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 Team9 min read
A complaint is a research lead, not yet an AI use case. When an operations team says client onboarding briefs take too long, a remembered process map may make drafting look like the obvious target for an AI summariser. Completed cases can tell a different story: difficult briefs may wait for missing commercial decisions and then be rewritten when contradictory fields are resolved. Automating the visible writing step would leave that constraint in place. The sounder starting point is a bounded sample of real work, examined closely enough to locate the friction, connect it to an operational consequence and preserve what remains uncertain.
What to carry into the field
Define the workflow by its trigger, end condition, output, downstream user, participating roles and relevant case types before collecting evidence. Start with recurring work and real examples, then state the questions the field round must answer. For the onboarding example, the boundary runs from an approved sales handover to a brief accepted by a delivery lead, covering routine, incomplete and changed-scope cases. Record where handovers occur, how outputs vary and which dependencies sit just outside the boundary. Keep the boundary provisional: observation may show that an upstream decision or downstream correction belongs inside it.
Build the evidence plan around the question, not around a fashionable research technique. The method here is a practical synthesis rather than a standardised protocol: interviews reconstruct experience, observation captures work in context, artefacts expose working state, diaries extend the view across time, and records test patterns at greater scale. Use the smallest proportionate mix that can answer the question and offset obvious blind spots. A recurring queue may need interviews plus records; an undocumented judgement may require a recent-case interview, a demonstration and the artefact used.
Ask the participant to reconstruct one specific, recently completed case from trigger to accepted output. Concrete stories are more useful than a polished account of how the process is meant to work, while still respecting the participant's report as evidence of experience. Establish what arrived, what happened next, which decisions were made, who received each handover, what was uncertain and where the case departed from the expected path. Open, neutral prompts keep the account grounded: “What happened next?”, “What told you that?”, “Who did you contact?” and “Can you show me what you used?”
Probe the thinking behind visible actions as well as the actions themselves. Ask which cues mattered, what information was missing, which alternatives were considered and what made a judgement difficult. This is compact cognitive-demand probing, not the formal ACTA protocol. Demonstrations and artefacts can surface task detail omitted from verbal descriptions, but one case remains illustrative. Contrast a routine onboarding brief with a difficult one and include coordinators, sales operations and delivery leads, whether separately or in a suitable joint session, so that the reconstruction crosses the handovers.
Watch the work with its usual tools, documents, data, interruptions and dependencies, while treating the workflow rather than the employee as the subject of research. Note source switching, waiting, verification, rework, handovers, workarounds and moments when incomplete or conflicting information requires interpretation. Choose the observation mode explicitly. 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. Record the chosen mode so later reviewers do not mistake elicited behaviour for an untouched routine.
Use informed participation, minimise researcher influence, reconfirm consent before introducing any additional recording and handle personal data securely. Collect only what the research question requires, restrict access and avoid covert monitoring, individual productivity scoring or continuous capture. Where employee, customer, confidential or regulated information is involved, bring in the organisation's appropriate privacy, security, legal, labour, accessibility and domain owners. These safeguards do not settle local obligations. Close the session by asking the participant to correct the reconstructed flow and flag interpretations that do not match their experience.
Keep observed action separate from interpretation in the field notes. “Coordinator opened three sources” is an observation; “the system creates unnecessary complexity” is an interpretation requiring corroboration. Contextual inquiry is granular and subjective, and observation can alter behaviour. Participant confirmation creates a stronger shared reconstruction, but it does not make a small sample representative of the workforce or prove that the inferred mechanism caused the outcome.
Use artefacts, short task diaries and operational records to extend the field view where interviews or scheduled observation leave gaps. Ask participants to explain the templates, checklists, spreadsheets, messages, drafts, paper notes and queue views used in the case. Markings and formats can reveal organising, tracking, memory and coordination practices absent from an official map, but their meaning must be confirmed. A coloured flag might indicate priority, ownership or merely personal preference; the artefact cannot explain itself or show how frequently the practice occurs.
A brief diary is useful when relevant events are intermittent, distributed or unlikely to occur during a booked session. Tie entries to real task instances and follow up for clarification, remembering that diaries remain self-reported evidence. Where usable case, activity and timestamp records exist, inspect waiting, rework, duplicated handling, deviations, missed deadlines and recurring quality problems. Record what happens outside those systems. Less common exceptions still deserve attention because their effort or consequence may be material, but no illustrative distribution should become a universal threshold.
| Evidence method | What it can reveal | What it can miss | How to corroborate |
|---|---|---|---|
| Recent-case interview | Sequence, experience, decisions, uncertainty and handovers | Unremembered detail and recurrence across cases | Inspect artefacts, observe a task and sample contrasting cases |
| Contextual observation | Real tools, interruptions, workarounds and support activity | Rare events, private reasoning and behaviour outside the session | Ask neutral questions, review notes with participants and use diaries |
| Workflow artefacts | State, priority, memory cues, revisions and coordination practices | Why, when and how often an item is used | Ask the user to explain it in a specific case |
| Short task diary | Intermittent work close to the point of occurrence | Unreported events, ambiguity and declining participation | Follow entries with interviews, artefacts or observation |
| Operational records | Recurrence, timing, queues, variants, rework and returns | Unlogged work, motives and a reliable cross-system case definition | Reconcile patterns with people, artefacts and field observations |
A reported pain point becomes a credible bottleneck hypothesis when multiple relevant cases or evidence sources place a recurring constraint at the same workflow point and connect it to an observable consequence. It remains a hypothesis, not causal proof. Compare researcher interpretations with participant understanding and reconcile operational records with observation and artefacts. Keep cases that avoid the problem, differences between roles, missing coverage and alternate explanations visible. Do not collapse disagreement into a neat confidence score that hides what the team still does not know.
For the onboarding workflow, the evidence may shift the hypothesis from slow summarisation to missing decisions and contradictory fields in difficult cases. Routine briefs are completed quickly, and longer briefs are not consistently slower. Available timestamps omit offline clarification, so they cannot establish why a delay occurred. That pattern justifies a discriminating test, not a declaration of cause. If recurrence, location or consequence is still missing, retain the complaint as an open research lead and gather more evidence rather than forcing it into an opportunity pipeline.
Do not ask where AI fits until you can show where the constraint recurs, what follows and what remains uncertain.
An evidence-backed opportunity card records the bounded workflow, observed sample, stated pain, observed pattern, operational consequence, counterevidence, constraints, alternatives and next test. Separate what participants reported from what the field round observed. Reference field-note identifiers, redacted artefact identifiers, diary entries and record queries for substantive assertions, or label an unverified statement as an assumption. Document the trigger, end condition, output, downstream user, roles, case types, period and method mix so another reviewer can see what the evidence covers and what it cannot support.
State the opportunity as who needs help with which bounded task and what better outcome is sought. Compare AI with deterministic rules, clearer ownership, workflow redesign, training, better information and no intervention against the same evidence. Then choose one decision: test AI, test a non-AI change, gather more evidence or stop. For onboarding, first test required intake fields, an explicit handover owner and a visible exception queue. Consider drafting assistance later, only for fact-complete cases, with defined success and failure evidence rather than workshop enthusiasm standing in for operational value.
Propose AI only when the evidence supports a bounded task, a recurring constraint, an observable consequence, workable information, appropriate oversight and a testable advantage over simpler alternatives. Before observing sensitive work or handling employee, customer, confidential or regulated information, involve the relevant organisational owners and seek qualified professional guidance where obligations or high-stakes decisions arise. This field method supports better questions and smaller tests; it is not a legal, labour, security or compliance determination.
Start with recurring work, define its trigger, output, roles and case types, then examine real cases through a proportionate mix of interviews, observation, artefacts, diaries and records. A credible opportunity links a recurring constraint to an observable consequence, records uncertainty and compares AI with non-AI alternatives before a controlled test.
A pain point is a valid report of frustration or difficulty and should be treated as a research lead. An AI opportunity is narrower: evidence places a recurring constraint within a defined workflow, connects it to an operational consequence, identifies relevant constraints and supports a test in which AI has a plausible advantage.
Frame the activity as workflow research, secure informed participation and collect only what the research question requires. State the observation mode, restrict access, avoid individual scoring or covert capture, let participants correct the reconstruction, and involve appropriate privacy, security, legal, labour, accessibility and domain owners.
No. The combined method is a practical synthesis, not a mandatory checklist. Choose the smallest mix that answers the workflow question and compensates for known blind spots, such as pairing interviews with records when recurrence matters or adding observation when remembered descriptions omit consequential steps.
No. Observation and operational records can support a bottleneck hypothesis by showing recurrence, location and associated consequences, but they do not establish causation on their own. Preserve negative cases and alternate explanations, then run the smallest test that can distinguish among plausible mechanisms.
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 South African guide to mapping real workflow behaviour, testing approvals, defining handoffs and resolving exceptions before automation.
A practical South African guide to mapping tasks, testing AI assistance, tracing hidden effort and redesigning roles after a measured pilot.
A practical guide to assigning AI decision rights, balancing central and business-unit authority, and turning pilot evidence into strategy changes.
