The drafting pilot cuts an email from twelve minutes to three. Customers still wait eight days because missing evidence, approval availability and decision authority govern the case. The new output simply reaches the queue sooner.

This is the visible-task illusion. A task attracts attention because many people perform it, its labour cost is easy to count or a demonstration makes it look transformable. The system outcome remains unchanged because the task had slack or because another station controls the rate of completed good work. Local minutes saved become additional work in process, not throughput.

A bottleneck intervention begins with a unit of good throughput: an accurately resolved enquiry, a complete file ready for authorised decision, an investigated alert with an accepted disposition, or a fulfilled order meeting control requirements. “Good” prevents defects and rework from being counted as production. The relevant system extends from demand arrival to the enterprise or customer outcome, not from prompt to draft.

The current constraint is the condition whose marginal relief would most improve that system outcome. It may be service capacity, arrival or processing variability, poor first-pass yield, unavailable evidence, scheduled access to a dependency, or policy and decision authority. It can move when demand mix, staffing, rules, data quality or the proposed intervention changes. There may also be several interacting limits rather than one permanent bottleneck.

The paper's practical device is a constraint receipt. It records the observed limiting condition, the evidence and time window, the causal mechanism by which AI would alter it, the next place the constraint is predicted to migrate, the capacity and controls prepared there, the unit of good throughput, and the owner who can rebalance the system. The receipt makes task automation a falsifiable operations claim.

01 · System constraintSeparate observable effort from limiting work

A visible task is an activity people can point to; a constraint is a system condition that governs the attainable rate or flow time of good outcomes. The distinction is causal. If doubling a task's service rate leaves completed-good throughput unchanged, that task was not the active rate constraint under the observed conditions. It may still matter for quality, resilience or experience, but its value case must name that different outcome.

The busiest person is not necessarily the bottleneck. High utilisation may reflect useful flow, hidden batching or excessive rework. The longest task may have abundant parallel capacity. The largest payroll line may contain slack because work arrives unevenly. The most disliked task may be painful without controlling throughput. Diagnose the system through arrivals, service, queue formation, blocking and defects rather than proxies for frustration or cost.

Choose an observation boundary that follows work across organisational ownership. A customer case may wait in a shared mailbox, a document store, an analyst queue and an approver's diary, none of which appear in the application that records touch time. Reconstruct a sample of cases from arrival to accepted outcome. Distinguish active service, waiting for capacity, waiting for information, waiting for authority, blocked system time and customer time. This chronology usually reveals which apparently idle period is actually governed by a recurring dependency.

Little's queuing result relates long-run average number in a system, arrival or throughput rate and average time in the system under stated conditions: L = λW.1 The operational lesson is not that a single snapshot proves a bottleneck. It is that accumulated work in process, flow time and throughput are connected. If faster drafting raises the number waiting for approval while completions do not rise, the system has created inventory.

Spearman's congestion model connects bottleneck rate, raw process time and congestion in a production setting.2 Service workflows have their own structures, but the general caution transfers: cycle time is not simply the sum of touch times. Waiting and contention dominate as demand presses against effective capacity. Map scheduled availability, batching and blocking dependencies, not merely average minutes per task.

02 · Constraint receiptWrite the causal claim before choosing the tool

The receipt begins with the observation window and case mix. Record arrivals by time and class; active and scheduled capacity at major stations; work in process and ageing; first-pass yield, exceptions and rework; blocking dependencies; and the authority that releases a case. A weekly average can conceal a daily surge, a rare but consuming exception or a committee available only twice a week. Segment before concluding.

Next state the limiting mechanism in a sentence that can be wrong. “Underwriter evidence assembly limits completed-good credit decisions for standard small-business renewals during weekday demand because available analyst hours are consumed locating and reconciling authorised sources.” This is more useful than “underwriting is slow”. It names population, outcome, time, resource and mechanism. The team can look for contrary evidence, such as idle underwriter time or a downstream approval queue.

The AI mechanism is equally specific. It may retrieve policy passages with provenance, extract candidate values, reconcile source differences or assemble a reviewable evidence pack. The claim is not that the model is faster. It is that the configured service reduces the scarce work at the observed constraint without lowering first-pass yield or raising downstream review. Include integration, exception routing, human checking, recovery and the cases where the system must not act.

Then predict migration. If evidence assembly releases underwriter capacity, will independent review, committee availability, fulfilment or customer response become limiting? Estimate only enough to make a capacity decision; do not disguise uncertainty as a precise forecast. Name the next station, its present headroom, the safeguard if the queue grows and the owner authorised to rebalance staff, schedule or policy.

Finally specify the enterprise signal. Use completed-good decisions, outcome cycle time, ageing, abandonment, quality, loss or another governed measure tied to the process purpose. Local task minutes remain a mechanism check, not the success criterion. The receipt expires when demand mix, staffing, policy, system configuration, data source or decision authority changes materially.

A receipt should include disconfirming observations. If the suspected station sometimes has no queue while end-to-end delay persists, record those periods and examine another mechanism. If release from that station does not change downstream starts or completions, the claimed causal link is weak. A useful test precommits to what would make the team abandon the AI hypothesis. This protects the programme from interpreting every local improvement as confirmation after the fact.

03 · Queue mechanicsVariability can constrain a system before average capacity does

Average arrival rate below average service capacity is not evidence of healthy flow. Demand arrives in bursts; cases vary in effort; people take breaks, attend meetings and switch tasks; dependencies operate on schedules; and rework returns unpredictably. As effective utilisation becomes high, a small increase in variability can produce a disproportionate increase in waiting. Kingman's heavy-traffic analysis is a foundational account of this sensitivity in a single-server setting.3

Do not transplant a queue formula into an executive slide without checking its assumptions. Real workflows may have priorities, parallel servers, batching, abandonment, correlated arrivals, shared staff and re-entrant work. The robust decision is observational: compare time-series arrivals, starts, completions, WIP, ageing and service availability by case class. Use models to challenge a design and test scenarios, not to manufacture a precise wait forecast from weak inputs.

AI can reduce mean service time while increasing variability. Routine cases may become nearly instant, while ambiguous cases require long review and context recovery. AI can also increase arrivals by making it easier to submit work, or create correlated bursts when upstream automation releases batches. Measure the distribution and exception tail, not just the average generated output time.

Quality is a queue mechanism. If first-pass yield falls, nominal completions recycle into investigation, correction and approval. This consumes constraint capacity and obscures the real arrival rate. Count a case as good throughput only after its required control and outcome are accepted. Record rework loops separately so that a detector producing more alerts or a generator producing more drafts cannot appear productive by expanding defective inventory.

Case mix must remain visible as well. A station can appear to gain capacity because easy cases increase, while complex-case waiting deteriorates. Segment by the attributes that drive service, risk and outcome, then compare like with like. Report both throughput and ageing for important classes. Where priority rules allow urgent work to jump the queue, check whether ordinary work accumulates invisibly. An intervention that improves the headline average by starving a vulnerable or complex population has not improved the complete system.

04 · Constraint migrationLocal acceleration can move inventory without improving flow

The email decoy shows one migration pattern. Drafting accelerates, approvals do not. The queue before approval grows, reviewers face more context switching and ageing cases become harder to distinguish. Customer cycle time may worsen even though local touch time improves. The right action is to remove missing-evidence and authority constraints, rebalance review, or keep drafting assistance as an experience intervention with a separate value claim.

Faster alert production is a more damaging negative control. A detector raises alert recall or produces a larger candidate set while investigator capacity, false-positive review and escalation remain limiting. More alerts can lower the fraction investigated within the required window, crowd out high-risk cases and raise rework. The visible model metric improves as good investigative throughput deteriorates.

Before an intervention, draw the expected after-state. Show arrival into every affected station, effective service, quality loss, WIP guardrail and authority. If upstream output will exceed downstream absorption, decide in advance whether to throttle, prioritise, add capacity, change the operating schedule or reject the intervention. “We will monitor it” is inadequate unless an indicator, threshold, owner and response are named.

The strongest absorption test is controlled release. Run the intervention on a bounded population, preserve a comparable baseline and watch downstream queue age, first-pass yield, rework, abandonment and completed-good outcomes. Stop or throttle when the next station breaches its accepted boundary. This separates a real system improvement from a demo that merely shifts work.

05 · Worked counter-caseRelieve evidence assembly and prepare the next control

A commercial-credit renewal process receives applications, gathers evidence, assembles an analyst file, reaches an authorised underwriting judgement, passes independent control and enters fulfilment. Observation shows that standard renewals wait primarily for analysts to locate and reconcile financial documents, policy clauses and prior decisions. Underwriters have available decision slots but cannot begin on incomplete files. Control and fulfilment have measured headroom for a bounded release.

The constraint receipt identifies evidence assembly for that population and time window. The proposed AI service retrieves authorised sources, extracts candidate values with locations, flags inconsistencies and drafts a structured evidence pack. It does not decide credit. Analysts verify critical fields; unsupported items remain missing rather than being invented. The baseline is accepted files per analyst hour, time from complete intake to decision-ready file, first-pass control acceptance, rework and completed-good decisions.

A bounded test samples representative entity types, document quality, languages, financial periods and exception classes. It measures the configured workflow, including review and recovery, rather than model latency. The test stops if critical-source omissions, unsupported claims, rework or downstream queue age breach their accepted boundaries. Underwriting and control owners observe arrivals and can throttle the service.

The intervention qualifies only if evidence work was genuinely limiting, the service removes that scarce work, quality remains within boundary and downstream stations absorb the release. The constraint may migrate to independent control. That migration is not failure if predicted and governed. The organisation rebalances control capacity, changes scheduling or limits release so that completed-good decisions and cycle time improve together.

Release can be governed through a daily quota, case-class eligibility, queue-age trigger or service-health condition. The control owner receives forecast and observed arrivals before expansion. If independent review reaches its WIP guardrail, the system narrows eligibility or returns to the approved manual intake rather than continuing to push. This makes downstream capacity part of the intervention design. It also creates evidence about how much of the theoretical task-time saving survives contact with the whole regulated workflow.

This counter-case differs from automating a visible task. It begins with system evidence, preserves decision authority, measures good output and prepares the next station. It also remains conditional. A change in renewal mix, policy, staffing, data availability or committee schedule can invalidate the receipt and move the constraint elsewhere.

06 · Strongest rivalNon-bottleneck automation can still create value

The strongest rival says a high-volume task can deserve automation even when it is not today's constraint. It may reduce defects, improve employee experience, create resilience, lower harm or preserve option value before demand rises. A tedious task with abundant capacity can still cause fatigue. A safety check can be valuable without increasing throughput. Preventing a future constraint can be cheaper than reacting to one.

This rival is right. The rejection applies only to the unsupported inference that local task speed equals enterprise throughput or economic value. If quality is the outcome, define the defect and affected population. If experience is the outcome, measure burden and behavioural response. If resilience is the outcome, test degraded operation and recovery. If option value is claimed, name the future state, observation and staged commitment.

Task-level evidence remains useful within its scope. Noy and Zhang report experimental results for professional writing tasks in their study setting.7 Brynjolfsson, Li and Raymond report field evidence from customer support, including heterogeneous effects across workers.6 Neither result establishes that an arbitrary enterprise workflow's bottleneck will move. A system case must connect measured task change to local arrivals, queues, dependencies and good outcomes.

Acemoglu and Restrepo use a task-based economic framework that distinguishes displacement, productivity and the creation of new tasks.4 The operational implication is that work does not disappear uniformly when one task changes. Review, exception handling, coordination and new control work can arise. Count the residual workflow and examine where scarce capacity moves.

Complementary investment can also delay visible returns. The productivity J-curve argument highlights intangible investment accompanying general-purpose technologies.5 OECD work similarly discusses organisational, skills, data and process complements in AI adoption.10 These sources support patience where a causal operating plan exists. They do not excuse a task-speed pilot with no route to an enterprise outcome.

The investment case should therefore separate immediate throughput, learning and prevention. Immediate throughput requires an observed active constraint and absorptive downstream design. Learning requires a decision that the test will inform, with the value of information linked to a later commitment. Prevention requires evidence that demand, capacity or policy is moving towards a future limit and that early action is cheaper or safer. These cases can coexist, but each needs its own measure and owner. Combining them into one productivity percentage makes review impossible.

Five constraint types, five different responses

A capacity constraint requires evidence that a station's effective service rate is below relevant demand and that AI changes scarce work. A variability constraint may require smoothing, segmentation, staffing or better exception routing rather than faster averages. A quality constraint requires higher first-pass yield and less rework. A policy constraint needs a decision by the authorised owner, not a model. A dependency constraint may require availability, access, data repair or schedule redesign.

Constraints interact. Poor input quality can create variable service and consume capacity; scheduled authority can cause batching; an AI detector can increase arrival variability at investigation. The operating team should identify the dominant mechanism for the bounded population and preserve competing hypotheses. If observations cannot distinguish them, the correct action is to observe before intervening.

07 · Executable instrumentChoose the system action, not the most attractive task

The Constraint Targeting Console is a triage mechanism. Enter one output population and common unit of rate. Supply representative arrivals, variability, effective capacity and WIP for four major stations. Add quality, rework, dependencies, authority, proposed AI effects and downstream absorption. The console identifies the dominant constraint hypothesis, predicts migration and returns one action with missing evidence.

The calculation is deliberately conservative. Capacity and queue observations can nominate a station, but poor yield can make quality dominant, a blocking dependency can supersede capacity and a hard authority condition cannot be automated. The result does not optimise a complex queueing network or prescribe staffing. It makes assumptions visible so an operations team can decide what to observe or test next.

08 · Operating cadenceObserve, test, rebalance, measure and re-map

Start with a stable unit of good throughput and a representative observation window. Review arrivals, completions, WIP age, capacity availability, first-pass yield, rework and blocking time by case class. Walk cases across handoffs to find informal waits and workarounds. Agree the constraint hypothesis with the end-to-end owner and the owners of the affected controls.

Design the smallest test that could disprove the AI mechanism. Preserve a baseline or comparison, bound the decision population, and measure the configured service including human review and recovery. Predefine the next station, its absorption boundary and the throttle. Track task measures to diagnose the mechanism and enterprise measures to decide whether the system improved.

Rebalance when the constraint migrates. Capacity may move through scheduling, staffing, work design, prioritisation or policy, not only headcount. A queue guardrail should trigger a named response. If quality or risk deteriorates, stop and restore the fallback. If good throughput and outcome time improve without unacceptable harm, expand by another bounded slice and remap.

Hold a short constraint review at a cadence matched to volatility. The board needs the observation window, current population, leading constraint hypothesis, confidence and contrary evidence, active experiments, WIP guardrails and rebalancing decisions. It should remove stale receipts when conditions change. This is not a performance meeting for maximising every station's utilisation. Some protective capacity and slack are deliberate: they absorb variation, preserve recovery and prevent one local optimiser from destabilising the flow.

Production readiness matters because a model result can create a new operational constraint. Breck and colleagues emphasise testing and monitoring needs beyond an offline model result.9 Sculley and colleagues warn about entanglement, feedback loops and hidden dependencies in machine-learning systems.8 These mechanisms can change arrivals, failure correlation and recovery work, so they belong in the constraint map.

NIST's AI RMF Core asks organisations to map context, purposes, tasks, benefits and risks, measure and manage them, and make go or no-go decisions.12 The framework does not identify a workflow bottleneck. It supports the governance around the local causal test: accountability, risk tolerance, monitoring, response and retirement. ISO/IEC 42001:2023 likewise provides management-system requirements and continual-improvement context, not proof that a task automation creates throughput.14

09 · ConclusionImprove the flow of accepted outcomes

A task is not strategic because it is visible. The constraint is found in the end-to-end relationship among demand, capacity, variability, quality, dependencies and authority. Relieving it creates value only when the configured intervention changes scarce work, protects first-pass yield and the next station can absorb the release.

That standard keeps the portfolio honest: it rewards accepted outcomes and controlled flow, while exposing faster local activity that merely relocates delay, risk or unfinished work.

The constraint receipt makes that claim inspectable. It records the population, observation window, mechanism, evidence, AI effect, predicted migration, quality guard, downstream response, enterprise measure, owner and expiry. Negative evidence is useful: it can redirect attention from generation to evidence, from modelling to policy, or from upstream speed to downstream capacity.

This method leaves room for quality, safety, experience, resilience and option value. It simply requires those outcomes to be named and measured rather than smuggled into a throughput case. A non-bottleneck intervention can be good; an unsupported claim that local speed equals enterprise value cannot.

The executive question is: “Which condition currently limits completed-good outcomes for this population, how will the intervention change that condition, and where will the constraint move?” Answer it before approving the automation. Then observe, test, rebalance and ask again.

SourcesReferences

1. Little, J. D. C. (1961). A Proof for the Queuing Formula: L = lambda W. Operations Research, 9(3), 383–387.
2. Spearman, M. L. (1991). An Analytic Congestion Model for Closed Production Systems with IFR Processing Times. Management Science, 37(8), 1015–1029.
3. Kingman, J. F. C. (1961). The Single Server Queue in Heavy Traffic. Mathematical Proceedings of the Cambridge Philosophical Society, 57(4), 902–904.
4. Acemoglu, D. and Restrepo, P. (2018). Artificial Intelligence, Automation and Work. NBER Working Paper 24196.
5. Brynjolfsson, E., Rock, D. and Syverson, C. (2020). The Productivity J-Curve. NBER Working Paper 28233.
6. Brynjolfsson, E., Li, D. and Raymond, L. R. (2023). Generative AI at Work. NBER Working Paper 31161.
7. Noy, S. and Zhang, W. (2023). Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence. Science, 381(6654), 187–192.
8. Sculley, D. et al. (2015). Hidden Technical Debt in Machine Learning Systems. NeurIPS.
9. Breck, E. et al. (2017). The ML Test Score: A Rubric for ML Production Readiness and Technical Debt Reduction. IEEE Big Data.
10. OECD (2025). The Adoption of Artificial Intelligence in Firms. OECD Artificial Intelligence Papers.
11. OECD (2024). The impact of Artificial Intelligence on productivity, distribution and growth. OECD Artificial Intelligence Papers.
12. NIST (2023). AI Risk Management Framework Core.
13. NIST (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
14. ISO (2023). ISO/IEC 42001:2023, Artificial intelligence management systems.