How to Observe Real Work Before Proposing an AI Use Case

A complaint is a useful research lead, not a ready-made AI use case. Imagine an operations team saying that client onboarding briefs take too long and asking for an AI summariser. A workflow map drawn from memory may make drafting look like the problem, while completed cases may show that difficult briefs wait on missing commercial decisions and are rewritten when contradictory fields are resolved. The complaint is useful evidence of somebody's experience, but it does not by itself identify a recurring constraint, its location or its operational consequence. Observe a bounded sample before choosing the remedy.

What must be defined before observing a workflow?

Define the workflow by its trigger, end condition, output, downstream user, participating roles and relevant case types before beginning the field round. Official use-case discovery guidance frames discovery around recurring workflows and real outputs, including their triggers, frequency, difficulty and downstream users. GOV.UK interview guidance says planning should begin with the research questions and the processes or technologies the researcher needs to understand. Keep the boundary provisional: observation may reveal an upstream dependency or a downstream acceptance step that the original process map omitted.

• Boundary: from the event that starts the work to an output accepted by a named downstream user.
• Case mix: include routine work and relevant incomplete, changed or exception paths.
• Field questions: specify what must be learned about variation, handoffs, information, decisions and consequences.
• Onboarding example: follow an approved sales handoff through to a brief accepted by the delivery lead.
• Method choice: select the smallest proportionate mix that can answer those questions and expose known blind spots.

The proposed evidence mix is an editorial synthesis of complementary methods, not a single standardised protocol validated by one source. Teams should choose the smallest proportionate mix that answers the workflow question and compensates for known blind spots. A frequent, distributed event might justify a diary and record review; a judgement-heavy handoff may need recent-case interviews and observation. The point is not to perform every method. It is to collect enough different views to challenge a plausible but incomplete account.

How do recent-case interviews reveal what actually happened?

Recent-case interviews reveal the sequence, decisions and handoffs of a completed instance when the participant is asked for a concrete example rather than an ideal process. GOV.UK interview guidance recommends focusing on stories and real examples rather than generalities or statements about how work should happen, using open and neutral follow-up questions. Reconstruct one routine onboarding brief and one difficult brief with coordinators, sales operations and delivery leads. One case illustrates what happened; contrasting cases and roles help test whether the same account travels across the workflow.

1. Ask what triggered the case and what information arrived at the start.
2. Move step by step: “What happened next?” and “What told you that?”
3. Probe decisions, uncertainty, alternatives and the information cues used.
4. Ask who received each handoff, what they returned and where the path changed.
5. Request the actual redacted item used, then identify the output that was accepted.

Do not ask the participant to design an AI feature during this reconstruction. An original field study combined semi-structured interviews with demonstrations, observation, field notes and inspection of the forms and tools participants mentioned; some activities became sufficiently detailed only when demonstrated. GOV.UK guidance also says people who perform a service together may be interviewed in pairs or small groups and that interviews can be combined with observed task performance. The ACTA paper describes cognitive task analysis as a way to identify mental demands and skills that ordinary task descriptions can omit and presents streamlined interview methods for eliciting those demands. Use that principle to probe cues and difficult judgements without claiming to run formal ACTA.

What should you watch when work happens in its normal setting?

Watch the work with its usual tools, documents, data, interruptions and dependencies, while treating the session as workflow research rather than an assessment of the person. Contextual research observes activities in the participant's everyday environment with the real equipment, data, documents, devices and usual distractions, and it can expose barriers and the workarounds used to overcome them. For the onboarding brief, note source switching, waiting, verification, clarification, rework and the points where incomplete or conflicting information requires interpretation.

• Case context and the observation mode being used
• Observed action and the relevant artefact reference
• Decision, information cue or uncertainty
• Handoff, interruption or workaround
• Visible consequence for the case
• Researcher interpretation, recorded separately from observation

Choose and record how you will observe. Silent observation preserves more natural flow but can leave motives unclear; occasional questions add context with some interruption; continuous explanation provides depth while changing the activity. Silent, occasionally questioned and continuously explained observation offer different trade-offs between natural flow, contextual understanding and researcher influence. Ask only what is needed, and distinguish what was seen from what the researcher inferred so that participants and later reviewers can challenge the reconstruction.

Participation must be informed and evidence handling proportionate. GOV.UK guidance instructs researchers to obtain informed consent, minimise their influence, reconfirm consent before additional recording and store collected personal data securely. Observation should be framed as research into the workflow, not as individual performance scoring, with proportionate collection and controlled access. Involve the appropriate privacy, security, legal, labour, accessibility and domain owners where sensitive information or workplace obligations apply. At the end, invite the participant to correct the reconstructed flow. Contextual inquiry is granular and interpretive; participant clarification and review of the researcher's understanding can create a shared reconstruction without turning the result into population-level proof.

Which evidence methods reveal what interviews or observation miss?

Artefacts, brief task diaries and operational records can extend the field view across hidden state, intermittent events and repeated cases, but each remains partial. Templates, checklists, spreadsheets, messages, drafts, paper notes and queue views may carry state, priority, memory cues and coordination practices absent from an official process map. In an original workflow study, the markings and formats of artefacts prompted detailed explanations of participants' personal organising and tracking strategies. Always ask the participant what an artefact means and how it was used in the specific case.

An applied GOV.UK project used a diary to capture feedback during or just after real guidance use and combined it with interviews and observed tasks. Brief task diaries can extend evidence across intermittent real task instances, but their entries remain self-reported and need clarification. Event sequences and timestamps can support analysis of waiting, service time, rework, duplicated cases, deviations, deadlines and recurring quality problems when usable records exist. Event logs are incomplete examples of behaviour: they may omit negative cases and unlogged work, span disconnected systems or lack a simple shared case definition. A less common path should not be discarded before its effort and consequence are examined.

When does a reported pain point become a credible bottleneck hypothesis?

A pain point becomes a credible bottleneck hypothesis when evidence places a recurring constraint at a particular point, connects it to an observable consequence and preserves plausible counterevidence. Contextual inquiry supports comparing researcher interpretations with participant understanding, while its small, subjective samples require caution about broader inference. Operational event data can reveal recurring bottlenecks, deviations, waiting and rework, but incomplete records require reconciliation with observation, artefacts and practitioner interpretation. Use the following questions as a disciplined challenge, not as a scoring model.

1. Recurrence: does the same constraint appear across more than one relevant case, role or evidence source?
2. Location: where does the queue, wait, rework loop, information gap or judgement overload enter?
3. Consequence: is the pattern connected to observable delay, repeated handling, correction, dropped work, inconsistent output, avoidable effort or risk?
4. Mechanism: how might the constraint produce that consequence, and do people doing the work confirm or correct the account?
5. Counterevidence: which cases avoid the problem, what differs and what alternative explanations remain?
6. Actionability: could changing this point alter the outcome, and is AI more suitable than a rule, ownership change, process change, training, better information or no intervention?

The six-question bottleneck test is an editorial decision rule synthesised from source-supported research practices; it does not establish causation or constitute a formal statistical test. In the onboarding example, difficult briefs wait for missing commercial decisions and require contradictory fields to be reconciled, while routine briefs are drafted quickly and longer briefs are not consistently slower. Available timestamps omit offline clarification. The evidence therefore shifts attention from slow summarisation to an upstream information problem, but it supports a hypothesis rather than causal proof.

Do not ask where AI fits until you can show where a recurring constraint appears, what consequence follows and what remains uncertain.

What belongs on an evidence-backed AI opportunity card?

An evidence-backed opportunity card should be a bounded decision record linking the proposed intervention to observed work, counterevidence, constraints and the next smallest test. The NIST AI RMF Playbook's Map function calls for documenting intended purpose, users, operational setting, business value, limitations, impacts, application scope and human oversight. The NIST Playbook asks teams to consider non-AI and non-technology alternatives that may lead to more trustworthy outcomes; this voluntary guidance does not prescribe which alternative should win.

• Workflow boundary: trigger, end condition, output, downstream user and roles
• Observed sample: period, case types, roles represented and method mix
• Stated pain and the concrete cases behind that account
• Observed pattern, consequence and references to redacted evidence
• Counterevidence, gaps, disagreements and alternative explanations
• Opportunity statement: who needs help with which bounded task and toward what outcome
• Information, quality, permission, security, privacy, labour, accessibility, domain and oversight constraints
• Alternatives considered: AI assist, deterministic rule, workflow or ownership change, training, better information or no intervention
• Next test: assumption, relevant cases, success and failure evidence, owner and review date

Link each substantive claim on the card to a field-note identifier, redacted artefact identifier, diary entry or record query; otherwise mark it as an assumption. Official discovery guidance retains an evidence requirement, owner, next milestone and later test rather than treating workshop agreement as proof of value in real work. End with one of four decisions: test AI, test a non-AI change, gather more evidence or stop. For onboarding, first test required intake fields, an explicit handoff owner and a visible exception queue. Consider drafting assistance later, only for fact-complete cases.

Propose AI only when the evidence supports a bounded task, recurring constraint, observable consequence, workable information and oversight, and a testable advantage over simpler alternatives. Gather more evidence or stop when recurrence, location, consequence or actionability remains unclear. Before observing sensitive work or handling employee, customer, confidential or regulated information, involve the appropriate organisational owners and seek qualified professional guidance for obligations or high-stakes decisions. This field method supports better discovery decisions; it is not a legal, privacy, labour or compliance determination.