TLDR

1. Map units of work through states, queues, handoffs, decisions, controls and outcomes before selecting an AI task.
2. Separate touch time, waiting, rework and approval delay; local handling speed can rise while end-to-end time worsens.
3. Find the actual constraint with event evidence. Feeding a fixed bottleneck faster increases work in progress and ageing.
4. Keep removal, policy change, integration, deterministic software and operating redesign beside learned intelligence as valid interventions.
5. The Workflow Opportunity Lens routes a flow to repair, bounded AI testing, constraint redesign or a non-workflow method.

A team halves processing time at step three. Step five remains capacity-constrained and now receives more variable, lower-quality input. What happens to completed outcomes, work in progress and customer elapsed time?

Step three looks successful on its local dashboard. Its queue shrinks and its throughput rises. The system can still deteriorate. Work reaches step five faster than step five can finish it, so the downstream queue grows. Variable input increases review and rework. Customers wait longer even though one activity became faster. The intervention has shifted inventory and perhaps added failure demand rather than improving the end-to-end outcome.

This thought experiment captures the workflow reason many AI pilots disappoint. Opportunity lists identify visible cognitive work such as drafting, summarising, classifying or searching. A product then reduces handling time in that activity. Yet the economic or service outcome depends on the joins around it: when evidence arrives, who owns the next state, how exceptions route, where approval waits, whether context survives and how errors return.

A workflow lens follows a unit of work from trigger to outcome. It records authoritative states, work and wait, arrivals and ageing, decisions and required evidence, handoffs, exception classes, control points, rework loops, fixed constraints and feedback. The output is not a prettier process diagram. It is a testable intervention hypothesis tied to the completed flow.

The method remains solution-neutral. Some joins need an interface or shared state. Some need a policy decision, queue rule or ownership change. Some need deterministic validation. Learned intelligence earns a place where an ambiguous transition has usable evidence, measurable headroom and a safe consequence boundary. The discovery question is which mechanism moves the entire outcome, not which feature makes one station impressive.

Figure 1. Work moves across islands; elapsed time collects in the water

Workflow river

01 · Unit and statesTrace completed work, not departmental activity

Applied-AI opportunity is often created or destroyed in workflow joins, so discovery must map the unit of work across state transitions before optimising its most visible task. Define the unit first: one credit file, complaint, claim, order or service request. State the entry trigger and the completed outcome. A flow without a stable unit becomes a tour of organisational boxes.

Use states that an authoritative system or accountable role can recognise. “Documents received”, “evidence complete”, “analysis ready”, “exception under review” and “decision recorded” are stronger than “finance is working on it”. Record which system owns each state and which events move the unit. When state lives in spreadsheets, inboxes and memory, the map exposes a state-management problem before it proposes prediction.

Separate touch time from waiting. Touch time is active work on the unit. Waiting includes queue, batching, missing evidence and pending authority. Rework is repeated activity caused by an earlier defect or changed requirement. Approval delay deserves its own category because the official task may be brief while the calendar wait is long. These distinctions prevent a drafting-time gain from being confused with an elapsed-time gain.

Davenport and Short argued for information technology and business-process redesign across functional boundaries, with attention to outcomes rather than isolated activities.2 Hammer's famous warning against automating inherited work made the same provocation in sharper language.3 Their era differs from generative AI, but the operating lesson persists: technology can encode unnecessary handoffs as efficiently as it can remove them.

Map arrival rate, throughput, queue size and ageing where data permit. The busiest step is not automatically the constraint. A station may look busy because upstream work arrives in batches or because downstream rejection creates repeats. Age distributions are often more revealing than average queue length. A small set of old cases can represent a distinct exception path that needs ownership rather than general acceleration.

Little's queueing relationship connects the average number of units in a stable system, the effective arrival or throughput rate and average time in the system.1 Use it narrowly. The original proof relies on finite means and stationarity conditions. Real service systems experience changing demand, priorities, abandonment, rework and unstable capacity. The formula is a consistency check, not a complete process model or a licence to treat every observation period as steady.

If completed throughput does not change while work in progress rises, average time must rise under those conditions. That is the local-speed trap. Accelerating intake can create a larger visible pipeline while customers wait longer. The intervention should be assessed on completed, acceptable outcomes and end-to-end elapsed time, not units entering the next queue.

Build the baseline over a period that contains the demand and staffing variation the intervention will face. Report medians and tails, not only means; an improvement for routine cases can coexist with worsening old-case ageing. Reconcile arrivals with departures and abandonment so the denominator does not silently change. Note planned outages, policy changes, campaigns and staffing shocks. The objective is not a perfect process model. It is a baseline precise enough to show whether the completed flow improved and which population absorbed the change.

Figure 2. A fast station floods a fixed downstream gate

Local-speed trap

02 · Time and constraintsMeasure where the calendar goes

A work-and-wait barcode is a useful evidence artefact. Sample representative units and place every interval on one elapsed timeline: active handling, queue, missing information, approval, rework and idle time. Repeat by outcome and exception class. The result often overturns workshop beliefs. A step people complain about may consume substantial effort but little calendar time; an innocuous handoff may explain most delay.

Event logs help when state transitions are captured reliably. Observation and case review remain necessary because logged completion may differ from actual completion, workarounds may occur outside the system and timestamps can represent batch entry. Reconcile data with frontline accounts and inspect outliers rather than deleting them as noise. Exception cases frequently contain the discovery value.

Constraint confidence should be explicit. A stable queue before one specialist group with consistently high utilisation, long ageing and downstream starvation is a strong signal. A queue that moves unpredictably with missing evidence, policy revisions or seasonal demand may have several causes. Testing one cause at a time is better than labelling the visually longest bar the bottleneck.

Capacity also has a quality dimension. A reviewer may appear to complete ten units a day, but three return because upstream evidence is unreliable. Counting first passes overstates effective throughput. Define a completed unit as one that meets the required state and survives the relevant rework window. When quality and volume trade off, show both. An AI intervention that raises first-pass volume while increasing returns has not expanded usable capacity, even if the local productivity calculation improves.

Fixed constraints can be policy, authority, capacity, integration, physical operations or customer response. A model cannot remove a policy that requires committee review. Prediction might help schedule scarce capacity or route material exceptions, but only if the owner can change the queue rule. The workflow record therefore includes change authority beside the constraint.

The second thought experiment contains ten manual steps. Nine are deterministic lookups; one is an ambiguous exception judgement. Should automation potential be measured by step count? No. The nine lookups may be removed through integration or rules, producing a simpler and safer flow. The remaining judgement might benefit from evidence retrieval or a bounded prediction test. Counting automated steps would over-reward adding AI to work that ordinary software should eliminate.

Figure 3. Work, wait, rework and approval form an elapsed-time barcode

Case timeline

Exception concentration changes the design. Routine units may follow a straight-through path that deterministic validation can improve. A small residual can consume most expert attention because evidence conflicts, policy is unclear or consequences are material. Building one average flow for both produces either excessive automation or unnecessary review.

Classify exceptions by cause and routing need. Missing evidence, conflicting evidence, policy exception, customer vulnerability, calculation uncertainty and system failure should not share a catch-all queue. Record frequency, ageing, owner, consequence and whether a repeated exception indicates a process defect. Some “AI opportunities” disappear when the dominant exception is fixed at source.

Use an intervention ladder after diagnosis. First ask whether the state or step can be removed. Then test simplification, shared evidence, integration, deterministic validation and queue redesign. Only the unresolved ambiguous transition reaches learned intelligence. This ordering does not assume that simpler technology is always better; it makes the additional role of intelligence explicit. The final hypothesis should name the changed event, the expected effect on the downstream distribution, the affected exception class and the condition that would stop the test.

Figure 4. A small exception delta can consume most expert attention

Residual flow

03 · Handoffs and feedbackContext decays when ownership changes

Handoffs create value loss when the receiving role cannot see why prior choices were made. Evidence may be copied without provenance, a summary may omit uncertainty, or a state label may conceal unresolved exceptions. The next person reconstructs context by reopening documents, sending messages or repeating checks. This work appears necessary at the receiving station but originates in an earlier information design.

Map evidence fields across the handoff. Which items are required? Which are authoritative? Which are transformed, summarised or dropped? Which uncertainties and exclusions survive? Which role can correct the record? A generative summary may improve reading speed and worsen context loss if it hides citations or unresolved contradictions. A structured evidence pack can be more valuable than fluent prose.

Failure demand is work created because the flow failed earlier: repeat contacts, re-opened complaints, resubmitted files, correction, remediation and escalation. Treat it as a loop, not fresh demand. If a chatbot closes a conversation that returns through another channel, local completion rises while the workflow outcome worsens. Link identifiers where lawful and assess closure across the relevant window.

Feedback closes the map. Outcomes should update policy, evidence requirements, queue rules and model monitoring. A decision that produces no observable downstream signal cannot support rapid learning. Where the outcome lag is long, intermediate receipts must be chosen carefully and their relationship to the final objective tested. Prediction quality, action taken and workflow outcome remain separate measures.

A practical evidence pack contains a state dictionary, event-quality audit, sampled timelines, queue-age distribution, exception taxonomy, handoff-field comparison, rework links and owner map. Preserve the raw case identifiers under appropriate access controls so claims can be traced. Where linking is restricted, use approved aggregate or privacy-preserving analysis and state what cannot be observed. Missing observability is a discovery result because it can redirect investment towards state capture before model development.

Figure 5. Evidence fields fall away and are rebuilt after each handoff

Context decay

04 · Worked workflowsCredit review and complaint handling

Commercial-credit review

A useful credit map begins when documents arrive and ends when the decision and conditions are recorded in the booking system. Between those points sit completeness checks, extraction, calculation, policy interpretation, exception escalation, memo preparation and committee review. Approval remains with the authorised officer or committee. The map is for opportunity discovery, not an autonomy design.

Missing evidence at entry often creates late rework. A file can appear active while analysts extract what is present, only to restart when a missing period or schedule is discovered. The loop consumes preparation time, delays committee readiness and causes context switching. An early completeness service can combine versioned requirements, classification and evidence highlighting. It changes the timing of the request without making the credit judgement.

Calculation is a mixed state. Locating candidate inputs can involve document understanding; the authorised formula should be deterministic. Period comparability and policy relevance require judgement when evidence conflicts. The workflow should preserve the difference, routing uncertain cases with sources and reasons. A single “covenant automation” label would hide these transitions and make evaluation incoherent.

Committee review may be the constraint, but its queue can be caused by poor readiness rather than seat capacity alone. Measure how many files arrive incomplete, how often conditions change, which exception classes consume time and how ageing varies. Faster memo drafting could worsen the queue if it sends more variable files forward. A readiness gate or better evidence pack may move end-to-end time more.

A bounded credit experiment can phase the completeness intervention across comparable intake teams or file groups while keeping approval unchanged. Pre-register the eligible population, required-evidence version, late-request outcome, rework definition and safety exclusions. Track whether alerts lead to earlier customer requests, not merely whether the classifier is accurate. Review false alerts by document and customer type. Stop if burden shifts to customers without reducing downstream restarts, or if analysts bypass the official evidence record.

Figure 6. Credit review exposes an early evidence loop and retained authority

Worked flow

Customer complaint handling

The complaint unit begins with a customer expression of dissatisfaction and ends with a durable remedy, communication and recorded closure. Authentication, issue classification, evidence gathering, vulnerability detection, remedy calculation and communication approval sit between. Repeat contact, reopened cases and ombudsman escalation are outcome signals, not separate demand to ignore.

Response drafting is visible and suitable for demonstration. It may be far from the constraint. If evidence gathering and remedy authority create most delay, faster prose arrives at an approval queue sooner. Generated variation can increase review. The negative control therefore requires end-to-end elapsed time, first durable resolution, reopened cases, complaint quality, vulnerability outcomes and review effort beside drafting time.

Vulnerability detection is an asymmetric exception path. A signal can route the unit to specialist support, but false negatives risk harm and false positives can be intrusive. The map identifies the exact transition, evidence, reviewer, recovery and appeal. Local containment targets cannot push these cases into a lower-control path.

Remedy calculation may be deterministic once facts and policy are established. Communication can be drafted with cited evidence and controlled templates while a named role releases it. Classification may assist routing, but a statutory or policy complaint label should remain auditable. Each intervention is evaluated by its effect on the completed complaint outcome rather than feature usage.

For complaint experiments, preserve a contemporaneous comparison where operationally and ethically feasible. Slice by issue, channel, vulnerability route and complexity. Measure review edits, elapsed time, durable resolution, repeat contact, reopening, escalation and correction. A generated response should carry the evidence and policy version used so reviewers can distinguish fluent language from supported content. If drafting speed improves but release waits, the next experiment belongs at the authority or evidence join rather than in a larger language model.

Figure 7. Complaint flow protects vulnerability, remedy and release authority

Outcome flow

05 · Strongest rivalTask precision beats workflow theatre

The strongest rival is task or decision analysis. Broad process mapping can consume weeks, produce walls of sticky notes and document a standard process nobody follows. It can delay a straightforward experiment on an isolated expert task. The rival wins when the work has no material queue, handoff or downstream coupling and the outcome can be tested directly.

Accept the criticism by imposing evidence and closure rules. Build the map from event records, observed cases and representative exceptions. Mark uncertainty rather than filling gaps with consensus. Every proposed intervention must reduce to a specific state transition, decision, queue rule or evidence handoff with a baseline and owner. If the map cannot change an investment decision, stop mapping.

Set a time box and a decision audience before discovery begins. A two-week reconstruction of a bounded flow can be more useful than an enterprise process taxonomy that is never maintained. Publish the disputed points, missing events and rival explanations alongside the diagram. The review question is whether evidence supports a different intervention or measurement plan. This keeps mapping accountable to a decision and gives task specialists a clear point at which their narrower method takes over.

Do not map highly emergent research as a linear flow. Genuine inquiry, novel strategy and crisis response may have recurring practices but no stable unit or state sequence. Use event reconstruction after episodes and bounded probes before the next one. Manufacturing a neat process would hide uncertainty and create false baselines.

Current adoption research supports attention to operating complements. The OECD notes that firms often underestimate process change, skills, data and organisational implications in AI adoption.4 Census evidence shows differing depth at firm, function and worker-task layers.5 Task-level field effects remain tied to specific populations and settings.67 These sources justify local workflow evidence, not generic transformation claims.

06 · Executable instrumentChoose the next workflow intervention

Run the Lens on one bounded flow. Include the end-to-end owner, people at upstream and downstream states, evidence and system owners, control specialists and people affected by failure. Bring event samples and exception cases. The inputs are diagnostic categories, not a scorecard for choosing a model.

Workflow Opportunity Lens

Assess one unit of work from trigger to completed outcome. The result chooses the next intervention hypothesis; it does not authorise automation.

Unit and end-to-end outcomeNo stable unit or completed outcomeDepartmental activity onlyUnit, trigger and outcome namedVersioned population and outcome owner

Event and state observabilityProcess cannot be observed crediblyWorkshop account onlyEvent samples reconciled with practiceAuthoritative states and exceptions logged

Work-to-wait profileOnly local handling time knownElapsed time known; categories mixedWork, wait, approval and rework splitDistribution sliced by outcome and class

Bottleneck confidenceBusiest step assumed to constrainQueue visible; cause uncertainFlow evidence supports locationConstraint response tested

Exception concentrationAll cases forced into one pathExceptions known but unclassifiedMaterial classes and owners mappedResidual flow measured by consequence

Rework and failure demandRepeat work counted as fresh demandLoops suspectedRework source and burden measuredRoot-state defect and outcome linked

Evidence loss at handoffsAuthoritative context routinely lostReconstruction observed, not measuredMissing fields and burden mappedProvenance and uncertainty preserved

Ambiguous decision headroomTransitions are deterministic or undefinedAmbiguity assertedBaseline errors expose testable headroomEvidence and action response demonstrated

Downstream capacity and authorityFixed gate; no change authorityCapacity concern unresolvedOwner can change downstream flowConstraint capacity and queue rule secured

Consequence and lawful boundaryConsequential action lacks lawful controlExposure named; recovery incompleteAuthority, appeal and recovery definedBoundary tested with severe exceptions

Choose next routeState defectAmbiguous transitionFixed constraintEmergent work

Assumptions, hard overrides and limits

• Unit: answers cover one repeatable unit of work, entry trigger, observation window and completed outcome.
• Repair route: missing state, integration, ownership or deterministic logic should be fixed before adding probabilistic work.
• AI-test route: one ambiguous transition with evidence, headroom and measurable downstream response can receive a bounded experiment.
• Constraint route: local acceleration stops when downstream capacity or authority cannot change; redesign around the gate.
• Non-workflow route: emergent or unobservable work uses bounded probes and event reconstruction rather than a fabricated linear process.
• Hard overrides: no scale recommendation without a defined unit, completed-outcome owner, measured baseline and lawful boundary for consequential action.
• Limits: the Lens does not prove the bottleneck, calculate capacity, authorise a model, set an autonomy level or replace impact assessment.

Figure 8. Four routes prevent local optimisation from masquerading as value

Opportunity board

“Repair the workflow first” is a valid discovery result. It directs investment to state, integration, policy, ownership or deterministic controls. “Run a bounded AI test” applies only to a named ambiguous transition and measures the downstream action and outcome. “Redesign around the constraint” stops a local feature from feeding a fixed gate. “Do not force a workflow” protects research and one-off work from invented process certainty.

NIST's AI RMF reinforces the need to define context, affected populations, human configuration, monitoring, feedback, appeal and deactivation.10 HM Treasury's Green Book links rationale, options, lifetime effects and evaluation to end outcomes.11 Neither makes a workflow diagram sufficient. The local flow record makes their questions operational.

07 · Negative controls and boundaryJudge the completed outcome

The first negative control is a drafting assistant that cuts handling time by 30 per cent while creating a review queue. Rework and waiting raise total cycle time. The local task result may be valid, but the workflow case fails. The intervention should be redesigned around evidence quality, release authority or the actual constraint.

The second negative control automates intake and raises units entering a fixed specialist bottleneck. Work in progress and ageing increase while completed outcomes stay flat. Higher intake throughput is not value. Limit arrival, improve readiness, change the specialist constraint or alter the service design before accelerating supply.

The counter-case checks evidence completeness at entry. It prevents downstream restarts and removes avoidable handoffs without making the consequential decision. Its model component may be modest; its system value can be large because it changes the timing and quality of flow. Measure late requests, rework, ageing and completed outcomes across the target population.

Emergent work remains the boundary. Research, novel negotiation and crisis response may be reconstructed after events, but forcing stable states can suppress the very adaptation required. Use bounded information probes, decision records and after-action evidence. Return to workflow mapping only when repeatable units and transitions appear.

A maintained workflow lens is a living evidence asset. Policy, volume, staffing, systems and customer behaviour change the constraint. Update event definitions, exception classes and owner authority. A previous AI opportunity can disappear after an integration fix; a new ambiguous transition can emerge after policy changes. Discovery continues with the flow.

08 · ConclusionImprove the river, not one island

The workflow lens shifts attention from visible cognitive tasks to the movement of completed work. It makes queue, state, context, exception, authority, rework and feedback part of the opportunity. That wider view explains how a genuine local speed gain can increase customer delay and operating burden.

Begin with the unit, trigger and completed outcome. Reconstruct authoritative states from event data, observation and exceptions. Separate touch, wait, rework and approval. Locate the constraint with evidence. Preserve context across handoffs. Decide whether to remove, simplify, integrate, make deterministic, augment or test prediction.

The method does not compete with task and decision analysis. It tells the team when downstream coupling makes those units insufficient. When the work is isolated, test the task directly. When it crosses queues and authorities, judge the intervention by completed outcomes and the whole elapsed path.

The executive question is not “Which workflow step can AI perform?” It is “Which state transition, queue or evidence join prevents value from moving, and what intervention will improve the completed outcome without flooding the next gate?” That question turns opportunity discovery from feature placement into operating design.

SourcesReferences

1. Little, J. D. C. (1961). A Proof for the Queuing Formula: L = λW. Operations Research, 9(3), 383–387.
2. Davenport, T. H. and Short, J. E. (1990). The New Industrial Engineering: Information Technology and Business Process Redesign. Sloan Management Review.
3. Hammer, M. (1990). Reengineering Work: Don't Automate, Obliterate. Harvard Business Review.
4. OECD (2025). The Adoption of Artificial Intelligence in Firms. OECD Artificial Intelligence Papers.
5. Bonney, K. et al. (2026). The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks. US Census Bureau, CES Working Paper 26-25.
6. Brynjolfsson, E., Li, D. and Raymond, L. R. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942.
7. Dell'Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier. Organization Science.
8. Cui, Z. et al. (2026). The Effects of Generative AI on High-Skilled Work. Management Science.
9. Kleinberg, J., Ludwig, J., Mullainathan, S. and Obermeyer, Z. (2015). Prediction Policy Problems. American Economic Review, 105(5), 491–495.
10. NIST (2023). Artificial Intelligence Risk Management Framework Core.
11. HM Treasury (2026). The Green Book 2026.
12. Kaplan, R. S. and Anderson, S. R. (2004). Time-Driven Activity-Based Costing. Harvard Business School Working Paper 04-045.
13. OECD (2025). The Effects of Generative AI on Productivity, Innovation and Entrepreneurship. OECD Artificial Intelligence Papers.
14. NIST (2023). AI RMF Manage Playbook.