Clear, source-led guidance for accountable business AI.
Map local tasks, test a fixed AI set-up, count hidden effort and redesign roles only after workload, ownership and expertise needs become visible.
ModelFold Editorial Team8 min read
Faster AI-assisted production does not automatically create usable team capacity. A service agent may draft a reply sooner while a team lead inherits more reviews, a specialist receives harder escalations and a knowledge owner must correct the material feeding the assistant. Leaders therefore need to map the work below the job title, test a fixed AI set-up and trace every hour of effort that disappears, remains, moves or is newly created before changing roles, targets or staffing.
What decision-makers should retain
Role redesign should begin with tasks because a job combines activities that AI may affect in different ways. Some may be automated within a narrow boundary, some complemented, some retained as human work and some introduced only after deployment. One label for the entire post hides these differences and can obscure where judgement, authority, hand-offs and consequences actually sit.
The ILO's refined index uses task-level occupational information to assess potential generative-AI exposure and identifies transformation as the more likely general outcome than replacement. The OECD similarly reports that workers may encounter AI through changes to tasks and working conditions. Neither finding establishes local automation, productivity, job loss or staffing capacity. Exposure is a reason to inspect observable work, not permission to make an occupation-wide workforce decision.
A current-state ledger should describe what people actually do, why each task starts, what completes it and what happens when the normal path fails. Write each task as an action and object, such as “verify invoice details” rather than “handle finance operations”. Split steps when evidence, judgement, consequences or owners differ; combine small motions only when they change together and share the same control.
Build the baseline from observation, worker and manager interviews, work samples and relevant system records; no single source substitutes for the others. O*NET can supply occupation-linked task statements, ratings, activities and work-context data as a vocabulary and completeness check. It is a US occupational database, however, and cannot establish what happens, how often it happens or what it costs in an Indian organisation. This ledger format is an adaptable editorial tool, not an O*NET method.
Test each task against one identifiable AI configuration before assigning future work. Record the model, tools, prompts, approved data, controls, test cases and date. The result then describes that set-up rather than “AI” in general. Use real, appropriately handled cases or faithful test cases, and compare outcomes with the current process on quality, effort, failure behaviour and the ability to recover.
This breadth matters because apparently similar tasks may behave differently. In a preregistered experiment with 758 consultants, AI improved measured performance on selected tasks inside the tested capability frontier but reduced correctness on a selected complex task outside it. The study describes that frontier as uneven, changing and difficult to locate in advance; its findings remain specific to the people, model, tasks and period tested. NIST's voluntary AI RMF also calls for differentiated responsibilities, operator proficiency and defined oversight.
| Task configuration | Human responsibility | AI responsibility | Required operating design |
|---|---|---|---|
| Human-performed | Performs and owns the task | Absent or limited to unrelated support | Record why evidence, suitability or benefit does not justify a different arrangement |
| AI-assisted human | Frames the task, checks required evidence and remains the decision-maker | Provides bounded retrieval, analysis or drafting support | Define allowed inputs, prohibited delegation, review criteria and completion authority |
| AI-first with human decision or review | Reviews defined cases or makes the consequential decision | Produces or routes the first output | Name the queue owner, evidence, priority, authority, peak capacity and exception route |
| Bounded automation | Monitors performance and resolves exceptions | Completes a narrowly defined task within approved conditions | Set boundaries, stop conditions, observability, change control, recovery and fallback |
Measure five separate effort accounts: production, verification, judgement, exception handling and coordination. This ledger prevents quick drafting or retrieval from being mistaken for an equivalent reduction in total workload. For every change, record the destination role or team, the volume arriving there, the effort per case and the queue under ordinary and peak demand. The five accounts are an editorial accounting device, not a standard prescribed by OECD or NIST.
Review job quality alongside time. The OECD evidence review notes that AI may reduce tedious work but may also increase work pace, reduce autonomy or narrow a worker's task set, with effects varying by worker and setting. A faster task can therefore coexist with a poorer role design. NIST's calls for clear responsibilities, communication, feedback and monitoring also support making oversight and coordination visible, without prescribing these five categories.
AI rarely removes work in one clean block; it changes where effort, authority, exceptions and learning live.
Recompose changed tasks by assigning every necessary duty to a named role with criteria, authority, communication paths and enough capacity. A generic “human in the loop” label is not an operating control: it does not say what evidence is examined, which cases trigger review, who may act, what competence is required or whether the queue can withstand peak demand. NIST calls for documented roles, differentiated responsibilities, proficiency, oversight, feedback and monitoring, but mandates no single staffing model.
Consider a hypothetical customer-support team. AI-assisted retrieval and drafting may reduce production on tested routine cases, while agents retain diagnosis, contextual verification and decisions within delegated authority. Leads and specialists may inherit review, difficult escalations, knowledge correction and coaching. Evidence should be segmented by experience: a study of 5,179 support agents found materially different productivity effects across experience groups and only suggestive evidence that the assistant disseminated practices associated with more able workers. A separate experiment with 776 product-development professionals found changes in performance and expertise integration across tested individual and team configurations, while leaving longer-term expertise questions unresolved.
Before removing routine work, identify where people currently gain observation, repetition, feedback, causal understanding and exposure to difficult cases. If AI reduces that practice, create a deliberate route through sampled original work, shadowing, simulations, case review, rotation, coaching and progressively consequential decisions. These activities are editorial recommendations, not universally tested interventions. The practical test is whether developing staff can diagnose and recover when the assistant is unavailable or wrong, not merely whether they can accept a polished suggestion.
Evidence is strong enough only after a measured pilot reconciles the complete workload and shows reasonably stable demand, quality, ownership and recovery performance. For a stated period, estimate gross production effort reduced for observed demand, then subtract added verification, judgement, exception, coordination, learning, monitoring and knowledge-maintenance effort. Adjust for changed demand, service levels, quality, rework, peak queues and work transferred downstream. The remainder is provisional capacity, not an immediate staffing conclusion.
Human touch time and elapsed wait time should remain distinct: waiting affects flow and service, but it is not automatically labour effort. The OECD's evidence on differing worker and job-quality effects means an average production measure is incomplete. NIST calls for documented responsibility, feedback, measurement and monitoring, while the uneven capability frontier supports reassessment after configuration or context changes. None of these sources prescribes this reconciliation formula; it is a conservative operational synthesis.
Do not change staffing, job descriptions, performance measures or service commitments while the workload ledger, expertise plan, review design, exception capacity or representative evidence remains unresolved. Keep employment, labour-relations, accessibility, discrimination, privacy, safety, legal, regulatory and professional determinations with qualified and authorised organisational owners. The task ledger informs those decisions; it does not make them. Used this way, a pilot becomes a decision tool rather than a shortcut from faster output to premature structural change.
Observe the local work and break the job into tasks with demand, effort, judgement, dependencies and consequences. Test a fixed AI configuration, measure all five effort accounts and then recompose duties around named owners, authority, queues and expertise needs. Make structural changes only after a measured pilot reconciles the complete workload.
It is workforce planning that begins with locally observed task demand and working conditions rather than an occupation-wide automation estimate. It examines touch time, variation, judgement, consequences, dependencies and tested AI effects before drawing conclusions about roles or capacity. Exposure data can prompt analysis but cannot supply the local answer.
Choose among human-performed, AI-assisted human, AI-first with human decision or review, and bounded automation after testing representative cases. Each choice needs explicit inputs, actions, completion authority, review or monitoring duties, exception routes, stop conditions and fallback arrangements. The appropriate boundary belongs to the tested configuration and context.
Record production, verification, judgement, exception-handling and coordination effort for each task before and during the pilot. Attribute transferred or new work to its destination owner, then measure volume, effort, queues and peak demand there. Keep elapsed wait time separate from human touch time so workflow delay is not mistaken for labour.
Only provisionally, after the redesigned process has operated long enough to reveal side work and reasonably stable demand. Reconcile gross production effort reduced with added review, judgement, exceptions, coordination, learning, monitoring, knowledge maintenance, quality, queues and downstream transfers. If evidence, ownership or expertise development remains incomplete, retain the current staffing and role design.
This article was researched using the following sources:
We report on how AI actually lands inside a business. Our work starts from named sources, separates what we found from what we think, and uses AI assistance for research and drafting under documented editorial controls. We are not a substitute for individual expert review.
A practical field method for studying real workflows, testing bottleneck hypotheses and choosing AI only when evidence beats simpler alternatives.
A practical method to map real workflow behaviour, test approvals, define accepted handoffs and redesign exceptions before automating.
A practical guide to assigning AI decision rights, choosing an operating structure, and turning pilot evidence into durable portfolio learning.
