Prioritize AI enablement by work task, not by job title

Inventory work that can actually be observed

Start with a week of ordinary task examples, stripped of sensitive material. For each, note the initiating request, input sources, output, decision owner, frequency, current pain, and what the worker must verify. Separate a task that drafts internal discussion points from one that changes a customer commitment or a financial record. A task may be common but poorly documented; another may be rare yet valuable to learn from. The board should make that distinction visible rather than assign a single “AI suitability” score to a whole function.

The ILO's 2025 update on generative AI and jobs analyzes occupational exposure using task variability as well as mean exposure. Exposure estimates describe potential interaction with technology, not the quality of a particular organization's implementation. The NIST AI RMF Map playbook calls attention to context of use, human-AI configuration, and responsibilities for oversight. Its suggestions are voluntary, not a universal ranking formula. Together they support examining the work and its context before picking a cohort.

Sources: [1] [2]

Make a board that reveals trade-offs

Suppose a synthetic services firm proposes three pilots. Task A is summarizing non-sensitive internal project updates for a manager; the source folder is versioned, the manager can correct the draft, and no action is sent automatically. Task B is drafting customer renewal commitments from contracts and account notes; a wrong promise may create a dispute, and the contract owner has little review capacity this month. Task C is identifying inconsistencies in a synthetic monthly operations pack; the original spreadsheet and definitions are available, but the team first needs a reproducible calculation check.

The board has one row per task, not one row per employee. Columns read: output and reader; source authority; sensitive-data boundary; plausible material error; independent verification route; reviewer and available time; ability to reverse; learning objective; and stop condition. Task A can be a first, bounded pilot if its owner confirms that the internal source set is approved. Task B is deferred until contracts, approval authority, and review capacity are defined. Task C can run as a controlled exercise with synthetic numbers and a check against independently calculated totals. This is a hypothetical selection, not an outcome observed at Two Prune or a recommendation for every firm.

The ranking should not be “the role with the largest number of prompts wins.” An enthusiastic early adopter may already know the workflow; a team with low tool use may have the clearest learning need. Conversely, a difficult task can be unsuitable for an unattended pilot even if it offers a tempting time-saving story. Document why a task is selected, deferred, or excluded, and who can change that decision. Do not use the board to infer that a particular worker is more capable from their title or usage alone.

Define what would count as learning

For the chosen task, write a short pilot charter. Specify the same input type, desired deliverable, human review point, observed failure modes, and an exit decision: continue, revise, or stop. Capture baseline process evidence and the pilot's corrections, not just minutes spent drafting. A manager should be able to explain whether the worker found relevant evidence, verified material claims, surfaced uncertainty, and delivered something useful for the intended reader. Where the source set, reviewer, or outcome changes between trials, record that limitation rather than reporting a clean before-and-after comparison.

The NIST AI RMF Govern playbook asks organizations to clarify responsibilities and whether personnel have the skills, training, resources, and domain knowledge needed for their assigned roles. It does not establish that training alone improves a given team's output. This is why the charter names both the opportunity to practice and the conditions for responsible use. After a small trial, revisit the board with the evidence: if the chosen task creates a growing review queue, the next investment may be review design rather than another role cohort.

Two Prune can supply a realistic, synthetic assessment setting in which a participant handles a bounded task and receives a human-interpreted evidence report. That is one possible input to a programme conversation, not a substitute for a workplace pilot, a validated productivity estimate, or an automated personnel decision. The task-selection board remains the organization's own planning and review artifact.

Sources: [3]