust pipe.
What belongs
- Representative work samples and populations.
- Decisions, evidence and consequence handled.
- Verified outcomes and boundary conditions.
- Abstentions, escalations, incidents and recovery.
- Workflow, control, product or platform improvements.
- Knowledge transfer and observed coaching.
- Affected-party, accessibility and worker considerations.
- Independent challenge, versions and revalidation.
What does not prove level
- Prompt count or assistant usage.
- Licence activation or tool evangelism.
- Self-reported hours saved.
- A polished output without traceable evidence.
- Course or certificate completion alone.
- Popularity in a community channel.
- Stretch work performed on top of full load.
- An opaque algorithmic employee ranking.
Negative control: the prolific assistant user
A software engineer accepts generated code throughout the quarter and produces more pull requests than peers. Review time, escaped defects and operational burden do not improve; no reusable test, interface or platform capability appears. Usage is real, but the proposed promotion mechanism predicts advancement even when durable contribution is absent. That failed negative control invalidates usage as career evidence. The engineer may still deserve support or recognition, but the promotion claim needs a stronger portfolio.
Thought experiment: delete every digital trace
Suppose privacy rules remove prompt histories, keystroke data and activity dashboards tomorrow. Could the organisation still assess a person's career contribution from selected work, decisions, consequences, system changes, peer transfer and independent review? If not, the framework has mistaken surveillance convenience for evidence quality. A defensible architecture should become fairer, not blind, when behavioural exhaust is unavailable.
Psychological safety matters here because a credible portfolio includes abstention, uncertainty and failure. If promotion requires a flawless record, people learn to hide the very evidence that distinguishes mature judgement. Managers must separate good-faith learning and recovery from recklessness, concealment or repeated disregard of controls. Appeal is part of evidence quality because assessors and work samples can be wrong.
06 · Mobility requires a receiving systemThe Career Bridge Matrix
A list of adjacent roles is not a mobility programme. Each bridge needs transferable capability, missing capability, real work, a mentor, a sponsor, a receiving manager, protected learning time, grade and pay treatment, an assessment gate, an appeal route, a maximum duration and a fallback. Without these, retraining produces certificates for vacancies that do not exist.
| Bridge | Transferable capability | Missing capability to prove | Safe real-work assignment | Gate and protection |
|---|---|---|---|---|
| Practitioner → product | Domain judgement, user consequence | Requirements, outcome evidence, lifecycle ownership | Co-own one bounded workflow change | Receiving product manager, protected time, review by users and engineering |
| Practitioner → technical | Failure knowledge, data context | Engineering quality, security, reliability | Build a tested evaluation or integration component | Technical mentor, code and production gate, fallback to enriched practice |
| Product → assurance | Lifecycle and evidence knowledge | Independent challenge and conflict handling | Review a different product outside prior approval | Organisational separation, assurance sponsor, no self-review |
| Expert → capability leader | Tacit judgement, recovery | Transfer design and inclusive facilitation | Develop successors through observed cases | Recipient demonstration, attribution, workload relief and reward |
| Declining role → new path | Service, customer and domain knowledge | Path-specific capability | Paid bridge assignment tied to real demand | Income treatment, appeal, deadline and credible fallback |
Training transfer depends on learning design, worker conditions and the work environment. A worker cannot prove product judgement without access to a product decision, or independent assurance while accountable for shipping the challenged design.
Stretch work must displace ordinary workload; otherwise spare time, schedule flexibility and manager support become hidden access tests. The bridge therefore makes time, location, accessibility and grade treatment explicit for part-time staff, carers, disabled workers, contractors and distributed employees.
07 · Repair both ends of the pipelineThe Entry and Expert Continuity Test
Automation can hollow a career ladder from both directions. At the entry end, routine preparation disappears even though it once taught context, error patterns and customer consequence. At the expert end, remaining work concentrates into rare, ambiguous and emotionally demanding exceptions. The middle may appear efficient while the organisation quietly loses its way of producing future experts and exhausts the experts it already has.
| Continuity question | Failure signal | Repair evidence | Owner |
|---|---|---|---|
| Which entry tasks taught context and error recognition? | Juniors only approve outputs they cannot explain | Supervised reconstruction, evidence reconciliation and customer cases | Business and learning leaders |
| Is real work available without unsafe authority? | Only simulations or passive observation | Bounded cases with review, stop and remedy | Line manager and domain assessor |
| Has expert work concentrated? | Queue contains mostly conflict, distress and irreversible decisions | Capacity, recovery time, grade and system-change authority | Operations executive and HR |
| Is expertise transferred fairly? | Knowledge extracted without consent or attribution | Protected time, agreed reuse, reward and capable successors | Practice sponsor and worker representatives |
| Can the pipeline replace experts? | No credible first step or route through varied cases | Forecast demand, bridge cohorts and demonstrated succession | Workforce planning owner |
Spreadsheet preparation once taught juniors data origins, reconciliation failures and business context. Checking an assistant's totals does not recreate that learning. Supervised reconciliation across varied cases, explanation of discrepancies and gradually wider authority can.
Experts need parallel redesign. A permanent exception queue concentrates judgement and consequence despite fewer cases. Workload, emotional demand, recovery, authority and power to repair upstream systems need explicit treatment; adding coaching alone consumes the capacity continuity requires.
08 · Protection before consequenceThe Transition and Income Covenant
Not every role will remain unchanged. Some task bundles and some demand genuinely decline. Fair architecture does not promise otherwise. It makes the evidence, choices, distribution and support of transition explicit before grade, pay or employment consequences occur.
The covenant covers advance notice and transparent task evidence; worker consultation and representation; role, location, schedule and accessibility options; funded learning and workload relief; real bridge assignments; grade, pay and benefits treatment; fair selection and appeal; redeployment, voluntary exit or redundancy routes where applicable; review of disparate effects; and named executive, human-resources and business owners.
Opaque analytics cannot decide careers. Behavioural data may be incomplete or biased by schedule, disability, caring, location and access to visible work; a manager's signature does not cure that. Use inspectable task evidence, representative work, independent challenge and a route to contest facts and standards.
The Role Half-Life and Revalidation Rule
A career standard expires when the underlying work changes materially. Trigger review when the task bundle, AI capability, regulation, workflow, demand or evidence standard crosses a declared threshold. Role descriptions follow observed work, not product roadmaps. Existing incumbents receive a supported transition before a new standard is used in a consequential assessment. This protects people from moving goalposts while keeping the framework responsive.
Decision rule
Do not lower grade or pay because automation reduced routine steps. First verify durable task and demand change, reconcile hidden and shifted work, assess all six contribution dimensions, run entry and expert continuity, consult affected workers and apply the covenant. Only then can local employment and reward processes determine a fair consequence.
09 · Ten worked casesWhat the architecture changes in real decisions
1. Fewer drafts, harder customer remedies
A customer-service adviser produces fewer routine replies but handles disputed policy, distressed customers and remedy. The translation map shows removed drafting, enriched judgement and created evidence work. Decision judgement and consequence rise; system leverage begins when the adviser improves the remedy route. The practitioner and expert modes combine. The portfolio includes representative exceptions, appropriate escalation and customer outcomes. Entry continuity is repaired separately. Grade is retained while workload and remedy authority are reviewed. The service director and people partner own the decision, with worker appeal.
2. Junior analysts lose spreadsheet apprenticeship
Preparation work that taught data lineage disappears. Checking an AI-produced total would expose juniors to consequence without the knowledge to challenge it. The career decision pauses. A supervised evidence-reconciliation path supplies varied source systems, discrepancy explanation, abstention and reviewer feedback. The bridge has protected time and a named finance manager. Portfolio evidence begins with reconstruction rather than output volume. Entry continuity, accessibility and cohort effects are reviewed before staffing changes. Finance operations and learning leaders own the repair.
3. The underwriter becomes an exception queue
A senior underwriter receives fewer but harder cases involving ambiguity, recourse and emotional strain. The role shows greater judgement and consequence, yet poor system leverage because there is no authority to repair the routine path. The design moves to a higher-leverage expert-practitioner route: protected recovery time, authority over exception taxonomy, access to product owners and recognised coaching. The portfolio covers decisions, reversals, recovery and upstream improvement. Grade cannot fall with case volume. Underwriting and HR jointly own load and pay review.
4. Code-assistant activity without durable improvement
An engineer uses assistants heavily and opens more pull requests, but quality, review burden and reliability do not improve. Usage and speed fail the negative control. The portfolio lacks representative outcomes, reusable system leverage and recovery evidence. The result is not a denial of potential; it is a pause in career inference. The team redesigns work and assessment around verified code quality, test assets, operational learning and platform contribution. Engineering leadership owns the standard, and the employee may challenge case selection.
5. A domain expert crosses into product shaping
A claims specialist understands user consequence and exception patterns but has not owned requirements or lifecycle outcomes. A six-month bridge assigns one bounded workflow, workload relief, a product mentor, sponsor and receiving manager. Evidence gates cover user research, decision logs, evaluation, service effects and engineering handoffs. Grade and pay remain protected during transition; fallback returns to an enriched expert route. Product leadership and claims operations share ownership. The bridge yields a specialism only after demonstrated product contribution.
6. A product manager seeks assurance progression
The manager has strong lifecycle evidence but still owns the design they propose to assess. Boundary leadership cannot cancel the independence conflict. The bridge requires work on a different product, a separate reporting line, assurance methods and evidence of challenge that survives commercial pressure. Previous product knowledge transfers, approval authority does not. Career evidence excludes self-review. The chief risk officer receives the path, the product executive releases the individual from delivery decisions and an appeal route tests whether separation is genuine.
7. A frontline practitioner multiplies capability
A practitioner builds a reusable evaluation set from recurring failures and coaches peers to use it. Personal throughput falls, but system leverage, evidence responsibility and capability multiplication increase. Independent quality review confirms the asset works outside the source team. The portfolio includes exclusions, versions, peer demonstration and the product change it informed. The role moves to a higher-leverage practitioner path with protected coaching time. Operations and quality owners recognise contribution without counting community popularity.
8. A role genuinely declines
After a sustained demand change and a representative workload study, a manual classification role has materially less work. Hidden review and exception tasks are measured and insufficient to sustain the old population. The covenant activates: notice, worker consultation, paid learning, real bridge assignments, income treatment, fair assessment, appeal and a credible fallback including redeployment or lawful exit routes. An algorithm does not select individuals. The business executive and HR lead own the decision, with employee representation and disparate-impact review.
9. An opaque talent system ranks redeployment
A vendor model scores behavioural traces, message patterns and tool activity, then ranks workers for new roles. It cannot show representative task contribution, accessibility effects or stable standards. The output is rejected as the sole decision basis. The organisation redesigns selection around declared role evidence, sampled work, independent observations and appeal. Worker data use receives privacy and representation review. HR owns the process, the business owns demand evidence and assurance tests the instrument before any career consequence.
10. A small organisation uses explicit hats
A forty-person firm cannot maintain six specialist departments. It defines three broad paths: domain practice, product-technical shaping and capability leadership. People may wear compatible hats, but each consequential decision names the active hat, authority and evidence. External assurance reviews high-consequence work. Bridge assignments and succession are still explicit. The architecture enriches current roles and retains grade rather than inventing titles. The managing director and people lead own coherence; staff can challenge conflicts or invisible extra duties.
10 · Executable instrumentHuman-Agent Career Path Designer
The designer converts task change into a bounded career route. Use it with representative work evidence, worker participation and local employment expertise. The ten named scenarios exercise all five decisions. A custom result is a structured hypothesis for review, not an employee score, legal conclusion or automated promotion.
Custom career case
Worker and role
Task population
AI task change
Core contribution
Boundary and modes
Career evidence
Target path
Bridge support
Pipeline continuity
Income and fallback
Fair transition
Revalidation
The result preserves six separate dimensions and names missing evidence. It cannot establish legal compliance, individual suitability or organisational demand without local review.
Enrich the current role and retain its grade
Task-to-Career Translation Map: Routine work changes, while judgement, evidence and hidden work remain inside the role.
Contribution-dimension evidence: Preserve six separate judgements and document context instead of producing a composite score.
Mode and level: AI-enabled practitioner with a locally evidenced contribution pattern.
Career Evidence Portfolio: Representative cases, decisions, outcomes, abstentions, recovery and system changes.
Bridge plan: No forced path change; attach bounded development to real work.
Entry and expert continuity: Protect entry practice and monitor exception concentration.
Grade and pay consequence: Retain grade while task and demand evidence mature.
Transition protection: Consultation, accessible evidence, stable standards and appeal remain active.
Revalidation trigger: Re-open on material model, workflow, policy or demand change.
Appeal and owners: Business owner and people partner decide; the worker can inspect and challenge evidence.
11 · Strong rival and failure boundaryKeep career architecture lighter than the work it clarifies
The strongest rival: frameworks freeze fast-changing work
Detailed levels can become bureaucracy: title inflation, credentialism, grade disputes and performative portfolios. AI changes faster than job families, small organisations need generalists, and employment rules vary. False precision can consume manager time without improving mobility.
The answer is a few durable contribution dimensions and modes, evidence from real work and frequent re-observation of task bundles, not product-specific lists or new titles. Compatible hats, sampled portfolios and role half-life keep the structure proportionate and current.
Where this architecture fails
Do not use it without observable representative work, defined authority, independent challenge, worker consultation, a real bridge destination and stable standards, or where consequences are predetermined. Stop if it becomes automated ranking, extracts expertise without consent, makes stretch work unpaid or reduces grade solely because routine tasks were automated.
The designer is an organisational synthesis, not a validated psychometric instrument. It cannot estimate potential, predict redundancy, replace collective bargaining or determine lawful treatment. Field studies show bounded effects in particular populations and technologies; task economics describes aggregate change, not an individual's fair grade.
Start with one population and decision, with worker participation. Compare the map with observed work and pilot without employment consequence. Test explainability and access; give incumbents a supported path before consequential use.
12 · The durable ladderProgress should mean carrying better responsibility
AI shifts value from routine volume towards concentrated judgement, verification, recovery and stewardship. Careers that still reward volume will downgrade enriched roles, overload experts and erode apprenticeship.
The alternative begins with observed work and ends with a fair, evidence-bearing decision. The dimensions, translation map, modes, sampled portfolio, destination bridge, continuity tests and covenant connect changed tasks to protected progression.
A human-agent career ladder is worth climbing when each step means better judgement, wider but bounded consequence, stronger evidence, more reusable systems, clearer boundaries and greater capability in others. That remains durable contribution as machines perform more production.
Evidence and limitsSources checked 3 September 2026
ILO, OECD and UK sources frame tasks, labour, skills and worker participation. Field experiments and research on task economics, competency, training transfer, human-AI interaction and team learning provide bounded mechanisms, not a universal career score.
- International Labour Organization. Generative AI and Jobs: A Refined Global Index of Occupational Exposure (2025).
- International Labour Organization. Artificial Intelligence Adoption and Its Impact on Jobs (2025).
- International Labour Organization. Social Dialogue Critical to Ensure Positive Outcomes of AI in the Workplace (2026).
- OECD. The Impact of AI on the Workplace: Main Findings from Surveys of Employers and Workers (2023).
- OECD. Artificial Intelligence and the Changing Demand for Skills in the Labour Market (2024).
- OECD. AI and Skills (2026).
- UK Government. AI Skills Tools Package (2026).
- Skills England. AI Foundation Skills for Work (2026).
- UK Government. Government Digital and Data Profession Capability Framework.
- Brynjolfsson, E., Li, D. and Raymond, L. Generative AI at Work (NBER Working Paper 31161).
- Dell'Acqua, F. et al. The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise (NBER Working Paper 33641).
- Dell'Acqua, F. et al. Navigating the Jagged Technological Frontier (Organization Science, 2026).
- Autor, D. Why Are There Still So Many Jobs? The History and Future of Workplace Automation (2015).
- Acemoglu, D. and Restrepo, P. Automation and New Tasks: How Technology Displaces and Reinstates Labor (2019).
- Campion, M. et al. Doing Competencies Well: Best Practices in Competency Modeling (2011).
- Baldwin, T. and Ford, J. Transfer of Training: A Review and Directions for Future Research (1988).
- NIST AI Resource Center. AI Risk Management and Human-AI Interaction.
- Edmondson, A. Psychological Safety and Learning Behavior in Work Teams (1999).
The Human-Agent Career Architecture, six contribution dimensions, Task-to-Career Translation Map, Career Contribution Modes, Career Evidence Portfolio, Career Bridge Matrix, Entry and Expert Continuity Test, Transition and Income Covenant, Role Half-Life and Revalidation Rule and Human-Agent Career Path Designer are this article's synthesis. They require local validation, employment-law review and worker participation.
About this paperCareers should follow durable human contribution
This paper turns changing task bundles into visible decisions about progression, evidence, bridges, continuity, grade, pay, transition and revalidation while keeping human accountability and worker appeal explicit.