From Token-Max to Outcome-Max: How You Use AI Determines Its Productivity

Chen Xu; Mengqiao Liu; Beibei Li; Chenyan Xiong

cs.AIOct 5, 2026

From Token-Max to Outcome-Max: How You Use AI Determines Its Productivity

Authors: Chen Xu, Mengqiao Liu, Beibei Li, Chenyan Xiong

Organizations: Carnegie Mellon University · Amazon

Abstract

Generative artificial intelligence (AI) models can perform increasingly complex tasks, yet greater AI usage does not necessarily translate into proportional productivity gains. We identify token-max as one source of this inefficiency: when token consumption is treated as productive effort, agents are encouraged to over-exert and expend computation beyond what is necessary. We instead propose outcome-max, which rewards independently verified task completion per unit cost and induces a principled stopping rule. Then, to study these objectives, we develop a three-level simulation framework spanning immediate interaction, long-run behavioral adaptation, and organizational collaboration. Across all three levels, outcome-max improves the efficiency of AI-assisted production while largely preserving verified task performance. To further align these incentives with outcome-max, we introduce OutcomeShare, an incentive mechanism. Theory and simulation show that OutcomeShare can induce participation while generating shared gains for employees, firms, and LLM providers. Together, our results suggest that AI productivity not only depends on model capability, but also on how to construct the objectives governing AI use.

Figures & tables

Figure 1 : More tokens need not improve success. Blue: simulated token-max on SWE-bench. Red: real human coding sessions from SWE-chat.

Figure 2 : Simulation framework. L1: execution; L2: adaptation; L3: selection.

		Success (%) $\uparrow$				Tokens (k) $\downarrow$	Efficiency gain (%) $\uparrow$
Setting	Objective	Overall	Easy	Medium	Hard
L1: single-round human–AI interaction
SWE-bench	Token-max	48.8	62.2	43.9	39.8	21.1	0.0
SWE-bench	Outcome-max	48.2	66.7	48.9	25.0	9.4	+120.7
HumanEval+	Token-max	87.2	94.3	89.3	73.8	5.4	0.0
HumanEval+	Outcome-max	87.5	95.9	91.4	69.6	4.7	+15.3

Table 1 : Performance across three levels of human–AI simulation. Bold marks the better point estimate within each pair;

\dagger

denotes reconstructed math difficulty strata (Appendix A.1 ).

Figure 3 : Token-max versus outcome-max across policy implementations. Blue: success gap in percentage points (left axis); gold: efficiency gain in % (right axis). “With rules” states both the objective and continuation instructions; “Only” states the objective alone. Positive gaps favor outcome-max.

Panel of Figure 4 — Figure 4 : Where the efficiency gap arises. (a) Mean agent tokens by the turn at which conversations end; (b) token-max/outcome-max ratio of action counts (above 1: token-max repeats more); (c) a paired xarray attempt: token-max keeps requesting checks after a passing patch; outcome-max accepts.

Objective	Employee behavior after the first patch	Verified result; total tokens
Case A: both agents’ first patches already pass the task tests.
Token-max	Requests extra tests, import/changelog checks, then edge-case and documentation work.	Four passing patches; 33,602.
Outcome-max	Accepts the first patch and ends the interaction.	One passing patch; 7,696.
Why wasteful: repeated checks and edits increase tokens without improving verified task correctness.
Case B: both agents’ first patches fail the task tests.
Token-max	Questions backward compatibility of existing dim= calls, prompting a repair.	The next patch passes; 30,901.

Table 2 : Which follow-up work is wasteful, and which is useful? Two paired xarray attempts. The first-patch tests below are external; employees cannot see them. Appendix C.3.2 gives the traces.

Panel of Figure 5 — Figure 5 : Learned experience and interaction depth. (a) Employee playbook topics after early (light) and final (dark) learning; outcome-max emphasizes targeted work and stopping. (b) SWE-chat personas ( Baumann et al., 2026 ) and simulated objectives compared by normalized interaction length and mean-centered success. Human ratings and simulated pass rates are distinct measures; bars show 95% cluster-bootstrap intervals.

Figure 6 : Workforce selection and firm performance over five periods. Left: token-max firm; right: outcome-max firm. (a) Revenue (bars above zero), cost (bars below zero), and net profit (line) under common scenario accounts. (b) Twelve working employees per period, colored by ancestral profile. Each period ends with three replacements from top-ranked lineages; expanding segments include descendants.

	Deliveries / week		Weekly accounts (USD)
Production / contract	Correct $\uparrow$	Wrong $\downarrow$	Provider $\uparrow$	Firm $\uparrow$	Employee $\uparrow$
Token-max / token billing	9.8	15.6	2	$-$ 5,429	2,400
Outcome-max / token billing	38.3	61.3	3	$-$ 14,746	2,400
Outcome-max / OutcomeShare	19.4	2.9	535	275	3,443

Table 3 : Delivery and weekly accounts. OutcomeShare reduces defective shipments while retaining more correct deliveries and improving all three participant accounts relative to token-max. Selective delivery trades some correct-output volume for defect control; the design is described in Section 5.3.1 , with calibration details in Appendix C.3.4 .

Panel of Figure 7 — Figure 7 : Acceptance, allocation, and defect control. (a) Cumulative acceptance by turn with and without fine instructions; (b) weekly payoffs at fixed production versus fee share $\phi$ ; (c) correct and defective deliveries on matched policies; (d) success and defects versus fine level $\lambda$ .

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Dimension	Low propensity	High propensity
Iterates and refines	Accepts initial output	Requests revisions
Clarifies goals	Leaves goals implicit	States goals and constraints
Provides examples	Gives no examples	Supplies examples
Specifies format	Leaves format open	Specifies output structure
Sets interaction mode	Uses no explicit role	Assigns a collaboration role
Communicates tone	Leaves style open	States style preferences

Appendix

Table S1 : Employee behavior dimensions. The first six dimensions control description; the remaining five control discernment. Low and high archetypes use propensities of 0.1 and 0.9; single-behavior profiles elevate one dimension.

Parameter	Scenario value
Annual backlog	2,400 tasks
Task-value tiers	$16,$ 40, $200,$ 640
Loaded wage	$80 per hour
Employee-time anchor	15 minutes per attempt
Input, output, cache prices	$0.27,$ 1.10, $0.07 per million tokens
Employees and periods	12 employees; five selection periods

Appendix

Table S2 : Population-accounting assumptions. Monetary values define the stated scenario, not measured workplace prices or returns.

Panel of Figure 13 — Figure S1 : Interaction-depth coverage. Both panels compare WildChat coding conversations with simulated interactions, conditioned on multiple user turns. The tail view groups turns 2–12; it is not an additional sample. Overlap supports behavioral coverage, not causal productivity validity.

Setting	Token (%)	Outcome (%)	Gap [95% CI] (pp)
(a) Primary SWE benchmark
Overall	51.3	50.3	-1.0 [-8.2, +5.2]
Easy	62.0	67.6	+5.6 [+0.0, +11.4]
Medium	46.8	50.0	+3.2 [+0.0, +7.4]
Hard	43.6	26.9	-16.7 [-24.5, +0.0]
(b) Model and persona robustness

Appendix

Table S3 : Matched success rates and task-cluster uncertainty. Gaps are outcome-max minus token-max, with pointwise 95% task-cluster bootstrap intervals. Bold marks the higher success rate; ties are unbolded.

Setting	Objective	Success (%)	Tokens (k)	$\eta$
Haiku 4.5	Token-max	51.9	15.5	33.57
	Outcome-max	61.6	8.3	73.86
DeepSeek v4 Pro	Token-max	58.1	281.3	2.06
	Outcome-max	61.0	78.5	7.78
Detailed guidance	Token-max	45.8	30.2	15.16
	Outcome-max	50.0	7.4	67.33

Appendix

Table S4 : Paired means underlying Figure 3 . Within each pair, bold marks higher success and efficiency or lower token use; ties are unbolded.

Different backbones and human persona
SWE-bench; human: Opus 4.6
Agent	Human behavior	Overall	Easy	Medium	Hard	Tokens (k)	Eff. gain (%)
Haiku 4.5	Minimal input	26.7	20.0	40.0	20.0	9.8	0.0
	Detailed guidance	26.7	40.0	20.0	20.0	11.6	-15.5
	Active checking	46.2	40.0	66.7	40.0	23.1	-26.1
	Guidance + checking	28.6	40.0	25.0	20.0	25.0	-57.8

Appendix

Table S5 : Backbone and human-persona results with an Opus employee.

Different backbones and human persona
SWE-bench; human: Haiku 4.5
Agent	Human behavior	Overall	Easy	Medium	Hard	Tokens (k)	Eff. gain (%)
Haiku 4.5	Minimal input	40.0	60.0	40.0	20.0	8.6	0.0
	Detailed guidance	54.5	60.0	66.7	33.3	43.2	-72.8
	Active checking	25.0	40.0	33.3	0.0	38.4	-86.0
	Guidance + checking	50.0	60.0	66.7	25.0	49.9	-78.4

Appendix

Table S6 : Backbone and human-persona results with a Haiku employee.

Policy	Objective	Success (%) $\uparrow$	Tokens (k) $\downarrow$	Eff. gain (%) $\uparrow$
Haiku 4.5 agent; Sonnet 4.6 employee
Haiku 4.5	Token-max	51.6	15.5	0.0
	Outcome-max	60.5	8.4	+117.4
DeepSeek v4 Pro agent; Haiku 4.5 employee
DeepSeek v4 Pro	Token-max	55.8	262.7	0.0
	Outcome-max	63.0	91.5	+224.2

Appendix

Table S7 : Objective comparisons under learned policies. Each model pairing uses its own training and validation protocol.

Persona	Objective	Success (%)	Tokens (k)
Guidance and checking	Token-max	50.0	34.8
Guidance and checking	Outcome-max	54.2	11.3
More guidance	Token-max	42.3	29.7
More guidance	Outcome-max	50.0	7.4
More checking	Token-max	54.5	17.7
More checking	Outcome-max	52.2	11.4

Appendix

Table S8 : Objective comparisons across human personas.

	Token-max		Outcome-max
Ending	Runs	Solved	Runs	Solved
Verified 50-turn cap	18	5	46	6
Employee acceptance	0	0	32	14
Confidence threshold	2	2	4	1
Other agent ending	73	30	10	2

Appendix

Table S9 : Stopping reasons in the Hard cohort.

		Success (%) $\uparrow$				Tokens (k) $\downarrow$	Efficiency improvement (%) $\uparrow$
Adaptation	Objective	Overall	Easy	Medium	Hard
Employee adapts	Token-max	58.1	73.4	57.3	25.0	281.3	0.0
Employee adapts	Outcome-max	61.0	80.6	56.7	28.3	78.5	+276.5
Agent adapts	Token-max	62.0	78.1	59.7	30.4	207.3	0.0
Agent adapts	Outcome-max	61.2	79.6	58.0	26.8	94.6	+116.3
Both adapt	Token-max	56.2	72.5	55.0	20.4	261.7	0.0

Appendix

Table S10 : Adaptation outcomes on matched tasks and personas.

(a) Adapting role
Role	Objective	Success (%)	Tokens (k)	Efficiency (solutions per M tokens)
Neutral (zero profile)	—	57.0	54.2	10.51
Employee	Token-max	57.6	281.0	2.05
Employee	Outcome-max	60.9	78.9	7.72
Agent	Token-max	62.4	206.8	3.02
Agent	Outcome-max	61.0	96.0	6.36

Appendix

Table S11 : Adaptation and verification outcomes.

Panel of Figure 23 — Figure S2 : Adaptation-role comparison. Each of the six regime–objective arms is restricted to the same 327 task–persona cells, so these matched estimates differ from the valid-run cohorts in Table S11 . Each panel compares objectives under a common adaptation role.

Panel of Figure 24 — Figure S3 : Verification-instruction variants. No fine, gate check, and audit-fine use the same 129 matched task–persona cells. External correctness is distinct from employee acceptance. These prompted components were tested separately from the payment contract.

Objective	Success (%)	Tokens (k)	Efficiency $\eta$	Continued (%)	Later tokens (%)
Token-max	51.3	21.5	23.8	19.7	5.6
Outcome-max	50.3	9.4	53.6	3.5	1.3

Appendix

Table S12 : Continued interaction after a correct artifact is available. Bold marks the better objective for success, token use, and efficiency; continuation metrics describe behavior.

Panel of Figure 26 — Figure S4 : Post-correct-patch continuation. Both panels use the same 310 matched primary SWE pairs as Table S12 . The later-token share is descriptive, not an estimate of removable cost; subsequent verification may still be useful.

Initial patch	Objective	Employee decision and result	Patches	Tokens
Correct	Token-max	Requests tests and further checks; all saved patches pass.	P, P, P, P	33,602
	Outcome-max	Accepts the first correct fix.	P	7,696
Incorrect	Token-max	Questions backward compatibility; the next patch passes.	F, P, P, P	30,901
	Outcome-max	Accepts a patch that breaks existing dim= calls.	F	4,456

Appendix

Table S13 : Paired stopping cases on xarray’s dim -to- coord API change. P/F denote external pass/fail results for successive saved patches. Each pair shares a task, persona, and repetition; interactions are generated independently.

Aspect	Evidence or scope
Efficient stopping	High-guidance/low-checking profile, repetition 0. Both first patches pass; outcome-max stops, while token-max continues without improving correctness.
Premature stopping	Revision-only profile, repetition 0. Token-max asks about backward compatibility and obtains a passing patch; outcome-max accepts a patch failing both task-specific integration tests.
Selection	The efficient example is the first task–profile repetition within this task with complete replays, two correct first patches, and earlier, cheaper outcome-max acceptance. The failure example is a selected counterexample.
Verification	All final patches match the tested artifacts, with complete logs and no identified disk-space installation failure. Shared background failures are distinguished from task-specific resolution.
Interpretation	Approval is not verification. The examples show possible stopping mechanisms, not their frequency or whether extra interaction would repair the same outcome-max trajectory.

Appendix

Table S14 : Case interpretation and selection.

Decision	Token-max	Outcome-max
Scope	Search related methods and compare alternative fixes.	Request a minimal diff with the relevant context upfront.
Verification	Enumerate edge cases after every change; challenge confident answers.	Check a specific expected outcome; target the exact failure on retry.
Stopping	Revisit assumptions or request explanations after a solution is accepted.	Accept a plausible complete fix; cap failed attempts.

Appendix

Table S15 : What employees learn under each objective. Representative rules paraphrased from generation-3 employee playbooks in L2.

Metric	Token-max	Outcome-max
Initial-period net profit (USD)	-6,917	+2,141
Final-period net profit (USD)	-15,778	+3,826
KPI–efficiency rank correlation	-0.46	+0.44

Appendix

Table S16 : Firm accounts under population selection.

Panel of Figure 31 — Figure S5 : Costs and workforce composition under selection. Left: API and supervision costs over the five periods. Right: workforce share descended from example-providing or missing-context profiles. Each firm maintains twelve employees. Main-text Figure 6 reports the full lineage composition and resulting accounts.

Component	Result and interpretation
Production	Lower token use per resolution supports more attempted work under the capacity assumption.
Provider	Outcome fees raise revenue by sharing verified production value; revenue is before compute costs.
Employee	A bonus on verified output raises modeled income.
Firm	The residual value yields a positive profit estimate relative to the token-regime loss; its uncertainty interval still includes losses.
Scope	These accounts illustrate feasible value sharing, not a field deployment of the full payment-and-liability contract.

Appendix

Table S17 : How outcome payments share the value of efficient production. Capacity, value, and wage follow Table S2 ; these allocation-only accounts exclude defect losses.

Panel of Figure 33 — Figure S6 : Fee allocation under fixed observed production. The left panel shows recipient transfers; the right panel shows firm profit and its pointwise 95% task-bootstrap interval. All curves use the same valuation and capacity assumptions. The figure is not a behavioral fee intervention and does not identify provider net profit.

Production / payment	Verified successes/wk	Provider revenue ($)	Firm profit ($)	Employee income index
Token / token fees	$9.8$	$2.2$	$-1511.6$	$100.0$
Outcome / token fees	$38.3$	$2.9$	$1113.2$	$100.0$
Outcome / outcome fee	$38.3$	$703.2$	$412.9$	$100.0$
Outcome / equal surplus	$38.3$	$373.5$	$371.3$	$115.5$
Outcome / provider–firm	$38.3$	$559.2$	$557.0$	$100.0$
Outcome / employee–firm	$38.3$	$2.2$	$557.0$	$123.2$

Appendix

Table S18 : Fixed-production payment baselines. Weekly scenario accounts; all Outcome rows reuse the same production. Sharing rules redistribute the same observed value.

Production / payment	Verified successes/wk	Provider revenue	Firm profit	Employee income index
Token / token fees	$[6.6,13.2]$	$[2.1,2.3]$	$[-1988.9,-871.8]$	$[100.0,100.0]$
Outcome / token fees	$[24.9,53.9]$	$[2.6,3.3]$	$[-793.9,3592.4]$	$[100.0,100.0]$
Outcome / outcome fee	$[24.9,53.9]$	$[321.8,1199.0]$	$[-1112.9,2396.2]$	$[100.0,100.0]$
Outcome / equal surplus	$[24.9,53.9]$	$[0.0,1199.9]$	$[-791.1,1197.7]$	$[100.0,149.9]$
Outcome / provider–firm	$[24.9,53.9]$	$[0.0,1798.7]$	$[-791.1,1796.5]$	$[100.0,100.0]$
Outcome / employee–firm	$[24.9,53.9]$	$[0.0,2.3]$	$[-791.1,1796.5]$	$[100.0,174.9]$

Appendix

Table S19 : Uncertainty in weekly payment scenarios. 95% percentile intervals from 20,000 joint resamples of 50 task clusters, retaining both production arms within each sampled task. Prices, capacity, and review-time calibration remain fixed. Income is indexed to the fixed salary (

100

); provider revenue and firm profit are in USD per week.

Figure S7 : Participation and break-even are different requirements. At fixed outcome-max production, the light region preserves all three baseline accounts; the dark region also gives the firm nonnegative profit. Provider participation compares gross revenue, not net profit. Regions use point estimates, excluding audit costs and defect damages. The marked reference allocates 20% to the provider and 10% as employee bonus.

	Deliveries / week		Weekly accounts (USD)
Policy	Correct	Wrong	Provider	Firm net profit	Employee
Production baselines: submit every output
Token-max / token billing	9.8	15.6	2	$-$ 5,429	2,400
Outcome-max / token billing	38.3	61.3	3	$-$ 14,746	2,400
Alternative submission rules on outcome-max outputs
Random screen	3.2	5.2	179	$-$ 1,774	549

Appendix

Table S20 : What screening and allocation change. Same $240/reference-hour valuation, 80% inspection accuracy, and weekly USD accounts as Table 3 . The employee-heavy allocation balances delivered quality and participant payoffs. Screens use 20% provider fee and 10% employee bonus; all screened rules use damage-matched liability.

Task value / reference hour	Provider	Firm net profit	Firm 95% interval	Employee
$80	178	-1,508	$[-1,992,-880]$	2,748
$160	357	-617	$[-1,584,640]$	3,095
$240 (central)	535	275	$[-1,176,2,161]$	3,443
$320	713	1,166	$[-768,3,681]$	3,790

Appendix

Table S21 : Task value and OutcomeShare profitability. Only task valuation changes; correct/wrong deliveries stay at 19.4/2.9 per week. Accounts are USD per week; salary is $2,400. Higher task value improves firm returns; intervals quantify conditional task-sampling uncertainty. Intervals are conditional 95% task-cluster intervals, not uncertainty over the valuation itself.

Figure S8 : Inspection quality and selective delivery. The blue curve uses the employee-heavy OutcomeShare allocation, with the same $240/reference-hour valuation. Better signals improve the quality–quantity tradeoff; alternative screens and equal-thirds sharing illustrate dependence on the rule and allocation.

Fine level $\lambda$	Success (%)	Tokens (k)	Efficiency $\eta$	Acceptance (%)	Bad delivery (%)
Matched tasks (same cohort as the main figure)
0	40.0	54.9	7.3	90.0	50.0
0.25	40.0	58.1	6.9	90.0	50.0
0.5	30.0	60.6	5.0	90.0	60.0
1	40.0	72.5	5.5	70.0	40.0
2	40.0	61.6	6.5	70.0	30.0

Appendix

Table S22 : Fine sensitivity on matched and available tasks.

Explore similar work

May 12, 2026cs.GT

Human-AI Productivity Paradoxes: Modeling the Interplay of Skill, Effort, and AI Assistance

Generative Artificial Intelligence (AI) tools are rapidly adopted in the workplace and in education, yet the empirical evidence on AI's impact remains mixed. We propose a model of human-AI interaction to better understand and analyze several mechanisms by which AI affects productivity. In our setup, human agents with varying skill levels exert utility-maximizing effort to produce certain task outcomes with AI assistance. We find that incorporating either endogeneity in skill development or in AI unreliability can induce a productivity paradox: increased levels of AI assistance may degrade productivity, leading to potentially significant shortfalls. Moreover, we examine the long-term distributional effect of AI on skill, and demonstrate that skill polarization can emerge in steady state when accounting for heterogeneity in AI literacy -- the agent's capability to identify and adapt to inaccurate AI outputs. Our results elucidate several mechanisms that may explain the emergence of human-AI productivity paradoxes and skill polarization, and identify simple measures that characterize when they arise.

Ali Aouad, Thodoris Lykouris, Huiying Zhong

Massachusetts Institute of Technology

Human-Ai Collaboration Artificial Intelligence Literacy Artificial Intelligence Systems Generative Artificial Intelligence Skills Latent Confounders

Ask PDF

Apr 3, 2026cs.HC

The Augmentation Trap: AI Productivity and the Cost of Cognitive Offloading

Experimental evidence suggests that AI tools raise worker productivity, but also that sustained offloading can erode expertise. This creates a tradeoff when expertise is a complement to AI. To explore the consequences of this tradeoff, we develop a dynamic model in which a decision-maker chooses the intensity of practice-displacing AI offloading for a worker over time, trading immediate productivity against the erosion of worker skill. We decompose the tool's productivity effect into two components, one independent of worker expertise and one that scales with it. The model produces three main results. First, a decision-maker who fully anticipates skill erosion still rationally adopts AI when front-loaded gains outweigh long-run skill costs, lowering long-run productivity. The decomposition sorts deployments into five regions by their long-run effect, separating beneficial from harmful adoption. Second, the tradeoff introduces the potential for misaligned incentives to harm workers. When managers are short-termist or worker skill has external value, AI use can leave the worker worse off than with no AI, the outcome we call the augmentation trap. Third, when AI substitutes strongly enough for expertise, offloading can generate permanent divergence, with high-skill workers realizing their potential and low-skill workers deskilling. Small differences in managerial incentives can determine which path a worker takes.

Michael Caosun, Sinan Aral

Artificial Intelligence Deployment Artificial Intelligence Systems Offloading Trade-Off Augmentation

Ask PDF

May 18, 2026cs.AI

Generative AI and the Productivity Divide: Human-AI Complementarities in Education

Generative Artificial Intelligence (GenAI) is transforming how firms create, process, and apply knowledge, yet little is known about the heterogeneity of its productivity effects across users. We report results from a randomized controlled experiment in which participants-analogs of early-career knowledge workers-were assigned to self-study a technical domain using either traditional resources or large-language-model (LLM) assistance. On average, GenAI access significantly increased task performance, but the distribution of gains was highly uneven. Improvements were not predicted by GPA or prior knowledge, but by \textit{AI Interaction Competence (AIC)} -- the ability to elicit, filter, and verify model outputs. High-AIC participants realized outsized gains; low-AIC participants saw limited or even negative marginal returns. A scaffolding intervention (conceptual maps) reduced outcome variance, indicating that standardized workflows can mitigate inequality in AI-mediated performance. We interpret these findings through the lens of human-AI complementarities: GenAI raises mean productivity while introducing a new axis of capability inequality. Managerially, firms should pair GenAI access with short AIC micro-training and simple standard operating procedures to capture value consistently and avoid uneven adoption outcomes.

Lihi Idan, Bharat Anand

Industrial and Systems Engineering Department, Texas A&M University · Leonard N. Stern School of Business, New York University

Artificial Intelligence Literacy Generative Artificial Intelligence Human-Ai Collaboration Competence

Ask PDF