Practical intelligence for accountable AI programmes.
A practical field-research method for testing workflow complaints, locating recurring constraints and deciding whether AI deserves a controlled trial.
ModelFold Editorial Team8 min read
A complaint is a research lead, not an AI use case. When an operations team says client onboarding briefs take too long, an AI summariser may sound like the obvious response. Yet a workflow map drawn from memory can hide the real constraint: difficult briefs may be waiting for missing commercial decisions or being rewritten after contradictory fields are resolved. Observe bounded cases before selecting a capability, or the project may automate the visible writing step while leaving the recurring delay untouched.
Key takeaways
Define the workflow by its trigger, end condition, output, downstream user, participating roles and relevant case types before observing it. This boundary makes the field round manageable and gives researchers specific questions to answer. Treat it as provisional: an observation may reveal that an upstream decision or downstream acceptance step belongs inside the study. Use-case discovery guidance supports beginning with recurring work, real outputs and their dependencies, while interview guidance recommends planning around the research questions and processes to be understood.
For the onboarding example, set the boundary from an approved sales handover to a brief accepted by a delivery lead. Include routine, incomplete and changed-scope cases, then record what arrives, what the brief must contain, who contributes and where handovers occur. The method here is a practical synthesis, not a standardised protocol: choose the smallest proportionate mix of interviews, observation, artefacts, diaries and records that can answer the workflow question.
Recent-case interviews reveal the sequence, decisions and handovers of a specific completed case. Ask for one recent example rather than a description of how the process usually works. Begin with the trigger, follow what happened next, and finish with the accepted output. Open, neutral prompts preserve the participant's account while making it concrete. Interviewing people across connected roles can expose where their understandings of the same handover diverge.
In the onboarding study, reconstruct one routine brief and one difficult brief with coordinators, sales operations and delivery leads. Ask to see the intake material, questions, draft trail and acceptance feedback. Demonstrations and artefacts can elicit detail that verbal descriptions omit. Probe cues, information needs and difficult judgements as compact cognitive-demand inquiry, but do not describe these prompts as the formal ACTA protocol. Each case is illustrative until compared with other cases and evidence.
Watch how the work unfolds with its usual equipment, documents, data, interruptions and dependencies. Note source switching, waiting, verification, rework, handovers, workarounds and moments when someone interprets incomplete or conflicting information. Contextual observation can expose barriers and support activity that are hard to recall away from the task. It remains an interpretive sample, however, and cannot establish an objective account of an entire workforce.
Choose an observation mode and write it into the research record. Silent observation preserves more natural flow but can leave motives unclear. Occasional questions provide context with some interruption, while continuous explanation produces depth but changes the activity more substantially. Separate what happened from what the researcher thinks it means, so participants and reviewers can challenge the inference rather than inheriting it as fact.
Frame the work as research into the workflow, never as an assessment of individual productivity. Obtain informed participation, minimise collection, control access, reconfirm consent before introducing recording and handle personal data securely. Where employee, customer, confidential or regulated information may be involved, bring in the appropriate privacy, security, legal, employment-relations, accessibility and domain owners. End by asking the participant to confirm, correct or dispute the reconstructed flow.
Artefacts, short task diaries and operational records reveal parts of the workflow that a scheduled conversation or observation may miss. Ask participants to explain the templates, spreadsheets, messages, drafts, notes, checklists and queue views used in a case. Their markings and arrangement may expose organising and tracking practices absent from the official map, but the participant must confirm what each artefact means and how it was used.
Use a brief diary when relevant events occur intermittently or across time, and follow entries with clarification because diaries remain self-reported evidence. Where usable case, activity and timestamp records exist, inspect waiting, returns, duplicated handling, deviations and recurring quality problems. Document work outside those systems as well: an event log may omit offline clarification, span disconnected tools or lack a consistent definition of a case.
| Evidence method | What it can reveal | What it can miss | How to corroborate |
|---|---|---|---|
| Recent-case interview | Sequence, experience, decisions and handovers | Unnoticed actions and reliable recurrence | Compare cases, roles, artefacts and records |
| Contextual observation | Tools, interruptions, workarounds and situated judgement | Infrequent events and unspoken motives | Ask neutral questions and confirm the reconstruction |
| Workflow artefacts | State, memory cues, tracking and coordination | Why or how often an item is used | Ask the participant to explain it in a real case |
| Short task diary | Real instances distributed across time | Unreported events and independent verification | Follow up with interviews, observation or artefacts |
| Operational records | Timing, variants, rework and recurring patterns | Unlogged work and reasons behind events | Reconcile with people, cases and artefacts |
A pain point becomes a credible bottleneck hypothesis when evidence places a recurring constraint at a specific workflow point and connects it to an observable consequence. That is still a hypothesis, not causal proof. Use the following six-question test as an editorial decision rule, preserving negative cases and alternate explanations rather than compressing disagreement into a confidence score.
In the onboarding example, the evidence shifts the hypothesis from slow summarising to missing decisions and contradictory fields in difficult cases. Routine briefs are completed quickly, and longer briefs are not consistently slower. Available timestamps also omit offline clarification. That combination justifies testing the upstream handover, but not claiming that missing information has been proved to cause every delay or that drafting assistance has no potential value.
Do not ask where AI fits until you can show where a recurring constraint appears, what follows and what remains uncertain.
An evidence-backed opportunity card records the bounded work, supporting evidence, counterevidence, constraints, alternatives and next smallest test. It is a decision record, not a sales pitch. NIST's voluntary AI RMF Playbook supports documenting purpose, users, operational setting, expected value, limitations, impacts, scope and human oversight, while also considering non-AI and non-technology alternatives. The card should make unresolved assumptions visible rather than presenting them as findings.
Link each substantive statement to a field-note identifier, redacted artefact identifier, diary entry or record query, or mark it explicitly as an assumption. Compare every intervention against the same workflow evidence. Official discovery guidance also retains an evidence requirement, an owner, a next milestone and a later test; agreement in a workshop is not proof that value will appear in real work.
For onboarding briefs, first test required intake fields, an explicit sales-handover owner and a visible exception queue. Consider a bounded drafting assistant later, and only for fact-complete cases. Finish with one decision: test AI, test a non-AI change, gather more evidence or stop. If recurrence, location, consequence or actionability remains unclear, further research—or no intervention—is more defensible than forcing the complaint into an AI proposal.
Start by bounding recurring work, then compare recent cases, observations, artefacts and any usable operational records. Look for a constraint with a clear location and observable consequence, preserve uncertainty, and compare AI with simpler interventions before defining a controlled test.
A pain point is a valid report of frustration or difficulty. An opportunity is a bounded task supported by evidence of recurrence, workflow location and consequence, with known constraints and a plausible intervention that can be tested.
Make participation informed, frame the study as workflow research rather than individual assessment, and collect only what the research question requires. Restrict access, disclose the observation mode, invite participant corrections and involve the appropriate organisational owners before handling sensitive material.
No. The method is a practical synthesis, not a universal checklist. Choose the smallest combination that answers the workflow question and offsets known blind spots; some studies may not need diaries or have usable operational records.
No. Observation and operational evidence can support a bottleneck hypothesis, but both are partial and open to alternate explanations. Preserve counterevidence and run the smallest test that can distinguish between the leading explanations.
This article was researched using the following sources:
We report on how AI actually lands inside a business. Our work starts from named sources, separates what we found from what we think, and uses AI assistance for research and drafting under documented editorial controls. We are not a substitute for individual expert review.
A practical method for mapping real workflow behaviour, testing controls, defining handoffs and resolving exceptions before automation begins.
A practical guide to mapping tasks, testing AI assistance, tracing hidden effort and redesigning roles only when pilot evidence is sound.
A practical guide to assigning AI decision rights, choosing an operating structure, and turning pilot evidence into portfolio and strategy change.
