A workshop opens with a familiar matrix. Humans are described as empathetic, creative and contextual. AI is described as fast, consistent and scalable. The room then tries to allocate a real customer complaint. The case needs account retrieval, policy calculation, interpretation of a distressed customer, a remedy decision, an explanation and follow-up learning. The matrix cannot say who should do any of them, because its categories describe actors in general while the decision depends on this population, evidence, worker, model, consequence and moment.

Whole-job allocation is the wrong level. Jobs combine contributions with different information and failure boundaries: observe, retrieve, calculate, interpret, propose, challenge, decide, act, explain, learn and recover. A tested service may calculate a refund limit; an agent may retrieve prior contacts and draft options; a person may interpret context and accept an exception; a bounded agent may execute an approved low-risk remedy. One outcome can legitimately use several allocation modes.

Labour economics has long examined how technology changes task content rather than simply replacing occupations. Autor, Levy and Murnane distinguished routine from non-routine task demands, while Acemoglu and Restrepo described displacement alongside the creation or reinstatement of tasks.12 Evidence on industrial robots also warns that technological gains and labour-market effects cannot be inferred from a feature list alone.3 For human-agent teams, the practical implication is to decompose the outcome and test contribution-level allocation with actual operating evidence.

01 / Break the job apartA complete outcome is a chain of unlike contributions

A work contribution is a bounded input to a downstream outcome. It has evidence, an acceptance test and a consequence if wrong. “Handle support” is too broad. “Retrieve the customer's last three verified interactions and label unresolved promises” is a contribution. “Choose whether policy permits a discretionary remedy and accept the financial and customer consequence” is another. The first may favour an agent with authorised retrieval tools. The second may remain human-owned.

Decomposition should follow changes in information, method, authority and reversibility. Retrieval ends where interpretation begins. Calculation ends where a policy exception begins. Proposal ends where action authority begins. Explanation is not an afterthought if it affects understanding, contestability or future behaviour. Recovery is a contribution because someone or something must detect a defective outcome, contain it and restore service.

Decomposition also exposes interaction cost. If an agent drafts a response but a person must reconstruct the evidence, correct the policy and rewrite the tone, preparation may add rather than save work. If a deterministic calculation supplies a tested result directly to the decision maker, a review step may be waste. The team contract therefore evaluates both the contribution and the hand-off it creates.

02 / Local advantageCapability is a field, not a permanent trait

Comparative advantage asks who should perform a contribution after opportunity cost is considered. The actor with the highest standalone accuracy does not automatically receive every case. An expert may be uniquely valuable on novel disputes, so assigning that expert routine retrieval carries a large opportunity cost even if the expert performs it perfectly. A calibrated agent may handle routine evidence preparation, freeing expert attention for cases where context and consequence matter.

The answer is local. It varies by case type, worker cohort, model version, source availability, intervention time and downstream cost. Field evidence makes the heterogeneity concrete. Brynjolfsson, Li and Raymond studied a generative assistant in customer support and reported larger gains for novice and lower-skilled workers than for experienced and highly skilled workers, alongside evidence consistent with learning.4 That does not establish a universal novice route. It establishes the need to estimate effects by worker and case population.

Technology also changes the field. Dillon and colleagues found individual time savings in a broad workplace experiment without detecting a corresponding change in task composition.5 Dell'Acqua and colleagues found that generative AI altered performance and the expression of functional expertise in a team experiment.6 These studies answer different questions, but together they caution against equating tool use with workflow redesign. Local gains may coexist with unchanged coordination or a newly shifted bottleneck.

03 / The strong rivalFull automation is sometimes the better team design

Human review is not automatically safer. It adds latency and cost, can create automation bias and may fail when reviewers lack evidence, time or stop authority. A stable calculation with controlled inputs, tested rules and reversible low-consequence output has no defensible ceremonial approval step. Deterministic execution with exception handling is the stronger design.

Full agent execution can also be warranted inside a permission envelope. The case population must be bounded, evidence and policy explicit, performance demonstrated on representative cases, confidence calibrated, actions reversible or low consequence, and exception routes operational. Supervision then occurs through monitored exceptions, sampling and stop controls rather than routine clicks on every case.

Preserve human authority when relevant information cannot be reduced to the tested envelope, legitimacy and care affect the outcome, negotiation changes preferences, or consequence is broad and difficult to reverse. The decision is not a tribute to human uniqueness. It is an evidence boundary. When the boundary later becomes testable and safe, allocation can change.

04 / Seven fieldsThe comparative-advantage team contract

For every contribution, the team needs a compact contract. The first field names the contribution and downstream outcome. It prevents local speed from becoming the objective when the true result is resolution, safety, customer understanding or retained capability. The second records information held uniquely by the person, agent or system. Complementarity potential often comes from information asymmetry, not from two actors seeing the same evidence.8

The third field contains the capability and error profile by case type. It separates routine from novel, novice from expert and current model version from prior evidence. The fourth defines authority and consequence: who proposes, decides, acts, stops and accepts the result. The fifth prices interaction and review cost, including waiting, reconstruction and duplicated attention.

The sixth records learning, deskilling and fallback effects. Assistance may transfer expert practice to novices, but broad drafting may erode independent judgement. A team that performs well only while the tool is available has a resilience liability. The seventh sets the routing rule, acceptance test and reassignment trigger. Model drift, changed case mix, rising rework, expert degradation or failed outage exercises must be able to reopen allocation.

05 / Realised complementarityBeat the strongest feasible component, not an average

A team is complementary only if it outperforms the strongest feasible component baseline on the chosen outcome and within the risk boundary. The baseline may be the best human cohort, the current agent, a deterministic service or a routed policy that already sends different cases to different actors. Beating the average employee proves little if the agent alone or a simpler route performs better.

Hemmer and colleagues formalise complementary team performance as performance that neither component can attain individually and identify information and capability asymmetry as potential sources.7 Work on complementary policies likewise treats routing as a decision problem rather than assuming that fixed collaboration improves every case.9 User expertise and system tuning can change joint performance, reinforcing the need for cohort-specific evidence.10

The measurement window must reach downstream. Faster drafting is not a gain if correction queues, complaints, refunds or decision delay rise elsewhere. Benefit capture matters too. If saved handling time becomes unattainable workload expectations while workers absorb more exception stress, the operating claim is incomplete. The team contract names who receives the benefit and who carries repair.

An evaluation protocol for the allocation

The unit of evaluation is a contribution performed for a defined case population by a defined actor or route. Record the model and tool version, worker cohort, source environment, policy version, consequence class and downstream outcome. This makes results portable only where their conditions match. A gain found in routine billing questions does not establish the route for vulnerability complaints, and a result for new starters does not establish the effect on experienced specialists.

Construct baselines before introducing the proposed team. The human baseline should represent the feasible cohort and operating conditions, including its current routing and support. The agent baseline should include the deterministic tools it would genuinely receive in production. The service baseline should cover complete stable rules. A routed baseline should use any existing risk, case or skill segmentation that could deliver the outcome more simply. Otherwise the experiment can credit collaboration for an advantage already available through basic allocation.

Define an outcome vector rather than one productivity measure. It can contain resolution quality, decision error, harm, reversals, completion time, interaction cost, worker workload, customer understanding, downstream correction and retained unaided capability. The principal chooses the acceptance boundary for each measure before reviewing results. This prevents a large local speed gain from hiding a smaller but unacceptable increase in consequential error.

Evaluate heterogeneity directly. Break results out by case type, worker experience, source quality and consequence. Examine where the person corrects the agent, where the agent corrects the person, where both fail and where the route sends the case to the wrong actor. Complementarity potential exists when the actors possess different useful information or capability. Realised complementarity additionally requires the interface, timing and incentives to exploit that difference.

Calibration is a property of predictions on a defined population, not a confidence label displayed beside text. If confidence is used for routing, compare stated confidence with observed outcomes for each route and consequence band. Test distribution shifts and rare cases. A high-confidence answer should never bypass a protected authority boundary merely because the interface makes uncertainty appear numerical.

Finally, price the interaction. Count time spent interpreting suggestions, opening sources, correcting drafts, resolving disagreements, recovering failures and maintaining routing rules. Include attention displaced from other work. A contribution may have positive gross assistance value but negative net team value after review. The allocation clause should report both so leaders do not mistake hidden human effort for machine productivity.

06 / Failure controlsFour ways a plausible team becomes weaker

Generic strengths and weak baselines

A generic matrix routes emotionally worded cases to people and data-heavy cases to agents without population evidence. It ignores calibrated performance, case rarity and opportunity cost. A related evaluation compares assisted workers with average unassisted workers. If the current agent-only route or the best feasible routed policy is stronger, the apparent team gain is a weak-baseline artefact. The repair is stratified evaluation against every feasible component and route.

Ceremonial review

A person approves hundreds of cases after seeing only a polished suggestion. The review queue rewards speed, sources are hidden and the reviewer cannot stop execution. Parasuraman, Sheridan and Wickens described multiple levels of interaction with automation rather than one generic role for the person.11 Meaningful review needs sufficient information, intervention time, manageable workload, authority and feedback. Without those conditions, remove the ritual and redesign the decision boundary.

Expert suppression

A support assistant based on historical best practice helps novices but nudges experts towards conventional replies. Average performance rises while the highest-performing cohort slightly declines and novel solutions disappear. The allocation should differ by cohort. Novices may receive suggestions with learning checks. Experts may receive retrieval and evidence tools without prescriptive text. Novel cases should preserve expert discretion and feed back into the knowledge base only after review.

Downstream transfer and correlated error

Draft production accelerates, but policy correction and complaint handling rise. Both person and agent also miss the same unusual condition because they rely on the same summary. Interaction has amplified shared blind spots. The repair is unique information, an external verifier or a human route with raw evidence. Confidence should never route cases unless calibration has been demonstrated for the relevant population. High confidence from correlated components is not independence.

07 / Learning and fallbackThe allocation changes the people as well as the output

Assistance can transfer practice when it exposes reasoning, sources and feedback. It can also deskill when it supplies finished text that workers accept without constructing a judgement. Measure learning separately from assisted performance. Useful tests include delayed unaided cases, novel-case transfer, source reconstruction and outage exercises. A worker who is faster with the tool but unable to recover during an outage may have gained throughput while the organisation lost resilience.

Feedback design matters. If workers are rewarded only for acceptance speed, they learn to comply. If disagreement creates extra administrative work, expert correction becomes invisible. Capture accepted, edited and rejected suggestions, but do not interpret agreement as correctness. Link feedback to eventual outcomes and give experts a route to challenge the system, policy or training evidence.

NIST's AI RMF Core treats human-AI configuration, roles, monitoring and affected-party evidence as governance concerns.13 The EU AI Act similarly makes human oversight meaningful through competence, authority and the ability to intervene in relevant settings.14 Organisational studies also point to complementary skills and capability as conditions of adoption rather than side effects.15 A team contract should therefore fund training, observation and fallback as operating work.

Govern reassignment as ordinary work

A named owner should review allocation evidence, but affected workers need a protected route to challenge it. They often see emerging case types, awkward hand-offs and unsafe workarounds before aggregate metrics move. A challenge should identify the contribution, route, evidence and observed consequence. It should not require a worker to prove a complete alternative before the organisation investigates.

Triggers can be operational, technical or human. Operational triggers include rising exception age, correction, complaint, reversal or downstream delay. Technical triggers include a model release, tool change, calibration drift, source loss or a new permission. Human triggers include expert suppression, declining unaided quality, excessive review workload, workarounds or failure in an outage exercise. Any trigger can narrow authority while evidence is refreshed.

Model change does not automatically mean improvement for every contribution. Re-run the relevant component and team baselines, including cohorts that previously benefited least. A model that drafts better may anchor decisions more strongly. A model that retrieves more may increase review burden through surplus evidence. Version-level evaluation keeps the allocation attached to observed behaviour rather than vendor release language.

Worker development should influence routing. Novices may move from preparation support to independent cases as their unaided performance improves. Experts should not become permanent exception handlers whose routine knowledge is extracted while difficult workload accumulates. Rotate exposure, preserve practice opportunities and monitor whether the distribution of work is sustainable. Comparative advantage includes the opportunity cost of exhausting scarce expertise.

Outage exercises provide a direct fallback test. Remove or degrade assistance for a representative period, supply the documented contingency and observe quality, recovery time and escalation. The goal is not to prove that every worker can reproduce the agent's throughput. It is to confirm that the organisation can maintain safe service, recognise its limits and recover without improvising authority.

Reassignment should be reversible and legible. Record the old clause, the evidence that triggered change, the new allocation, transition support and a review date. Tell workers which cases moved and why. Preserve receipts so downstream effects can be compared. This turns team composition into a controlled operating decision rather than an informal accumulation of prompts and exceptions.

Incentives belong in that review. A worker paid for rapid acceptance may appear to trust the agent while actually responding to the scorecard. A supervisor rewarded for automation volume may hide exception labour in another queue. A product team measured on drafting time may never see complaint handling. The contract should identify these pressures and give outcome owners access to the complete value stream. Sample rejected advice, examine whether dissent changes performance assessment and check whether vulnerable customers or difficult cases are being displaced. Comparative advantage can be measured only when the actors are free to exercise the role the design claims they hold. Otherwise observed behaviour reflects constrained incentives rather than capability, and the next routing rule will learn from a distorted operating system.

08 / Worked teamCustomer support with cohort-aware assistance

Consider a support operation handling account, service and vulnerability cases. The outcome is not a fast reply. It is a correct and durable resolution that the customer understands, with policy compliance, appropriate care and retained worker capability. The contribution chain begins with case classification and evidence retrieval, moves through options and decision, then ends with action, explanation and learning.

For novices handling routine reversible cases, an agent retrieves verified account facts, shows source links and proposes response options. The novice selects or edits an option and records why. The system surfaces policy boundaries rather than hiding them. For experts, the agent provides retrieval and calculation but suppresses prescriptive wording by default. This protects scarce expert judgement and avoids dragging strong performers towards historical averages.

Stable fee calculations remain deterministic. A bounded agent may execute a pre-approved low-value remedy after input validation, with sampling and a hard stop on exception signals. Novel, vulnerable, disputed or high-consequence cases route to an authorised person. The person sees raw evidence, the agent's uncertainty and rejected alternatives, and can pause action. This is human decision authority, not a signature after the answer is fixed.

The evaluation compares this team with the strongest human-only, agent-only and feasible routed alternatives. It stratifies novice and expert effects. It includes handling time, resolution durability, corrections, complaints, customer understanding, policy exceptions and unaided outage performance. If drafting improves but complaints rise, the route fails. If novice performance rises while experts decline, the allocation changes by cohort. If the agent becomes reliable on a bounded routine population, supervised execution can replace per-case approval.

09 / Interactive allocationHuman-Agent Team Composer

The composer returns the smallest defensible allocation clause. It has no maturity score. It asks whether the contribution is stable, who holds unique information, how errors correlate, whether authority and intervention are credible, what review costs, where downstream effects land and whether workers continue to learn and recover.

How the controls route

A fully specified calculation routes to deterministic execution. A mixed contribution with human consequence routes to agent preparation and human decision. Demonstrated differences across cohorts and cases route dynamically. A calibrated agent can execute reversible low-risk work under exception supervision. Novel, high-consequence, unbounded or correlated-error work remains human-owned. Ceremonial review, weak baselines, rising downstream harm and missing fallback prevent expanded agent authority.

Every route is provisional. A new model version can change capability and calibration. Experienced workers can learn, novices can become experts, policy can tighten and the case population can shift. The clause therefore names a reassignment trigger before launch. Reallocation is ordinary operating work, not an admission that the original team failed.

10 / Design ruleCompose around the outcome, then keep learning

The useful human-agent team is neither a permanent partnership nor a contest between species. It is a changing allocation of contributions inside one accountable outcome. People, agents and deterministic services bring different information, capability, authority and cost to different cases. Those differences must be observed rather than assumed.

Begin with the contribution chain. Establish the strongest feasible baselines. Identify unique information and correlated errors. Give stable transformations to tested services. Let agents prepare evidence and reversible drafts where that lowers opportunity cost. Route by proven advantage. Permit execution only inside a calibrated and reversible envelope. Preserve human ownership where legitimacy, care, negotiation or unbounded consequence requires it.

Finally, measure what the allocation does to the wider system. Follow rework and harm downstream. Separate novice from expert effects. Test unaided performance and outage recovery. Protect disagreement and expert correction. Recompose when evidence changes. Realised complementarity is not the presence of both human and machine. It is the demonstrated ability of the whole team to produce an outcome that its strongest feasible component or route could not safely produce alone.

SourcesReferences

1. Autor, D. H., Levy, F. and Murnane, R. J. (2003). “The Skill Content of Recent Technological Change: An Empirical Exploration.” Quarterly Journal of Economics, 118(4), 1279–1333. DOI.
2. Acemoglu, D. and Restrepo, P. (2019). “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives, 33(2), 3–30. DOI.
3. Acemoglu, D. and Restrepo, P. (2020). “Robots and Jobs: Evidence from US Labor Markets.” Journal of Political Economy, 128(6), 2188–2244. DOI.
4. Brynjolfsson, E., Li, D. and Raymond, L. R. (2025). “Generative AI at Work.” Quarterly Journal of Economics, 140(2), 889–942. Journal DOI; working-paper record.
5. Dillon, E. W., Jaffe, S., Immorlica, N. and Stanton, C. T. (2025). “Shifting Work Patterns with Generative AI.” NBER Working Paper 33795. DOI.
6. Dell'Acqua, F. et al. (2025). “The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise.” NBER Working Paper 33641. DOI.
7. Hemmer, P., Schemmer, M., Kühl, N., Vössing, M. and Satzger, G. (2024). “Complementarity in Human-AI Collaboration: Concept, Sources, and Evidence.” arXiv record.
8. Hemmer, P., Schemmer, M., Kühl, N., Vössing, M. and Satzger, G. (2022). “On the Effect of Information Asymmetry in Human-AI Teams.” arXiv record.
9. Gao, R. et al. (2023). “Learning Complementary Policies for Human-AI Teams.” arXiv record.
10. Inkpen, K. et al. (2022). “Advancing Human-AI Complementarity: The Impact of User Expertise and Algorithmic Tuning on Joint Decision Making.” arXiv record.
11. Parasuraman, R., Sheridan, T. B. and Wickens, C. D. (2000). “A Model for Types and Levels of Human Interaction with Automation.” IEEE Transactions on Systems, Man, and Cybernetics Part A, 30(3), 286–297. DOI.
12. Bainbridge, L. (1983). “Ironies of Automation.” Automatica, 19(6), 775–779. DOI.
13. NIST. AI Risk Management Framework Core. Official resource.
14. European Union (2024). Regulation (EU) 2024/1689, Artificial Intelligence Act. Official Journal text.
15. OECD, BCG and INSEAD (2025). The Adoption of Artificial Intelligence in Firms. OECD Publishing. Official publication.