Part one · Diagnose before teachingThe course arrives before the context does

A customer-service team seldom opens its new AI assistant. Management commissions another confidence course. In the classroom, the assistant answers a neat sample in seconds. At work, advisers must copy an account history from another system, remove protected data, wait for a response, verify unsupported citations and decide where an exception goes. No one owns the final queue. The team does not need more confidence in the demonstration. It needs a workflow that can complete the job.

Training is a tempting first response because it is visible, purchasable and assignable. A course has a budget line, a calendar and a completion report. Integration repair crosses teams. Authority repair requires executives to decide. Capacity repair admits that promised savings created new work. Product repair may challenge a launch date or supplier. Calling the condition a capability gap places the action on the user and keeps those harder owners out of view.

A person cannot learn around absent context, duplicate authentication, unusable latency, inaccessible design, unsafe recommendations, responsibility without override, an orphaned exception or verification work with no capacity. Teaching a workaround may keep a service moving briefly, but it must be labelled as temporary protection, not repair. Otherwise the organisation turns preventable system effort into a permanent human competence requirement.

Training becomes justified after two claims survive evidence. First, the capability fits the real task, including its difficult exceptions, affected populations, authority and workload. Secondly, a defined knowledge, procedural skill, perceptual cue, judgement pattern or coordination behaviour still explains material performance variation. The proof comes from representative work samples and transfer into work, not from attendance, satisfaction, self-confidence or a perfect demonstration.

The unit of success is a consequential work episode. The person obtains the right information, uses or refuses the AI appropriately, exercises real authority, handles an exception, recovers from error and produces a better outcome inside sustainable workload. Usage may be high or low. Appropriate abstention can be success. A fast visible step can fail if downstream rework, queue or harm grows.

Part one · Diagnose before teachingTen planes surround one work episode

The Workflow Fit Map is this paper's synthesis of ten interdependent planes. It maps actual work and its exception path, not the procedure manual. The planes remain separate because one missing condition can govern the whole episode.

Outcome and decision specifies intended result, decision population, consequence, permissible action, baseline and improvement evidence. Task sequence traces trigger, steps, dependencies, queues, batching, variation, rework and where AI enters or exits. Information and data follows required context, provenance, meaning, access, freshness, completeness and what the user must supply or verify.

Capability and interaction examines model behaviour, interface, explanation, uncertainty, controls, input burden, output form and adaptation. Timing and attention covers latency, interruptions, cognitive load, concurrency, urgency, frequency, alerting and review time. Handoff and exception identifies upstream and downstream owners, ambiguous cases, failure states, escalation, fallback, recovery and closure.

Authority and accountability names who may decide, approve, override, abstain, correct, escalate, stop and accept consequence. Workload and capacity reveals checking, correction, documentation, queues, emotional labour, support, staffing and downstream absorption. Physical, social and accessibility context includes location, device, connectivity, language, disability, privacy, collaboration, shift, professional norms and safety. Feedback and learning connects outcomes, incidents, user and affected-party feedback, versions, support and retirement.

Do not average the planes. A usable interface cannot compensate for missing authority. A fast model may create a worse workflow if correction shifts to another team. An accurate response can still be irrelevant when it arrives after the decision. The map's output is a condition, owner and consequence, not a number.

Part two · Attribute frictionReproduce the job before assigning the learner

The Workflow Friction Attribution Protocol proceeds: observe, reproduce, trace, compare, isolate, repair, retest, then train if necessary. It treats user behaviour as the start of causal inquiry.

Observe records the work event, population, environment, frequency and consequence without assuming reluctance or incompetence. Reproduce uses the real device, data, access, timing, authority and collaboration conditions. Trace follows information, actions, waits, checks, rework, handoffs, escalation and outcome from trigger to closure.

Compare puts successful beside failed episodes, users beside non-users, ordinary beside exceptional cases, and affected populations beside the convenient average. Isolate states the leading cause, credible rivals and evidence that would weaken each. Repair changes the smallest responsible capability, integration, workflow, authority, capacity, control, support or boundary. Retest examines outcome, burden, safety and recovery. Only then does train if necessary specify a residual capability gap and its transfer environment.

One reproduced case can be more revealing than a survey. A field worker may report low confidence, yet a field reproduction shows that the mobile connection resets authentication halfway through the task. The self-report is genuine but not causal proof. Conversely, a smooth interface may reveal a real judgement gap when novices repeatedly miss a defined uncertainty cue that proficient peers recognise under the same conditions.

Part two · Attribute frictionTwelve conditions before training becomes the repair

The Training Necessity Gate makes the burden of proof visible. Training may provide urgent safety orientation while a gate is unresolved, but it must not be presented as the primary fix.

The target event, population and outcome must be stable enough to teach. The capability needs adequate evidence and a defined operating boundary. Information, integration, interface, timing and device must permit the task. Authority, accountability, override, abstention, escalation and stop rights must be real. Exception, fallback, incident, correction and recovery routes must work.

Workload, staffing, attention, support and downstream capacity must absorb the changed work. Accessibility, language, privacy, safety and professional conditions must fit. Work-sample evidence must identify a specific knowledge, skill or judgement gap rather than inferred reluctance. The proposed learning must plausibly change it. The workplace must provide access, opportunity, coaching, feedback and stable versions. Transfer and adverse consequences must be measurable. Product, workflow, authority and capacity owners sign the conditions they own.

The gate is proportional. A low-consequence drafting aid may need a light reproduction and clear boundary. A clinical, financial, employment or rights-affecting episode demands deeper evidence, exception tests and independent challenge. The organisation should not wait for perfect fit, but it must not knowingly call a system defect a learner problem.

Part three · Prove the remaining gapRepresentative work includes the correct refusal

A Representative Work-Sample Protocol recreates the decision, authority and consequence of work. The sample set includes ordinary, difficult, rare, adversarial, justified non-use and recovery cases.

Each sample identifies task, trigger, desired outcome, population, affected party and consequence. It records current and proposed workflow, time, handoffs, information, authority, access, support, model version, configuration, data and interface. Expected behaviour includes correct use, justified abstention, override, escalation and stop. The visible product and downstream outcome are both observed.

Error consequence, detectability, recovery and residual exposure travel with the sample. Workload, attention, language, device and accessibility conditions match the environment. Novice, proficient and expert participants may be compared when relevant. The evidence must separate product, workflow, authority and skill causes. Every sample has a retest threshold, owner, expiry and change trigger.

A Capability-Gap Specification is written only after these samples isolate a residual gap. It names the role population; precise knowledge, procedural skill, cue, judgement or coordination behaviour; observable distinction between proficiency and error; prerequisites; included and excluded cases; appropriate use and non-use; coached practice; consequence-linked threshold; real-work transfer observation; booster, retraining, version-change and expiry rules.

Broad AI literacy can support informed participation, but it cannot prove role proficiency. A general course may help an underwriter understand uncertainty. It does not show that the person will detect the particular cue, select the correct evidence, override appropriately and recover inside the production process.

Part three · Prove the remaining gapGive every condition an owner and a sequence

The Fit-to-Training Decision Docket prevents a learning team from inheriting product, workflow, authority and capacity defects. It records the evidence, choice, owner, closure and expiry for one work episode.

The docket begins with observed behaviour and a neutral problem statement, then names outcome, workflow, population and consequence. It includes the ten-plane map, reproduced case, end-to-end trace, leading cause, rivals and confidence limits. Product, integration, data, workflow, authority, capacity, accessibility and support findings remain separate. A specific capability gap appears only if one survives.

Options include product repair, workflow redesign, changed authority, funded capacity, a learning intervention, continued manual operation and non-deployment. The docket names one owner for each repair, chosen sequence, interim protection and non-action consequence. When training is justified, it records objective, work sample, practice, feedback, assessment, transfer conditions, manager obligation and support route. Outcome, burden, safety and affected-party measures define closure.

Closing the docket means that the work condition changed and the episode passed retest. Training assigned, a backlog item closed or a policy published is not closure. If an unresolved condition requires temporary instruction, the docket labels it, names the interim protection and keeps the responsible owner and deadline visible.

Part four · Follow learning into workCapability decays when the environment changes

The Training Transfer and Decay Test follows behaviour beyond practice. It asks whether the workplace still offers the access, opportunity, authority, support and stable version used to learn.

Observe opportunity to perform; access to the same system, information and tools; manager and peer support; workload and time; representative work behaviour rather than self-report; appropriate reliance, non-reliance, override, escalation and recovery; downstream outcome, error, delay and rework; differences by subgroup, shift, language, device and accessibility; drift in model, interface, data, policy or workflow; retention over time and after rare exceptions; and shortcuts or shadow work.

A version change can invalidate evidence without anyone becoming less skilled. If a vendor changes model behaviour and renames interface controls, old practice may teach the wrong cue. If a manager removes protected review time, the behaviour may disappear although knowledge remains. If a rare task occurs twice a year, memory decay and absent practice may make an AI route unjustifiable. The response may be product repair, coaching, a booster, renewed work samples or withdrawal.

The Workflow Fit Evidence Panel keeps seven views apart: task and outcome; ten-plane fit; friction attribution; specified capability; transfer; burden and inclusion; lifecycle. A completion figure cannot cancel evidence of growing downstream rework. High use cannot compensate for incorrect reliance. A confident cohort cannot stand in for excluded language or accessibility groups.

Part four · Follow learning into workTen cases, only one clear training-first result

The cases apply the same diagnostic sequence. The point is not that training is rarely useful. It is that its proper scope becomes visible only after other causes are tested.

Part five · Exercise the decisionWorkflow Fit Before Training Lab

The lab routes a work-system diagnosis. It does not calculate readiness, confidence or adoption probability. Start with a named case, inspect the evidence fields and alter the controls to test a local episode.

The lab can challenge a proposed intervention, but it cannot validate the model, adjudicate rights or observe a real outcome. Its output should become a Fit-to-Training Decision Docket owned by the people who control the implicated conditions.

ConclusionGive people work worth learning

Workers deserve systems fit to learn and learning fit for the work they are actually expected to perform. That standard places training after diagnosis, not outside responsibility.

The discipline is simple to state and demanding to operate. Map ten planes around one consequential episode. Reproduce the work in its real device, clock, authority and social setting. Trace information, checks, handoffs, exceptions and outcome. Compare successful, failed and non-use cases. Isolate the leading cause and credible rivals. Repair the smallest responsible condition and retest the whole episode.

If a knowledge, skill or judgement gap remains, specify it precisely. Practise ordinary, difficult, rare, adversarial, justified non-use and recovery cases. Observe behaviour inside real authority and consequence. Follow transfer into work, including burden and affected-party outcome. Expire the evidence when model, interface, data, workflow, policy, role or population changes.

A time-boxed diagnosis need not become bureaucracy. A small set of representative episodes can expose whether the obstacle sits in the product, the sequence of work, decision rights, operational capacity or a learnable human capability. That evidence also makes action faster: each failure has an owner, an interim protection, a retest condition and a closure claim. Where immediate risk exists, provide safety orientation at once while the responsible owners repair and verify the conditions around it.

This sequence does not diminish learning. It protects learning from becoming an alibi for poor design. The final decision is not whether to buy a course. It is whether the work system can support correct action and, within that system, what a person genuinely needs to learn.