Practical intelligence for accountable AI programs.
Practical framework for mapping tasks, testing AI assistance, tracking hidden effort, and redesigning roles only after workload evidence is stable enough.
ModelFold Editorial Team9 min read
Redesign roles around the tasks that actually change, then rebuild the complete role only after the new workload is visible. An AI assistant may help a service agent draft faster while sending more reviews to a team lead, more ambiguous cases to a specialist, and more knowledge corrections to an operations owner. Counting only the faster draft turns displaced work into invisible overhead and makes a fragile pilot look like permanent capacity.
The operating rules
Tasks are the useful unit of evidence because a job title bundles activities that can change in different ways. One task may remain fully human-performed, another may gain drafting support, a third may begin with an AI output but end with a human decision, and a tightly bounded fourth may be automated. The eventual decision still concerns a complete role, including its workload, authority, relationships, and development path.
Occupational exposure is a prompt to investigate, not proof of local automation or staffing capacity. The ILO's refined index uses task-level occupational information and says transformation is generally more likely than replacement, but it does not measure one employer's staffing effects. The OECD likewise reports that many workers may encounter AI through changed tasks and working conditions, with AI automating, complementing, or introducing activities within a job.
A current-state ledger should describe what people do, what starts and finishes the work, how much demand arrives, and what makes performance difficult. Write each task as an action and object, such as “classify an incoming request,” then record its trigger, inputs, output, completion condition, current owner, receiver, and exception route. Split steps when their evidence, judgment, consequences, or owners differ.
Keep touch time separate from elapsed wait time: an item sitting in a queue affects flow but is not automatically human labor. Combine interviews, direct observation, work samples, and system records because each reveals different gaps. O*NET can supply occupation-linked task statements, ratings, work activities, and work-context vocabulary, but its occupational data cannot establish what happens, how often it happens, or what it costs in a particular operation.
The task format and baseline fields are an adaptable editorial method, not an O*NET standard. Review the draft ledger with frontline workers, managers, downstream recipients, and people who handle exceptions. Consultation does not settle governance questions, but it can expose informal work, missing handoffs, seasonal peaks, and learning activity that system logs overlook. Record disagreements instead of forcing false precision into a single average.
Test each task against a fixed, recorded AI configuration before choosing its future owner. Capture the model, enabled tools, prompts, approved data, controls, test cases, and test date. This creates evidence about one identifiable setup rather than a permanent label for “AI.” NIST's AI Risk Management Framework supports differentiated human-AI responsibilities, documented operator proficiency, and defined oversight, while remaining voluntary and use-case agnostic.
Nearby tasks can produce sharply different results. In a preregistered experiment with 758 consultants, AI assistance improved measured performance on selected tasks inside the tested capability frontier but reduced correctness on a selected complex task outside it. The researchers describe that frontier as uneven, changing, and difficult to locate in advance. Those results belong to the study's people, model, tasks, design, and period, not to every workplace.
| Task configuration | Human responsibility | AI responsibility | Required operating design |
|---|---|---|---|
| Human-performed | Performs and owns the task | Absent or limited to unrelated support | Record current ownership and why the arrangement remains appropriate |
| AI-assisted human | Frames the task, inspects required evidence, and makes the decision | Provides bounded retrieval, drafting, transformation, or analysis | Define allowed inputs and actions, review criteria, completion authority, and prohibited delegation |
| AI-first with human decision or review | Reviews defined cases or makes the consequential decision | Produces or routes the initial output | Define queue ownership, evidence, review scope, priority, authority, exceptions, and peak capacity |
| Bounded automation | Monitors performance and resolves exceptions | Completes a narrow task within approved conditions | Define boundaries, stop conditions, observability, change control, exception ownership, and fallback |
For every configuration, name who has completion authority, what inputs and actions are allowed, which evidence must be checked, where exceptions go, and what stops or reverses the process. Evaluate both output quality and operational burden. A review step is not an effective control merely because a person appears in the diagram; that person also needs criteria, competence, authority, queue ownership, time, and a usable fallback.
Measure five separate effort accounts: production, verification, judgment, exception handling, and coordination. This ledger is an editorial accounting device, not a research standard. Production covers the initial work product. Verification covers checks of evidence, calculations, completeness, policy fit, and downstream usability. Judgment covers framing, interpretation, choice, authority, and responsibility. Keeping the accounts separate reveals whether work disappeared, remained, moved, or was newly created.
A faster production step can coexist with a poorer work design. The OECD evidence review notes that AI may reduce tedious work but may also increase work pace, reduce autonomy, or narrow the remaining task set, with effects varying by worker and setting. NIST calls for clear responsibilities, communication paths, feedback, and monitoring, which supports making oversight and coordination visible without prescribing these five accounts.
AI does not remove work in one clean block; it changes where effort, authority, exceptions, and learning live.
Measure transferred work where it lands, not only where the AI-assisted task began. If agents produce more responses, a team lead may receive more reviews and a specialist may receive more escalations. Track arrival patterns, service expectations, and queue age for those destinations. Otherwise, an apparent gain in one role can become a bottleneck, overtime burden, or quality risk somewhere else in the operation.
Recompose changed tasks by naming every required duty, its authority, and its capacity. Depending on the design, that may include an output reviewer, exception owner, knowledge maintainer, system steward, performance monitor, learning owner, and escalation authority. NIST calls for documented risk-management roles, differentiated human-AI responsibilities, proficiency, oversight, feedback, and monitoring, but it does not mandate a particular staffing or review model.
Protect the work through which expertise develops. Identify tasks that currently provide observation, repetition, feedback, causal understanding, supervised practice, and exposure to difficult cases. If AI removes some of that practice, create deliberate alternatives through sampled original work, shadowing, simulations, case review, rotation, coaching, and progressively consequential decisions. This learning design is an editorial recommendation, not a universally tested intervention.
Consider a hypothetical customer-support team. An assistant may reduce retrieval and first-draft effort on tested routine cases, while agents retain diagnosis, contextual verification, delegated decisions, and the customer-facing action. Team leads or specialists may inherit review, exception, knowledge-correction, and coaching demand. The example illustrates the ledger; it does not borrow a productivity percentage or staffing conclusion from another organization.
Segment pilot results by experience instead of relying only on a team average. A study of 5,179 customer-support agents found materially different measured productivity effects by experience and suggestive evidence that the assistant spread practices associated with more able workers. Separately, a field experiment involving 776 product-development professionals found changed performance and expertise integration across tested individual and team configurations while leaving longer-term expertise development unresolved. Neither study supplies a universal role design.
There is enough evidence only after a measured pilot reconciles the complete workload and shows a reasonably stable operating pattern. For a stated period, estimate gross production effort reduced for observed demand, then subtract added verification, judgment, exception, coordination, learning, monitoring, and knowledge-maintenance effort. Adjust for changed demand, service levels, quality, peak queues, rework, and work transferred to downstream owners.
Do not count elapsed waiting time automatically as labor saved, and do not call the remainder permanent capacity. Operate the redesigned process long enough to reveal side work and establish reasonably stable demand. The OECD reports that AI-related task changes can affect job quality and workers differently, while NIST calls for documented responsibilities, feedback, and ongoing monitoring. An average production measure therefore leaves important redesign questions unanswered.
Reassess the design when the tested configuration or operating context changes because capability boundaries can shift and nearby tasks can behave differently. Keep staffing, job descriptions, performance measures, and service commitments unchanged when representative evidence, workload reconciliation, ownership, review capacity, exception handling, or expertise development remains incomplete. The pilot gate is an operational synthesis, not a formula prescribed by OECD, NIST, or the cited experiments.
Use the ledger as a decision tool, not a shortcut to a workforce conclusion. Employment, labor-relations, accessibility, discrimination, privacy, safety, legal, regulatory, and professional determinations belong with qualified and authorized organizational owners. A task-redesign exercise can show where work, authority, and evidence move; it cannot make those determinations or replace the required governance and consultation.
Decompose locally observed work into tasks, baseline their demand and effort, and test a recorded AI configuration on representative cases. Measure production, verification, judgment, exception, and coordination changes before recomposing duties, authority, queues, and learning paths into a complete role.
It is workforce planning based on task demand, effort, judgment, consequences, dependencies, and tested local AI effects. It avoids treating an occupation-wide exposure estimate as proof of automation, productivity, job loss, or staffing capacity.
Choose among human-performed, AI-assisted human, AI-first with human decision or review, and bounded automation configurations. Each choice needs defined authority, allowed actions, evidence, review or monitoring duties, exception routes, stop conditions, and fallback arrangements.
Record production, verification, judgment, exception-handling, and coordination effort by task for a stated period. When work moves, attribute its demand, queues, and peak load to the destination owner rather than leaving it as hidden overhead.
Only provisionally, after a measured pilot reconciles the complete workload, changed demand, quality, queues, transferred work, monitoring, and expertise-development needs. If representative evidence or stable ownership remains incomplete, do not change staffing, job expectations, performance measures, or service commitments.
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-research method for verifying workflow pain, locating recurring constraints, and deciding whether AI or a simpler change merits a test.
Practical guide to mapping real workflows, redesigning exceptions, testing approvals, defining handoffs, and deciding when automation is ready to deploy.
Build an operating model by assigning decision rights, evidence, escalation, and review across standards, funding, delivery, risk, operations, and reuse.
