How to develop and assess task interpretation in AI-enabled work

Reconstruct the assignment before acting

NASA's stakeholder-expectations guidance treats understanding users, intended use, needs, goals, assumptions, constraints, and decision authority as the foundation for later work. GOV.UK discovery guidance similarly recommends understanding the problem, users, context, policy intent, and constraints before committing to a solution.

For a knowledge worker, interpretation begins with the supplied record rather than an imagined ideal brief. A task may be distributed across an email, document, spreadsheet, meeting note, and deadline. Some elements will be explicit, some implied, and some genuinely unresolved. The worker should separate those states rather than letting an AI assistant fill gaps with fluent assumptions.

Interpretation is not the same as decision framing. Decision framing defines the choice that analysis should support. Task interpretation establishes what the assignment itself asks the worker to deliver, for whom, with which evidence and limits. Some tasks support a decision; others produce a reconciliation, explanation, review, or operational artifact.

Create a TRACE task reading

Use one TRACE record before drafting or analysis:

| Field | Interpretation question | Evidence to retain | | --- | --- | --- | | Trigger | What event, request, or deadline created the task? | Originating message, date, and urgency | | Recipient | Who will use the output, and what do they already know? | Named audience, role, and intended use | | Artifact | What exact deliverable, format, and level of detail are required? | Output description and acceptance conditions | | Constraints | Which sources, policies, time limits, exclusions, and approval routes apply? | Constraint with its originating source | | Evidence gaps | What is missing, conflicting, or only assumed? | Question log and chosen treatment |

For every field, mark the evidence as explicit, inferred, or unknown. An explicit instruction can still conflict with another instruction. An inference should include the clue that supports it. An unknown should lead to a clarifying question, a bounded assumption, or a limitation in the output.

After completing TRACE, write a one-sentence task contract: “For [recipient], produce [artifact] that enables [use], using [evidence], within [constraints], while making [gaps] visible.” Compare every major section of the planned output with that sentence.

Worked example: a synthetic supplier-incident review

Imagine a synthetic procurement analyst receives an email saying, “Can you review the supplier incident and give us a view by Thursday?” The evidence pack contains an incident log, contract extracts, service reports, and a finance note. A calendar invitation names the operations director, procurement lead, and security adviser. This example is invented and does not describe a Two Prune customer.

An AI assistant initially proposes a full supplier-performance report with market alternatives and a termination recommendation. The analyst pauses and creates a TRACE record:

The task contract becomes: “For the operations director, produce a concise evidence-led incident review for Thursday's governance meeting, using the approved pack, separating verified events from unresolved claims, and identifying rather than inventing missing supplier and remedy evidence.”

The analyst asks whether a quantified remedy calculation is required. While waiting, the draft includes the relevant contract fields and missing data but does not calculate a figure. Market comparison and termination advice remain outside the interpreted assignment. AI is used to organize the chronology and test whether every statement has a source, not to decide the supplier outcome.

The result is different from both a generic incident summary and a broad sourcing strategy. Its form follows the intended meeting use, evidence limits, and unresolved questions found in the task record.

• Trigger: a service interruption and a Thursday governance meeting;
• Recipient: the operations director, advised by procurement and security;
• Artifact: a short incident review that reconstructs events, identifies contract-relevant facts, records unresolved evidence, and lists decisions needed at the meeting;
• Constraints: use the supplied records, distinguish allegation from verified event, do not provide legal conclusions, and preserve restricted security details;
• Evidence gaps: two log timestamps differ, the supplier response is incomplete, and the email does not say whether financial remedies should be calculated.

Practice with split instructions

Give a learner an assignment whose requirements are divided across an inbox message, a document, and a spreadsheet tab. Include one conflict, one contextual clue, and one missing acceptance condition. Ask for a TRACE record and task contract before any substantive work.

Then reveal a late message that changes either the audience or intended use but not both. The learner should identify exactly which parts of the artifact plan change. A change in audience may alter terminology and detail. A change in use may alter analysis, evidence threshold, or structure. Neither should cause unrelated expansion.

Useful practice steps include:

• locating every instruction and tracing it to a source;
• separating explicit requirements from reasonable inference;
• identifying the intended recipient and use from context;
• turning ambiguity into a question or bounded assumption;
• checking that the proposed artifact matches the required format;
• rejecting AI-generated additions that have no task basis;
• revising the task contract when authoritative information changes.

Assess observable interpretation behavior

In an assessment, provide a realistic brief with several information channels, incomplete wording, and a tempting but unsupported interpretation. Observe whether the participant:

Do not reward agreement with a hidden ideal that the evidence never supported. Interpretation quality depends on how faithfully the participant reads the available task and handles uncertainty. A participant can interpret the task well and still execute it poorly; those are separate observations.

• finds instructions across the available evidence rather than relying on the first message;
• identifies the recipient, intended use, artifact, constraints, and deadline;
• labels explicit, inferred, and unknown elements accurately;
• notices conflicting instructions and seeks or records a resolution route;
• asks targeted questions rather than requesting the entire brief again;
• uses bounded assumptions when clarification is unavailable;
• keeps AI suggestions subordinate to the source environment;
• produces an artifact that satisfies the recovered task contract.

Verify with a reverse brief

Give the completed artifact to a reviewer who has not seen the original assignment. Ask the reviewer to infer the recipient, intended use, required artifact, evidence base, constraints, and unresolved questions. Compare that reverse brief with the TRACE record and original sources.

Next, trace each material section of the output to a task requirement or justified inference. Flag sections with no route back to the assignment and requirements with no corresponding output. Recheck conflicts and assumptions to ensure the final artifact does not silently present an unknown as an instruction.

Task interpretation is bounded by the evidence and context available in the exercise. It does not establish general comprehension or performance in another setting. In a Two Prune assessment, the TRACE record, source navigation, clarifying behavior, AI interactions, revisions, and final artifact can contribute bounded evidence about Context Framing, Evidence Navigation, AI Orchestration, and Communication & Delivery. They do not make an employment decision.

Sources

NASA Stakeholder Expectations Definition guidance

• GOV.UK guidance on how the discovery phase works

Sources: [1] [2]