Fieldwork worksheets
Choose a worksheet, fill in your notes and download a copy. You can also download a blank template to work offline.
Choose a problem worth testing
Use one card for one decision. Start with a real case, examine the current alternative, and mark missing evidence as unknown.
Which role needs to make what decision, and when? What would a useful result allow that person to do?
Describe one recent occurrence. Record its date and an authorised evidence reference. Separate what happened from your interpretation.
How often has this happened over a stated period, and what followed? Record the evidence, denominator where relevant, and any unverified estimate.
How is the decision handled today? What already works, what fails, and could an existing feature or simple rule address the gap?
Which inputs and local rules are essential? Note availability, permitted use, freshness and missing information; identify who can confirm correctness.
Who can produce, check and repeat the result? Estimate the full effort, including preparation, review and corrections, and identify missing expertise.
State the most consequential unverified assumption. What evidence would contradict it, and how will you check it before making a commitment?
Choose investigate, test or stop. Record the reason, the smallest useful next step, its owner and review date.
Explore a B2B buying decision
Understand past behaviour before describing an offer. Capture evidence and objections separately from your interpretation; leave an unagreed next step blank or marked none.
Which part of the work does this person know directly? Are they a user, process owner, buyer or adviser? Record the limits of their knowledge.
Ask: 'Tell me about the last time this happened. What did you do next?' Record the sequence and evidence references without proposing a solution yet.
Ask what they use today, what they have already tried and why they kept or abandoned it. Include the option of accepting the current situation.
Ask what actually happened as a result and how often it recurs. Separate documented facts, the participant's estimates and your own assumptions.
Ask when this is not a problem, what the current approach does well and why outside help might be unnecessary. Record objections in their own terms.
After understanding the case, describe one small proposed output. Ask what would make it useful, unsuitable or unnecessary. Separate the response from your interpretation.
Who can authorise a paid trial, what must they see, and what approval or procurement steps apply? Mark budget authority and funding as confirmed or unknown.
What specific action, owner and date were actually agreed? Record none if nothing was agreed; distinguish a further conversation from a paid commitment.
Compare manual work and AI fairly
Define the task, cases and acceptance criteria before seeing results. Compare the whole workflow, including preparation, checking and corrections; add a simple rule-based option when relevant.
What task and unit of work are being compared? Define usefulness, unacceptable errors and the quality/time trade-off that would justify a change before testing.
Describe the current manual method, any simple rule-based alternative and the AI-assisted method. Record operators, tools, versions, instructions and allowed support.
Define eligible cases, the selection method, period and exclusions. Include ordinary, difficult and ambiguous cases; record counts and any missing case types.
Who can establish the correct or acceptable outcome without relying on the AI answer? Record the supporting evidence and leave unresolved cases visibly unknown.
Use comparable case sets, or independent operators on the same cases. Alternate order where feasible; record training, practice, feedback and changes that could favour either method.
Record errors, unresolved items and useful outcomes before and after corrections. Give counts and denominators: for example, missed confirmed critical cases / all confirmed critical cases. A zero denominator is not perfect accuracy.
Record preparation, execution, checking and correction time for every role. Separate person-time from elapsed waiting; disclose setup work and which effort recurs.
Compare results with the predeclared criteria. Record the useful trade-off, unresolved limitations, changes needed and the next test on fresh cases; do not move the criteria after seeing the winner.
Review delivery, use and results
Review the pilot against its original agreement and a later opportunity to use the result. Keep delivery, acceptance, actual use and observed benefit as separate findings.
Record the intended decision, user role, scope, exclusions, baseline, acceptance criteria and review date. Identify any changes since agreement.
What was delivered, in which version and when? What was accepted, rejected or left unresolved, and by which accountable role?
What did the user actually do with the result? Record date, action and evidence; report used opportunities / eligible opportunities. Distinguish opening a file from taking a decision.
Could another intended user repeat the process on fresh inputs? Record assistance, rule changes and missing inputs. If no new opportunity occurred, mark repeat use as untested.
Compare the intended outcome with the baseline using consistent periods and denominators. Record other possible causes, adverse effects and uncertainty; keep satisfaction separate from performance.
Record effort by role across scoping, preparation, production, review, corrections, handoff and client work. Separate setup from recurring effort and compare it with the original estimate.
Which issues remain, how serious are they and who owns the next action? Identify maintenance, refresh and approval responsibilities; flag conclusions beyond the reviewer's expertise.
Record the decision and supporting evidence, the next owner's commitment and a review date. If unused, distinguish lack of opportunity, access, fit or need before drawing a conclusion.