Part I · The bottleneck after accelerationFaster preparation can create a slower operation

Imagine a credit-assessment agent that cuts evidence assembly from forty minutes to twelve. The bank responds by accepting three times as many applications. Every recommendation still needs one of four authorised people to commit the decision.

The model looks productive when measured at its own step. The operation does not. Prepared files accumulate before the commitment gate, evidence ages, applicants call for updates and authorised staff alternate between hurried approvals and weekend clearance. Adding another extraction agent makes the backlog worse. The capability removed one constraint and exposed another, while cheaper preparation expanded demand. The relevant economic object is therefore not the model task. It is the complete flow from eligible request to reconciled business outcome.

The central claim is that human-agent productivity must be governed as a queue system in which automation changes arrivals, service demand and routing at the same time. If machine throughput grows faster than accountable human service, delay has moved rather than disappeared. The design decision changes from buying faster generation to controlling admission, creating work-ready tokens, separating service types, protecting scarce capacity and stopping demand before oversight becomes nominal.

This is narrower than bottleneck diagnosis and broader than staffing. It asks what arrives, which contribution remains, who can legally and competently provide it, how long consequences can wait, and what proves completion. It also refuses to model people as interchangeable processors. A relationship specialist, an accredited lawyer and a payment authority may all be available, yet only one can serve a particular token. Fatigue, interruption, recovery, learning and professional challenge are part of production design, not personal inefficiency.

Part II · The service contractA queue should receive a contribution-ready token

An inbox says that something is waiting. A Work Token Contract, introduced here as part of the Human-Agent Queue Operating Model, says what service is required and why this recipient can provide it.

The token links a case and workflow identifier to the business outcome, eligible population and authoritative state. It carries the evidence manifest, provenance, freshness and omissions; agent work already performed; unresolved conflicts; consequence and reversibility; affected party; required skill, accreditation, language, relationship context and authority; deadline, evidence expiry and consequence clock; permitted actions; prior side effects; receiving owner; and closure proof. A recipient can accept or reject it for a declared reason. That acknowledgement starts the service clock and prevents work from vanishing between teams.

A customer-support drafting agent illustrates the difference. Sending a fluent reply and a ticket number to a reviewer saves little if the person must search six systems to reconstruct the claim. A proper token supplies the source passages, customer history, product state, policy version, unsupported statements, permitted remedy and communication boundary. The agent has then changed the work, not merely generated prose. If reconstruction remains, its time belongs in the agent-assisted route.

Part III · Evidence before optimisationCount arrivals, time and work in the same system

The Queue Equation Ledger keeps a small piece of arithmetic honest by defining the boundary around every number.

Little’s long-run relationship connects average work in a system, effective arrival rate and average time in that system under stated conditions.[1] Written compactly, it is useful:

It is an accounting relationship, not a staffing prescription and not a promise that an unstable launch has reached steady state. The arrival definition must say whether rejected, duplicate, abandoned and retried cases count. Work in process must include hidden email, agent rework and tokens awaiting evidence if they remain inside the service promise. Time must begin at the event that matters to the affected party, not at whichever queue entry makes the metric flattering. When definitions change, the comparison breaks.

Averages also conceal the mechanism that matters near saturation. Heavy-traffic analysis shows why variability becomes consequential as service demand approaches capacity, but its approximations should not be transplanted into a heterogeneous production system as a precise forecast.[2] Report service-time distributions by token type, the oldest unresolved case, evidence-expiry risk, rework and second review. A short median wait can coexist with an unacceptable tail.

Part IV · Capacity and demandConvert hours into protected service, then test arrival expansion

The Service Capacity Conversion begins with scheduled time and progressively removes what cannot safely serve the queue.

Subtract breaks, training, calibration, meetings, case conferences and protected recovery. Account for interruptions, handoffs and context switching. Estimate service-time distributions by token class, then include clarification, rework and second-look demand. Reserve headroom for novelty, vulnerable customers, outages and policy changes. Agent assistance changes the composition of the job, so confirm its effect through completed outcomes and voluntary workload evidence. A saved twelve-minute interval becomes capacity only when the work is reorganised and that interval can be used.

This matters when a legal review queue has one accredited authority. Adding three generalists does not increase the rate of reserved decisions. Capacity changes only if preparatory tasks move safely, another accredited person joins, authority is lawfully delegated, or demand is narrowed. Skill-based queue research models compatibility and abandonment under explicit assumptions; it supports making these edges visible, not treating every person as exchangeable.[6] Work on two-tier service systems likewise shows that referral, waiting and mistreatment costs interact under stylised conditions.[3]

The companion Arrival Expansion Test challenges a fixed-demand business case. Cheaper generation can create more requests, candidates and variants. Broader detection can surface previously invisible work. Agents can turn embedded uncertainty into explicit escalations. Parallel option generation multiplies comparison demand. Retries and missing correlation produce phantom arrivals. Easier access can attract a new population. The test separates useful new demand from duplication, avoidable rework and options nobody has capacity to compare.

Part V · Contribution and authorityRoute to a service that can change the outcome

The Skill-Authority Compatibility Graph separates capability, permission and availability. Routing is valid only where all required edges exist.

A procurement team can use several agents to propose suppliers in parallel. Each extra candidate is cheap to generate but costly to compare, reconcile and approve. Candidate admission should therefore be capped by the marginal value of another option and the available comparison service. A generalist can check completeness; a category specialist can interpret trade-offs; a delegated authority can commit. Sending the token to whoever is free creates movement without service.

The Review Value Test asks whether the human can add primary evidence, domain interpretation, independent validation, authority, affected-party voice, challenge, novelty recognition, repair or learning. Research on analysing and prioritising customer information shows that information collection itself consumes capacity and can affect later service.[4] Research on human-in-the-loop fidelity models a trade-off among service quality, queue length, time and a simplified cognitive state, but its homogeneous-task assumptions limit direct generalisation.[5] Both point to the same practical discipline: price the contribution, not just the click.

Reviewers who approve almost every output within seconds might be seeing excellent work. They might also lack evidence, time or authority to disagree. Approval rate cannot distinguish the explanations. Sample review episodes, observe whether the person inspects independent evidence, record meaningful changes and test outcomes. If review cannot change anything, remove the claimed control, redesign the commitment or stop the workflow.

Part VI · Priority and overloadProtect consequence clocks, customers and people

Priority is a governed allocation of waiting harm. It is neither model confidence nor customer revenue disguised as urgency.

The Priority and Fairness Charter orders work using consequence of delay, statutory or evidence deadline, vulnerability, reversibility, starvation age, blocked dependencies and incident relevance. It states when continuity outweighs pooled efficiency and reserves capacity for rare but severe cases. The classifier itself has a cost and error rate. Measure false urgency, missed urgency, triage time and distributional effects, then provide an appeal route. Queueing research on class prioritisation illustrates that performance and discriminatory effects can move together under model assumptions.[7]

A vulnerable customer is a decisive case. Direct access to a named specialist may look slower or less utilised than one pooled inbox, yet relationship continuity is part of the service. Breaking it imposes reconstruction, distrust and repeated disclosure. Preserve the route and judge it by timely, dignified resolution. The small-company negative control points the other way: low volumes may justify one shared queue and one named owner rather than a routing platform. It still needs consequence clocks, a capacity limit and a safe overload decision.

When demand crosses the safe envelope, the Overload Protection Ladder changes the system rather than blaming servers. First remove invalid and duplicate arrivals. Improve token quality and move safe deterministic checks out of human service. Protect focused work, reroute only across valid compatibility edges and activate trained surge capacity. Then defer work with tolerant clocks, narrow the service promise, shed low-value demand with transparent alternatives, suspend automated generation, and activate incident or executive risk control.

Engineering guidance on queue-based load levelling recommends buffering bursts while monitoring depth, bounding downstream consumption and handling failed work; it also warns that sustained production above consumption still requires throttling or shedding.[12] Google’s SRE guidance explains retry amplification, backoff, jitter and load shedding as controls against cascading failure.[13] These are engineering analogies, not permission to treat human service as elastic infrastructure. A person cannot be autoscaled, and recovery time is not idle waste.

The Human Capacity Protection Contract declares concurrent-token limits, focus blocks, recovery windows, minimum incoming evidence, the right to reject unsafe work, training, rotation through ordinary operations, independent challenge, consultation and privacy. NASA’s Task Load Index is a multidimensional subjective workload instrument covering mental, physical and temporal demand, performance, effort and frustration; it should support aggregate operational inquiry, never become an individual productivity rank.[8] Meta-analytic evidence on levels of automation distinguishes routine performance, failure performance, workload and situation awareness.[9] Bainbridge’s older warning remains material: people left with abnormal work may lose the practice and information needed to recover it.[10]

Part VII · Operating casesSixteen queues expose different economic mistakes

These compact cases use the same discipline: define the eligible population, issue a token, identify the service, observe the tail, protect people and decide what changes under overload.

Part VIII · Practical instrumentDesign the queue before expanding the agent

The Human-Agent Queue Economics Designer is an operational triage instrument. It does not calculate an optimal headcount or infer an employee’s productivity.

Start with a real eligible population and consequence. Enter whether demand expands after automation, whether duplicates or rework are material, which contribution remains, whether the token is ready, whether service capacity includes headroom, and whether the recipient has compatible skill and authority. The result recommends one of five routes and exposes the assumptions that require local evidence. Exercise ordinary, peak, outage and policy-change conditions before authorisation.

A positive result permits the named route only for the declared population and conditions. A negative result identifies the missing control: demand, token quality, compatible authority, capacity or protection. The instrument cannot estimate causal productivity from synthetic choices, determine legal compliance, observe cognitive state or replace workforce consultation. Its purpose is to make a consequential operating hypothesis testable.

Part IX · Rival and failure boundaryQueue arithmetic is a warning instrument, not a theory of people

The strongest rival says formal queue design simplifies the work it claims to improve.

Human service times are not stationary. Priorities can be political, data can be censored and the act of measuring a queue can change behaviour. Pooling and high utilisation may be efficient in some settings. Local teams coordinate through relationships that a token cannot fully represent. Added instrumentation can produce false precision, weaken discretion and enable surveillance. For rare work, an informal conversation among competent people may outperform a registry, graph and optimisation model.

The rival is right whenever the model is treated as reality. The answer is proportional evidence. Use ranges and scenarios where volume is low or conditions change quickly. Observe work and interview those who serve and receive it. Preserve direct access when relationship continuity is the service. Keep qualitative consequence and fairness beside queue measures. Retain simple manual coordination for the small-team negative control. A graph should reveal a missing skill or single authority, not assign a numerical worth to a person.

Current governance sources support this broader operating context without dictating a queue design. NIST AI RMF 1.0 is voluntary and under revision in 2026; its Core calls for context, defined roles, human oversight, monitoring, incident response, feedback and change management.[11] The UK Government AI Playbook applies accountable ownership and meaningful human control in its public-sector setting.[16] The EU AI Act has jurisdiction-specific and phased obligations for covered systems.[17] ISO/IEC 42001:2023 frames an organisational management system around objectives, resources, competence, operation, evaluation and improvement.[18]

Worker protection also requires evidence beyond flow telemetry. OECD’s 2025 report describes a survey of more than 6,000 firms across six countries and reports employer concerns including accountability, intelligibility and worker health, but its findings remain tied to those respondents and countries.[14] ILO and European Commission Joint Research Centre case studies in logistics and health care examine work organisation, efficiency, job quality and intrusive surveillance concerns in their studied settings.[15] Neither source licenses hidden monitoring. The Queue Operations Panel should show demand, tail, skill, authority, headroom, rework, outcome and fairness at the level needed to manage the service, with purpose, access, retention and worker consultation declared.

The Human-Agent Queue Assurance Docket records the outcome, eligible population, token, arrival counterfactual, service distributions, capacity conversion, compatibility edges, priority charter, Review Value Test, ordinary and surge scenarios, overload controls, worker and customer protections, full cost, owner and residual risk. The Queue Reauthorisation Rule reopens the decision after a material change to demand, population, model, prompt, evidence, routing, authority, staffing, deadline, fairness, overload action, law or affected group, and after hidden work, expiry, unsafe shedding, worker complaint or common-mode error.

ConclusionOptimise the verified outcome, not the accelerated step

An agent can prepare faster, detect more, generate alternatives and make a service easier to request. Each improvement can also enlarge the queue that scarce people must serve. The responsible response is not to reject automation or demand faster reviewing. It is to declare which work deserves admission, package it so the next contribution is ready, route it through compatible skill and authority, convert scheduled time into protected capacity, preserve consequence-sensitive fairness and reduce arrivals before overload destroys review value.

The decisive metric is controlled flow to a reconciled outcome. It includes the demand automation creates, the rework it transfers, the tail an average conceals, the headroom recovery requires and the human contribution that makes the result legitimate. When those terms are visible, leaders can choose straight-through control, agent preparation, specialist service, protected expansion or a stop. When they are hidden, a faster model can finance a slower and more dangerous operation.