Questions for designing AI-enabled work

What work is actually changing?

Describe the current task from trigger to consequence. What input arrives, what decision must be supported, what artifact leaves the workflow, and who relies on it? Split research, drafting, calculation, classification, recommendation, approval, and communication rather than calling the whole activity “content” or “analysis.”

Then ask what the AI will do: retrieve, transform, generate, compare, recommend, or act. State what it will not do. The NIST AI RMF Core calls for defining business context, specific supported tasks, system scope, human oversight, and affected actors. The framework is voluntary and does not prescribe one workplace design.

Sources: [1]

Which evidence and judgment remain authoritative?

List source systems and documents in order of authority. Which conflicts can be resolved mechanically, and which need domain judgment? Where can context disappear during retrieval, summarization, or transfer? Name the person who decides when evidence remains incomplete.

Ask what a reviewer must see to disagree with the output. A human approval step is weak if the source is hidden, time is inadequate, or rejection has no effect. Define the stop condition and preserve unresolved claims rather than allowing fluent text to erase uncertainty.

How will the design affect people doing the work?

Examine workload before and after the change. Which activities become faster, and which checking, coordination, or exception tasks appear downstream? Who handles failures? Does monitoring alter autonomy or create incentives to maximize visible activity? Can people report workarounds or harmful effects without penalty?

The International Labour Organization has emphasized social dialogue throughout AI design and deployment. That statement is a policy position, not evidence that consultation alone produces a good design. Use worker and manager input to surface operating facts, then test the resulting design.

Sources: [2]

Worked artifact: a synthetic executive-briefing workflow

Imagine a synthetic strategy team that prepares weekly executive briefings. The proposed assistant retrieves internal reports, drafts a summary, and suggests three decisions. The canvas separates evidence collection, synthesis, decision framing, and final release.

For evidence collection, the assistant may retrieve approved documents but must display version and date. For synthesis, it may draft claims linked to passages. A strategy analyst resolves conflicting forecasts and records why one source carries more weight. The director owns the recommendation and external circulation. No automatic sending is allowed.

The first trial reveals that drafting time falls but analysts spend longer checking duplicated and stale citations. A regional report is consistently omitted because its file metadata differs. The team changes the retrieval rule, adds a completeness check against the source register, and narrows the assistant’s decision suggestions to questions for the analyst.

The canvas records the before-and-after task steps, missing-source test, reviewer access, observed checking time, two escalations, and one unresolved forecast. It does not declare productivity improvement because total effort and decision quality are not yet established. The scenario is synthetic and does not demonstrate Two Prune outcomes.

What would make the workflow fail safely?

List material failure modes: missing authority, fabricated citation, incorrect calculation, prohibited data, inaccessible interface, excessive review load, or action beyond permission. Assign a detection method, owner, disposition, and stop route to each. Test degraded cases, not only clean demonstrations.

Ask what happens when the model, prompt, retrieval index, source system, user group, or deadline changes. A design that worked in a small trial may fail under real volume. Set review dates and thresholds for reopening the workflow.

How will learning and performance be interpreted?

Define what people need to understand and practice. Tool instructions are insufficient when the task requires evidence navigation, verification, judgment, and communication. Use representative cases that create observable opportunities to demonstrate those behaviors.

Separate absence of opportunity from weak performance. Do not infer capability from prompt length, tool activity, or a polished final answer alone. Record missing evidence and accessibility constraints. Keep assessment results task-bounded and subject to human interpretation.

Can someone reconstruct the final decision?

Before release, trace the output to sources, transformations, checks, exceptions, and accountable approval. Confirm that a second person can understand what changed and why. If the workflow depends on undocumented local knowledge, capture or deliberately preserve that human step rather than pretending it has been automated.

These questions help teams redesign work as a sociotechnical process rather than bolt AI onto a task. They do not establish safety, demonstrate compliance, or prove benefit. Two Prune can provide structured evidence from controlled work simulations; it does not design customer jobs or make employment decisions.