Redesign a job around tasks before making AI staffing claims

A practical workshop for mapping where AI helps, where review grows and which responsibilities stay with people.

A job title is a poor unit for deciding what AI should do. An operations coordinator may spend one morning collecting documents, another resolving a disagreement and an afternoon checking an exception that nobody anticipated. Replacing the title with a single automation score misses the differences that determine whether a new workflow will work.

Customer-support research reports uneven benefits by worker experience. HBS shows uneven performance across consultant tasks. METR's 2025 developer trial found slower completion, while a later update warns about measurement bias. This mixed evidence favours examining actual tasks rather than predicting staffing outcomes from a general AI headline. [Support](https://arxiv.org/abs/2304.11771), [HBS](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/), [METR](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/), [follow-up](https://metr.org/blog/2026-02-24-uplift-update/).

A fictional distribution company holds a workshop with its coordinators. They list the steps in arranging a supplier visit: collect requested dates, check site access, confirm equipment needs, communicate changes and handle a missed delivery. Each step has an input, a decision, an output and someone responsible. The exercise reveals that drafting a confirmation and resolving a last-minute safety concern are different kinds of work.

The team selects one experiment: drafting a visit checklist from approved information. AI may organise the dates and equipment requirements. It may not approve site entry, override a safety rule or promise unloading equipment that has not been booked. Those boundaries are practical decisions for this fictional workplace, not claims about a particular product or a universal policy.

In the worked case, a supplier requests a Tuesday morning delivery with a large vehicle. The site record says that the unloading area is unavailable that morning. A useful assistant highlights the conflict and drafts a question about an alternative slot. It should not smooth away the conflict to produce a completed checklist. The coordinator retains authority to agree a date after checking the necessary people and records.

Map the review effort explicitly. The coordinator must check the date, access requirement and equipment availability. If the AI produces a long narrative, that review may take longer than filling the original form. The team tests a structured draft beside the existing method. It asks whether the new process reduces clerical work without increasing time spent detecting errors.

New responsibilities may appear. Someone needs to maintain the source records, review suggested exceptions and investigate incorrect drafts. These duties belong in the task map and training plan. Describing them as free human oversight hides their cost. A claimed saving in drafting is incomplete if another colleague inherits an unmeasured queue of AI corrections.

Workers should participate in interpreting results. They can identify cases where the measured task time excludes preparation done earlier, or where a suggestion merely postpones a difficult conversation. An employee who declines an unsuitable AI draft may be protecting the business. The evaluation should reward correct outcomes and good judgment, rather than penalise low acceptance or tool usage.

Try a service interruption. The coordinator needs a usable manual checklist and access to the approved records if the AI is unavailable. If the system returns after someone finishes the work, it must not replace the approved version. The team also tests a change made after a draft is generated. A timestamp and source version help reveal when a checklist needs another review.

Track complete-cycle effort, scheduling errors, repeated supplier contacts and unresolved conflicts. Keep separate results for routine and unusual visits. Averages can make the new method look successful while masking problems in the cases where coordination matters most. A small trial supports a bounded decision about that task; it does not forecast the future of the entire occupation.

Budget for licences, record maintenance, training and review, using current commercial terms checked directly before purchase. No provider price or production integration is verified here. Research findings concern particular participants and technologies. They show that the relationship between AI and work needs measurement, rather than providing a justified number of positions to remove.

The workshop can end with a modest change: AI drafts the checklist, coordinators resolve the conflicts and the company invests in better source records. That outcome is valuable if it improves the actual service. Task-level redesign gives workers and managers a concrete basis for learning together without pretending that a job title tells them what the model can reliably do.