Practical guide
How a future-of-work researcher can build the case for applied AI capability evidence
A future-of-work researcher is usually an evidence architect and influential reviewer, not the automatic owner of assessment procurement. Their strongest business case does not begin with a tool category. It begins with a decision that current workforce data cannot answer, then shows how a bounded study could produce better evidence. Use the FIELD map: Focal decision, Indicators, Evidence gaps,.
A future-of-work researcher is usually an evidence architect and influential reviewer, not the automatic owner of assessment procurement. Their strongest business case does not begin with a tool category. It begins with a decision that current workforce data cannot answer, then shows how a bounded study could produce better evidence. Use the FIELD map: Focal decision, Indicators, Evidence gaps, Learning design, and Decision rights.
Focal decision: name the choice that needs evidence
Replace broad questions such as “Are our people ready for AI?” with a choice and a time horizon:
> By the next planning cycle, should the organization invest in task-specific AI practice, revise workflow controls, expand tool access, or run further study for a defined population and work context?
Write the current alternatives before recommending a new assessment:
This prevents the business case from treating a preferred product as the only option.
- maintain the current program and measure tool access or attendance;
- use self-report surveys about confidence and frequency;
- review work samples or manager observations;
- run a structured simulation with process and deliverable evidence;
- combine methods in a limited study.
Indicators: separate adoption, capability, and outcome
These constructs answer different questions.
| Construct | Example indicator | What it cannot establish alone | | --- | --- | --- | | Access | licensed seats, approved tools | whether people use AI well | | Adoption | active use, task frequency, reported use | quality, judgment, or risk control | | Confidence | self-rated comfort | observed performance | | Applied capability | behavior in a defined work task | organization-wide productivity | | Operational outcome | time, quality, rework, customer effect | why the outcome changed |
The OECD's review of AI measurement in business ICT surveys shows that official measures depend on explicit definitions and survey questions, and that different measures support different policy uses. The OECD also uses paired employer and worker surveys because organizational implementation and first-person experience provide complementary evidence. These sources support construct discipline; they do not supply a ready-made company benchmark. OECD: AI measurement in ICT usage surveys OECD: AI surveys of employers and workers
Sources: [1] [2]
Evidence gaps: map what the current portfolio misses
Interview the workforce analytics lead, learning owner, AI program lead, risk or governance reviewer, business-unit representative, and employee representative where appropriate. For each current source, record:
Then build a gap map.
| Decision question | Existing evidence | Material gap | Consequence of guessing | | --- | --- | --- | --- | | Which workflow needs practice? | usage by tool | no task or error context | training targets the loudest demand | | Which controls need repair? | incident count | no near misses or verification behavior | weak controls remain invisible | | Who needs which support? | confidence survey | no observed task evidence | support is assigned on sentiment | | Is expansion sensible? | participation and anecdotes | no comparison or operational burden | scale precedes learning |
NIST's AI RMF Core recommends defining the tasks and methods an AI system will support, documenting test methods and results, considering uncertainty, and monitoring the system in its operating context. A workforce study can borrow that evidence discipline without claiming to evaluate the AI system itself. NIST AI RMF Core
- population and sampling method;
- task context and date;
- construct measured;
- whether evidence is self-report, system telemetry, manager judgment, work product, or observed process;
- missingness and likely selection effects;
- privacy, access, and retention constraints;
- decisions the source is fit to inform.
Sources: [3]
Learning design: propose the smallest useful study
A credible proposal states what will be learned, not only what will be delivered.
### Population and task
Define the work context without turning a role label into a universal ability. Select one or two realistic tasks with materials, constraints, ambiguity, and a useful deliverable. State who is included, excluded, and underrepresented.
### Comparison
Choose a comparison that matches the decision. Options include a baseline task before practice, alternate task forms, groups receiving different support, or repeated measurement after a defined intervention. If no comparison is feasible, say which causal questions will remain unanswered.
### Evidence packet
Collect only what the study requires:
Predefine the scoring rubric, reviewer training, disagreement handling, minimum evidence for interpretation, retention period, access controls, and participant communication. A polished score is not a substitute for a traceable evidence record.
### Study gates
Use four dispositions:
- task interpretation and plan;
- information and source use;
- AI instructions, iteration, and critique where policy permits;
- checks, corrections, and risk handling;
- final deliverable and assessor rationale;
- time, missing evidence, and operational burden.
- continue: evidence collection works and the next stage answers a live decision;
- repair: the task, rubric, workflow, or sample has a correctable problem;
- hold: evidence is too incomplete or sensitive for the planned interpretation;
- stop: the decision no longer matters or the study cannot be run responsibly.
Decision rights: assign owner, influencer, reviewer, and collaborator
The future-of-work researcher can own research design and synthesis while another leader owns funding or deployment. Make the distinction explicit.
| Role | Typical responsibility | Named person | | --- | --- | --- | | Decision owner | chooses whether and how to proceed | | | Research owner | construct, design, analysis, limitations | | | Workforce analytics reviewer | sampling, linkage, interpretation | | | Learning collaborator | intervention and support design | | | AI program collaborator | tool and workflow context | | | Governance reviewer | data, risk, policy, retention | | | Business sponsor | task relevance and operating access | |
If these roles disagree, preserve the disagreement and the evidence needed to resolve it. Do not convert stakeholder enthusiasm into an evidence claim.
Copyable business-case brief
Decision: What choice must be made, by whom, and by when?
- Current alternatives: What can the organization do without this study?
- Evidence gap: Which construct is missing, for which population and task?
- Proposed study: Population, task, comparison, evidence, and duration.
- Interpretation boundary: What the study can and cannot support.
- Operating requirements: People, systems, review capacity, privacy, and records work.
- Gates: Continue, repair, hold, and stop criteria.
- Outputs: Evidence packet, findings, limitations, and decision meeting.
Worked example: from adoption dashboard to workflow evidence
A research team sees that 72% of licensed employees used an approved AI assistant during the quarter. Leaders want to fund a broad prompting course and set a 90% usage target. The future-of-work researcher applies FIELD.
The focal decision becomes narrower: should the operations function prioritize prompt practice, source-verification practice, workflow redesign, or no new intervention for monthly exception reporting?
Existing telemetry shows activity but not the task performed. A survey shows high confidence among frequent users, yet there is no observed evidence about source reconciliation or correction of unsupported output. The researcher proposes a 30-person study using a synthetic exception-reporting task. Participants receive a spreadsheet, two policy excerpts with a deliberate conflict, and an inbox clarification. They may use an approved AI assistant and must submit a short recommendation with evidence.
The study records source review, AI iteration, verification actions, final accuracy, reviewer disagreement, and completion time. It does not use the result for employment decisions. The pre-registered repair gate triggers if more than 10% of sessions lose event data or if reviewers cannot reach acceptable agreement after adjudication training.
The synthetic result shows that most participants produce a readable report, but many fail to reconcile the policy conflict. Frequent tool use does not distinguish stronger verification behavior in this small sample. The researcher's recommendation is to repair the workflow and test a source-conflict practice module before proposing broad expansion. The finding is bounded to this sample, task, and period.
Objections to address directly
“We already have usage data.” Usage is relevant to adoption. Show why the decision also needs task evidence, or do not add a study.
“A simulation is artificial.” It is. Explain which work conditions are preserved, which are simplified, and how transfer to live work remains uncertain.
“One score will make reporting easier.” A score can summarize evidence, but the decision should retain pillar-level results, observed examples, missingness, reviewer confidence, and limitations.
“The study will prove the program worked.” Only a design with a credible comparison, adequate implementation evidence, and appropriate analysis can address an effect question. Otherwise report association or descriptive change.
“The researcher should own the purchase.” Ownership follows organizational authority. Classify the researcher accurately as owner, influencer, reviewer, or collaborator.
Two Prune can support a structured AI-enabled work assessment and produce evidence about observed performance in that task. It does not measure every form of AI capability, prove workforce outcomes, or make employment decisions. A future-of-work researcher should treat the product as one possible evidence source inside a broader research design.
Sources
OECD: AI measurement in ICT usage surveys
- OECD: AI surveys of employers and workers
- OECD: Empowering SMEs in the age of AI, 2026 D4SME Survey
- NIST AI Risk Management Framework Core
Sources: [1] [2] [4] [3]
Sources
- 1.OECD: AI measurement in ICT usage surveys · www.oecd.org
- 2.OECD: AI surveys of employers and workers · www.oecd.org
- 3.NIST AI RMF Core · National Institute of Standards and Technology
- 4.OECD: Empowering SMEs in the age of AI, 2026 D4SME Survey · www.oecd.org