A claims team bought digital execution at a price that looked negligible beside loaded employee cost. The agent read submissions, drafted assessments and pushed files towards settlement. Draft volume climbed immediately. Settled claims did not. Assessors received more ambiguous exceptions, reopened source documents, corrected tool calls and reconciled duplicate work. The queue aged while model invoices, retrieval charges and retries multiplied. The price of a draft had fallen; the risk-adjusted cost of an accepted claim had risen.

The comparison failed because it treated a run as labour and a draft as output. Neither carried the full production boundary. The business outcome required context, tools, verified evidence, judgement, authority, exception handling, downstream absorption and accountability. Digital execution was abundant inside one step, but scarce human verification governed the whole system's throughput. The relevant economics appeared only after the unit of account moved from activity to accepted outcome.

01 · Unit of accountA cheap run is not a cheap outcome

A digital agent is a production arrangement. Model inference is one input. Retrieval, tool execution, storage, orchestration and retries add variable digital cost. Integration, evaluations, context engineering, permissions, observability and change capability form standing capital. People frame the work, specify boundaries, review ambiguity, resolve exceptions, maintain relationships and accept responsibility. Handoffs, queues and reconciliation impose coordination cost. Rework, delay, escaped error and expected loss adjust nominal output for quality and risk.

These components are non-additive in the managerial sense. A weak context service raises retry and review cost. A narrow evaluation regime allows correlated errors to reach downstream systems. An approval gate changes latency and may shift the bottleneck. The account therefore preserves causal relationships and ranges rather than presenting a neat “agent wage”. Its denominator is an outcome that the business can accept, use and defend.

A claims draft becomes accepted only after the necessary evidence is present, policy is correctly applied, exceptions are resolved, authority is valid and downstream state is consistent. Counting drafts rewards the agent for creating work. Counting accepted claims forces the production account to include the assessor, queue and repair path. The same logic applies to research briefs, reconciliations, customer follow-up, software changes and procurement decisions.

Quality belongs in the definition of the unit rather than in a later discount. If a baseline team settles a claim with complete provenance and the agent route settles one with a missing source, the two outputs are not interchangeable units. Define acceptance tests before observing the new system. Include timeliness, affected state, evidence, control receipt and recovery where they matter. This prevents a delivery team from improving its economics by weakening the output specification after deployment.

Latency needs the same boundary. An agent may finish its step in seconds while the business outcome waits days in review. Measure elapsed time from valid demand to accepted state, then separate active work from queue time. The split identifies whether faster digital execution creates customer value or merely shifts inventory downstream. For time-sensitive work, include the economic consequence of delay rather than reporting speed as a cosmetic service statistic.

02 · Task recompositionSubstitution, complement, reinstatement and rebound

Task economics provides a better lens than job exposure. Autor emphasises that automation both substitutes for and complements labour. Acemoglu and Restrepo distinguish displacement from the creation of new tasks in which labour has a comparative advantage. Technical exposure says that a system could affect some activities. It does not establish adoption, reliable performance, demand, organisational absorption, wage effects or employment change.

Direct substitution occurs when digital execution removes a stable task and the accepted output needs little additional human work. Augmentation occurs when people retain the task but complete it with better information or faster preparation. Skill compression occurs when less-experienced workers improve on supported cases while experts absorb harder exceptions and verification risk. New-task creation makes previously uneconomic work viable. Demand expansion increases volume when a service becomes cheaper, faster or broader.

A cheaper research brief illustrates rebound. Managers who once commissioned ten careful briefs may request many more once drafts arrive quickly. Review, reading and decision time then rise. Cost per draft falls while total spend and management attention increase. This may still create value, but it is a service expansion, not captured labour saving. Give it an explicit demand budget and measure which decisions improve.

New customer follow-up can reinstate human work. An agent identifies cases and prepares context that previously cost too much to assemble. The firm can now contact customers it ignored. Digital preparation creates demand for relationship conversations, exception repair and service design. Finance should record the new revenue or retention hypothesis alongside the extra human capacity it requires.

Task shares should be estimated from observed cases, not from a workshop list of capabilities. Sample complete trajectories and code each meaningful contribution: removed manual execution, faster human work, additional review, new service, exception repair and downstream work. One case can contain several categories. Report a range where classification is uncertain. This makes the recomposition visible without pretending that every minute carries equal economic or organisational weight.

Decision rights also move with tasks. If an agent prepares a recommendation while a person retains authority, labour has been augmented even when most keystrokes disappear. If bounded software executes a deterministic reconciliation under a standing rule, substitution can be genuine. If a reviewer must reconstruct the case and approve each output, execution has moved while responsibility and verification remain human. Costing must follow the actual allocation rather than the interface label.

03 · Original constructThe six-part production account

The account produces several views. Accepted outcomes per constrained human hour shows whether digital execution increases real capacity. Risk-adjusted cost per accepted outcome supports baseline comparison. Marginal cost at current congestion reveals the price of the next unit after queues begin to form. Released capacity, redeployed capacity and captured capacity remain separate because saved minutes have no financial value until work or spending changes.

Task shares show substitution, augmentation and new work without forcing a role-level conclusion. Demand elasticity records how volume responds to lower price, faster delivery or wider service scope. A value-distribution view follows gains to customers, workers, the firm and suppliers. These views can disagree. A service may lower customer waiting time while increasing platform rent and leaving firm cost unchanged.

Standing capital has a time profile. Early evaluation, integration and process redesign can depress measured returns before reusable capability appears, consistent with the complementary-investment logic of the productivity J-curve. Finance should not hide these costs, but it should distinguish one-off creation from recurring operation. Track which artefacts become reusable across cases, which require constant maintenance and which were abandoned. Amortisation should follow observed useful volume and life, not a presentation forecast.

Human complement cost includes learning and recovery. Reviewers need time to understand new failure modes, maintain unassisted skill and investigate incidents. Product owners need capacity to update policies, tests and context when the environment changes. These are not adoption frictions that automatically vanish. Some decline with experience; others recur because models, tools and business rules change. Record them by activity so later scale assumptions can be challenged.

Expected loss should use scenarios, not a false precision. Estimate frequency and severity ranges for escaped errors, correlated actions, delay, remediation and customer harm. Show the result with and without contested assumptions. A low mean can conceal an unacceptable tail. Conversely, a visible but recoverable error should not be costed like an irreversible one. The purpose is to make the economic boundary auditable, not to manufacture certainty.

04 · Marginal economicsThe token illusion and the accepted-output curve

Variable inference prices often fall, but the next accepted outcome may become more expensive. Harder cases trigger longer context, more retrieval, additional tools and retries. Acceptance rates can decline as scope expands beyond the well-specified core. Reviewers face concentrated ambiguity. When their queue saturates, delay and rework compound. Marginal cost must therefore be measured at current mix and congestion, not extrapolated from a pilot average.

The clean substitution case is the necessary counterweight. A reconciliation task has stable inputs, deterministic rules, reversible writes, high acceptance, low exception rates and complete receipts. Downstream systems absorb the output without new queues. Manual execution can be removed safely, and finance can capture capacity through reduced external spend, avoided hiring or an explicit workload change. The production account should approve this case rather than protecting unnecessary work.

The stop case is equally plain. When review, rework, delay, standing capital and expected loss push accepted-output cost above the baseline, continuing because tokens are cheap destroys value. Stop or reprice the service, narrow the scope, repair the evidence boundary or redesign the workflow. Sunk integration effort is not a reason to scale a failing unit.

Retries deserve causal inspection. Some arise from ordinary stochastic variation; others reveal unclear specifications, unstable tools, missing context or a model outside its useful boundary. Treating every retry as cheap digital labour can preserve a defective design. Segment first-pass acceptance, recoverable retry and human rescue. A rising rescue share is a signal that the next unit consumes scarce complement capacity, even if the digital invoice remains small.

Marginal analysis also protects good cases from bad expansion. A bounded reconciliation service may be economic for one ledger and uneconomic when extended to subsidiaries with inconsistent data. Keep cohorts and scope visible. Scale the region where the acceptance and recovery evidence holds, while pricing or rejecting the frontier separately. A blended average can cause management either to over-expand a fragile system or to abandon a useful core.

05 · Binding constraintQueue inversion

An agent can increase local productivity and reduce system throughput. In claims, abundant drafts arrive faster than assessors can verify novel evidence or policy exceptions. The exception share may look modest, yet absolute exception volume rises with total production. Experts receive a harder case mix, take longer per case and become the scarce factor. End-to-end delay increases despite faster generation.

The remedy may be better evidence packaging, narrower scope, redesigned exception categories, more expert capacity, independent automatic checks or a lower arrival rate. Adding general reviewers without relevant competence can increase reconciliation. The production account identifies which intervention lowers accepted-outcome cost rather than merely increasing upstream volume.

Queue disciplines should attach arrival control to downstream capacity. Set a maximum work-in-progress level, prioritise by consequence and age, and slow agent generation when the constrained station approaches its safe envelope. This can feel inefficient because digital capacity sits idle. In a flow system, however, unused upstream capacity is cheaper than excess inventory that ages, obscures exceptions and consumes recovery attention. The economic objective is accepted throughput, not machine utilisation.

Exception reduction must avoid gaming. Narrowing the label or routing difficult cases back to users can make the agent dashboard improve while total work grows elsewhere. Follow every rejected, deferred and abandoned case to its final state. Count repeated customer contact, manual reconstruction and downstream write-offs. The production boundary ends at the accepted business outcome, not at the component's success response.

06 · Response effectsDemand rebound, skill compression and new work

Rebound begins when a cheaper service lowers the threshold for asking. More briefs, analyses, messages or code changes become attractive. Some demand is valuable; some is speculative consumption of attention. Budget requests by their downstream decision or customer value. Expire unused outputs. Charge business units for constrained review where appropriate. A service catalogue should state what the agent produces and what humans must still absorb.

Field and experimental studies offer bounded evidence. Brynjolfsson, Li and Raymond reported heterogeneous productivity effects in customer support, with larger gains among less-experienced workers in their setting. Noy and Zhang studied professional writing tasks. Peng and colleagues studied a specific software-development task. These results justify local experiments and attention to worker learning. They do not supply universal productivity parameters for every role or agent.

Skill compression can improve access to performance while changing expert work. Experts may handle fewer routine cases and more ambiguous exceptions, coaching and quality risk. Their average handling time may rise even as the team's accepted output improves. Capacity plans and reward systems must recognise the harder mix. Otherwise, leaders interpret concentrated complexity as declining expert productivity.

Demand budgets need a decision rule. State the population served, maximum volume, value hypothesis and constrained human capacity. Attribute incremental requests to price, speed, scope or promotion where possible. If users do not consume the output or decisions do not improve, expire the service rather than storing unused production. If new demand creates measurable customer or revenue value, fund the complementary work and describe it as growth.

Worker effects require direct observation. Track task variety, interruption, autonomy, learning, exception burden and contestability alongside output. Interview people doing the changed work and compare their account with telemetry. An apparent productivity gain can coexist with unsustainable work intensity or lost tacit learning. Those conditions eventually reappear as turnover, quality variation or dependency, so they belong in the production case.

07 · DistributionWho captures the gain?

Apparent labour saving can become supplier rent. A workflow may depend on one model, proprietary context store, tool platform or data source. Usage charges rise, switching becomes expensive and evaluation knowledge stays with the supplier. Track concentration, portability, exit cost, price-change rights and the portion of value paid outside the firm. Cheap introductory inference is not a durable cost curve.

Workers may gain better tools, learning and higher-value tasks, or bear monitoring, work intensification and less autonomy. Customers may receive faster service or more automated friction. The firm may capture margin, growth or neither. Distribution is part of the business case because it affects adoption, retention, service quality and bargaining power.

Supplier exposure should be stress-tested. Reprice inference and storage, remove a proprietary data feed, change a model version, and estimate the work required to move prompts, evaluations, context and tool contracts. Portability means more than an API abstraction. The organisation must retain enough knowledge and test evidence to reproduce accepted outcomes on an alternative. If it cannot, the switching liability should appear beside the claimed saving.

Customer value can also be transferred rather than created. Faster automated responses may reduce waiting while increasing the effort needed to contest a mistake. A cheaper claims process may save the firm money but shift evidence collection to customers. Track resolution, repeat contact, complaints and successful appeal, not only first response. A production improvement should survive the perspective of the affected party.

08 · Worked accountClaims before and after the agent

Before deployment, assessors receive a bounded daily volume, review complete files and settle most cases end to end. After deployment, the agent prepares far more files, but acceptance falls at the edge of its capability. Retrieval and tool retries increase. Exceptions concentrate among experts. Downstream fraud and payment teams receive more reconciliation work. The nominal drafting cost falls while settled-case delay and expected loss rise.

The first response is not necessarily to abandon the agent. Narrow it to stable claim types, improve evidence packaging, cap arrivals at verification capacity and redesign exception categories. Compare risk-adjusted accepted cost again. If the bottleneck remains, do not scale. If bounded cases meet quality, latency and control thresholds, capture direct substitution there while using human complements for the jagged frontier.

The claims owner should run two matched cohorts over the same observation window. Reconcile intake mix, accepted settlements, assessor time, fraud referrals, payment corrections, customer contacts, delays and losses. Trace agent trajectories and platform charges. Separate routine claims from novel evidence. This design will not remove every confound, but it will expose whether the apparent gain persists after case mix and downstream work enter the account.

Finance should then classify the result. A cash saving requires reduced spend. Avoided cost requires an approved future expense that no longer occurs. Redeployed capacity requires a measured output elsewhere. Quality value requires an agreed outcome and valuation method. Growth requires incremental demand and margin. Keeping these categories separate prevents a single set of saved minutes from being counted several times.

09 · Interactive instrumentDigital Agent Unit Economics Console

The Console routes a case without inventing headcount equivalence or a maturity score. Its qualitative ranges are prompts for a finance-reconciled experiment. Replace them with observed ranges, named owners and an agreed baseline.

10 · Operating disciplineScale the production system, not the demo

Begin with a baseline cohort and an explicit accepted outcome. Instrument complete agent trajectories and human work. Segment routine and boundary cases. Reconcile operating data with finance rather than multiplying self-reported minutes by salary. Observe demand after the service changes. Follow queues into downstream functions. Test switching and failure recovery before committing the volume that creates dependency.

Baseline choice deserves governance. Use the current feasible process, including its real quality, backlog and loss, rather than an ideal manual procedure nobody follows. Record seasonal effects, case mix and policy changes that could distort comparison. Where the baseline is changing, run scenarios and state the uncertainty. A weak baseline can make any pilot look attractive, while an obsolete one can hide a genuinely better production design.

Run marginal analysis at successive volumes. The average pilot case is usually cleaner than the next case admitted at scale. Watch acceptance, retry depth, review minutes, exception difficulty, queue age, loss and supplier price. A curve that worsens identifies the current capacity frontier. Management can then narrow scope, invest in the complement or stop rather than mistaking more activity for progress.

Observation windows should cover learning and delayed consequences. A short pilot may include unusually attentive reviewers, supplier support and a protected case mix. Later, context changes, rare errors and maintenance arrive. Keep a launch view for fast learning and a longer cohort for economic judgement. Require the capture owner and finance partner to reconcile the same population so operating benefits and costs cannot be selected from different periods.

Capacity is captured only through a named action. Reduce external spend, avoid approved hiring, redeploy people to a measured service backlog, increase accepted volume within demand, or improve quality and loss. Each route has a different financial and workforce meaning. The owner should confirm the action in the same observation window and record any residual work that prevents release. Unallocated saved minutes dissolve into meetings, extra requests or hidden slack; they should not be booked as realised saving.

Redeployment needs an entry and an exit. Name the work people will stop, the work they will start, the competence transition, the receiving manager and the outcome that will evidence value. If old duties remain through precaution, social expectation or unresolved exceptions, capacity was not released. If the new work lacks demand, it was not productively redeployed. This discipline protects workers and finance from an ambiguous promise of future value.

Governance is a production input. Evaluation, monitoring, incident response, authority and evidence make accepted outcomes possible. Treating them as removable overhead understates cost and raises expected loss. Treating them as unlimited ceremony blocks valid substitution. Design the control proportionately and include its observed operating cost in the account.

The account should have a decision cadence. Operations reviews flow, acceptance, queues and exceptions weekly while a service is changing. Product and risk examine failure patterns, boundary cases and control performance. Finance reconciles the observation cohort, supplier charges and capture action monthly or at the agreed commercial cycle. Workforce owners review task mix, competence and redeployment. A quarterly executive view can then decide whether to scale, narrow, invest, reprice or stop using one shared outcome boundary. The cadence matters because costs emerge at different speeds: digital charges appear immediately, downstream reconciliation later, supplier dependency gradually and workforce effects across repeated cycles. A common ledger keeps those signals connected.

The labour economics of digital agents is therefore the economics of recomposed production. Measurement must follow the outcome. Cheap execution can substitute, complement, create or flood. Human work can shrink, move to exceptions, deepen into relationships or reappear as verification. Supplier rents and downstream queues can capture the apparent gain. Leaders should ask which outcome is accepted, which input is now scarce, who receives the value and what action makes the capacity real. They should keep the answer tied to an observed cohort, a credible range, a named owner and a reversible decision at the current scale. Only then does scaling have an economic meaning.

SourcesReferences

1. Coase, R. H. (1937), “The Nature of the Firm”, Economica. doi.org/10.1111/j.1468-0335.1937.tb00002.x.
2. Autor, D. H. (2015), “Why Are There Still So Many Jobs?”, Journal of Economic Perspectives, 29(3), 3–30. doi.org/10.1257/jep.29.3.3.
3. Acemoglu, D. and Restrepo, P. (2019), “Automation and New Tasks”, Journal of Economic Perspectives, 33(2), 3–30. doi.org/10.1257/jep.33.2.3.
4. Acemoglu, D. and Restrepo, P. (2018), “Artificial Intelligence, Automation and Work”, NBER Working Paper 24196. doi.org/10.3386/w24196.
5. Autor, D. H., Levy, F. and Murnane, R. J. (2003), “The Skill Content of Recent Technological Change”, Quarterly Journal of Economics. doi.org/10.1162/003355303322552801.
6. Brynjolfsson, E., Rock, D. and Syverson, C. (2021), “The Productivity J-Curve”, American Economic Journal: Macroeconomics. doi.org/10.1257/mac.20180386.
7. Brynjolfsson, E., Li, D. and Raymond, L. R. (2023), “Generative AI at Work”. arxiv.org/abs/2304.11771.
8. Noy, S. and Zhang, W. (2023), “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence”, Science. doi.org/10.1126/science.adh2586.
9. Peng, S. et al. (2023), “The Impact of AI on Developer Productivity”. arxiv.org/abs/2302.06590.
10. Dell'Acqua, F. et al. (2023), “Navigating the Jagged Technological Frontier”, Harvard Business School Working Paper 24-013. hbs.edu working paper.
11. Eloundou, T. et al. (2023), “GPTs are GPTs”. arxiv.org/abs/2303.10130.
12. Acemoglu, D. (2024), “The Simple Macroeconomics of AI”, NBER Working Paper 32487. doi.org/10.3386/w32487.
13. Kellogg, K. C., Valentine, M. A. and Christin, A. (2020), “Algorithms at Work”, Academy of Management Annals. doi.org/10.5465/annals.2018.0174.
14. NIST (2023), “AI Risk Management Framework Core”. airc.nist.gov AI RMF Core.
15. Autio, C. et al. (2024), Generative Artificial Intelligence Profile. doi.org/10.6028/NIST.AI.600-1.
16. OECD (2023), OECD Employment Outlook 2023. doi.org/10.1787/08785bba-en.
17. ILO (2023), Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality. doi.org/10.54394/FHEM8239.