Brief mapping workshops nearly double startup revenue and increase customer acquisition

Hyunjin Kim, Dahyeon Kim, Rembrand Koning, INSEAD

RCT, high evidence established

A preregistered randomized field experiment in a 3‑month accelerator (n=515 startups) finds a 90‑minute “map AI to production” workshop increases discovered use cases by 44%, tasks completed by 12%, the share acquiring paying customers by 11 percentage points, and roughly doubles revenue, without commensurate increases in headcount or funding—clear, low‑cost guidance for managers seeking reliable near‑term gains.

Partial human–AI collaboration often wins because near-perfect AI accuracy is disproportionately costly

Wensu Li, Atin Aboutorabi, Harry Lyu, Kaizhi Qian, Martin Fleming, Brian C. Goehring, Neil Thompson

theoretical, calibrated framework, medium evidence framework

A calibrated model links AI accuracy costs to automation intensity, showing convex costs and diminishing returns make partial automation (keeping humans in the loop) the cost‑optimal choice in many settings, implying slower displacement and higher returns to redesigning workflows and verification rather than chasing full autonomy.

[Weak-link complementarities in essential tasks slow AI-driven productivity explosions]() — Charles I. Jones, Christopher Tonetti

theoretical, calibrated growth model, medium evidence framework

A task‑based growth model, calibrated to U.S. data, attributes much historical TFP to automation but shows that aggregate growth remains constrained by essential “weak‑link” tasks until they are automated, tempering forecasts of rapid GDP acceleration from AI even if capabilities rise broadly.

Over 17,000 worker evaluations find broad, continuous AI capability gains—'rising tides' not abrupt waves

Matthias Mertens, Adam Kuzee, Brittany S. Harris, Harry Lyu, Wensu Li, Jonathan Rosenfeld, Meiri Anto, Martin Fleming, Neil Thompson

descriptive, medium evidence

Standardized human assessments on O*NET‑like tasks across domains show steady LLM gains rather than isolated spikes, providing the most comprehensive empirical baseline yet for tracking AI capability diffusion and informing task‑level workforce policy.

Expert forecasters expect substantial AI capability gains, higher GDP, and materially lower labor force participation by 2030

Ezra Karger, Otto Kuusela, Jason Abaluck, Kevin Bryan, Basil Halperin, Todd Jones, Connacher Murphy, Phil Trammell, Matt Reynolds, Dan Mayland, Ria Viswanathan, Ananaya Mittal, Rebecca Ceppas de Castro, Josh Rosenberg, Philip E. Tetlock

descriptive, structured elicitation

A structured survey of 69 leading economists, 52 AI industry experts, 38 superforecasters, and 401 members of the public finds median GDP growth forecasts of 2.5% (above CBO baseline), with rapid‑progress scenarios projecting 75% of national wealth held by the top 10% by 2030 and labor force participation dropping to 55% by 2050—half attributable to AI. The starkest consensus: inequality will widen regardless of scenario.

New industrial benchmark finds LLM agents complete only two-thirds of PHM tasks and fail at tool orchestration Ayan Das, Dhaval Patel (descriptive, high quality)

A 75‑scenario, 65‑tool industrial maintenance benchmark reports ≈68% success and systematic orchestration failures, underscoring the need for better tool use and cross‑asset generalization before high‑stakes deployment.

AI raises returns to augmentable cognitive skills in the formal sector but not for informal workers in Colombia Cristian Espinal Maya (correlational, medium evidence)

LLM‑derived task augmentability linked to household survey data suggests higher wage premia for augmentable cognitive skills among formal workers, highlighting distributional and formality divides in AI’s labor impact.

Ontology-constrained neurosymbolic agents improve accuracy and compliance in enterprise domains Thanh Luong Tuan (quasi-experimental, medium evidence)

A controlled 600‑run study shows that grounding agents in enterprise ontologies reduces hallucinations and improves compliance and role consistency, offering a practical path for safer enterprise deployment.

Conversational, code-aware assistants shift developer work toward iterative specification and delegated diagnostics Ningzhi Tang, Chaoran Chen, Zihan Fang, Gelei Xu, Maria Dhakal, Yiyu Shi, Collin McMillan, Yu Huang, Toby Jia-Jun Li (descriptive, high quality)

Analysis of 11,579 IDE chat sessions shows developers iteratively specify tasks and offload diagnostics/validation to assistants, signaling new collaboration patterns and verification needs in software teams.

China's AIIAPZ policy-linked AI adoption boosts firms' operational resilience, especially in coastal and capital-intensive firms Yiting Hu, Xu Yan, Chaofan Duan, Xiaodong Yang, Jiaoping Yang (quasi-experimental, medium evidence)

A staggered policy rollout is associated with higher operational resilience via reduced agency conflicts and better resource allocation, with gains concentrated in advantaged regions and firm types.

Structured intent templates cut cross-model variance, help weaker models most, and reduce interaction rounds by ≈60% Peng Gang (quasi-experimental, medium evidence)

Protocol‑like “5W3H” prompts reduce goal‑misalignment and stabilize outputs across models and languages, especially benefiting weaker models—useful for multi‑model production stacks.

Faster new-technology creation raises the college wage premium by favoring quicker adopters (theoretical, medium evidence)

A calibrated model attributes about one‑third of the 1980–2010 rise in the college premium to faster invention, predicting cohort effects as technologies diffuse—relevant as AI shortens diffusion lags.

Household ChatGPT adoption raises leisure browsing but not productive online time (quasi-experimental, medium evidence)

Using pre‑exposure instruments and browsing data, adoption is linked to more leisure activity online with little change in productive browsing, raising questions about consumer surplus and nonmarket productivity.

Heterogeneous-agent model shows AI can both raise and lower the equity risk premium depending on displacement and investor participation effects Rajan Raju (theoretical, low evidence)

A decomposition highlights productivity, participation compression, and alignment risk channels that can push the equity premium in opposite directions depending on market structure, guiding scenario design for investors.

Audit finds ELT-Bench underestimated agents; extraction/loading largely solved and many transformation failures are benchmark errors Christopher Zanoli, Andrea Giovannini, Tengjun Jin, Ana Klimovic, Yotam Perlitz (descriptive, high quality)

An Auditor–Corrector review shows benchmark flaws overstated agent failures, implying stronger current capabilities and the need for audited evaluation before procurement.

Routine displacement in Indonesia is episodic and gender-asymmetric, briefly narrowing then widening the wage gap Wulan Isfah Jamil, Bambang Brodjonegoro, Diah Widyawati (quasi-experimental, medium evidence)

Shift–share and stacked differences indicate women had higher routine exposure but often reallocated to interpersonal roles, producing temporary narrowing of the gender gap before reversals.

AI in research yields modest short-run returns but reshapes team size, budgets, and tasks toward human capital Moh Hosseinioun, Brian Uzzi, Henrik Barslund Fosse (correlational, medium evidence)

Observational proposal data show modest performance gains concentrated in the top tail alongside reorganized teams and budgets, consistent with general‑purpose technology patterns.

AI automates contiguous chains of steps, making adjacency and fragmentation key to realized automation Mert Demirer, John J. Horton, Nicole Immorlica, Brendan Lucier, Peyman Shahidi, NBER (theoretical, medium evidence)

A theory with task‑level evidence argues automation clusters in adjacent steps, creating thresholds and non‑linear labor demand shifts when AI quality passes key margins.

Prompting with operational constraints cuts runtime and CO2e in GenAI-assisted literature workflows; 'green' language alone doesn't help Andrés Alonso-Robisco, Carlos Esparcia, Francisco Jareño (descriptive, high quality)

Decision‑rule prompts reduce compute and estimated emissions without changing outputs, offering a no‑regrets tactic for “greening” research operations.

Batched contextual training cuts per-task token use up to 60% while maintaining or improving accuracy Bangji Yang, Hongbo Ma, Jiajun Fan, Ge Liu (descriptive, medium evidence)

Training models to solve problems jointly reduces token cost while preserving accuracy, revealing a tunable throughput–accuracy trade‑off that matters for scaling.

LLM-driven adaptive questionnaires cut questions and increase user preference but slightly lower risk-assessment accuracy Diogo Silva, João Teixeira, Bruno Lima (quasi-experimental, medium evidence)

Two in‑app tests show improved user experience and efficiency with a small accuracy penalty, guiding insurers on when to use adaptive flows versus fixed forms.

APEX enforces payment gating and policy controls for autonomous agents, cutting unnecessary spend by 27% Mohd Safwan Uddin, Mohammed Mouzam, Mohammed Imran, Syed Badar Uddin Faizan (descriptive, medium evidence)

An HTTP‑402‑like payment gate with policy controls limits wasteful API calls and resists replay, offering a simple spend‑governance layer for agent deployments.

Bigger LLMs produce better point estimates but are massively overconfident; conformal recalibration fixes interval coverage Luka Hobor, Mario Brcic, Mihael Kovac, Kristijan Poje (descriptive, medium evidence)

Across 11 models, prediction intervals are severely miscalibrated, but conformal methods restore coverage—critical for risk‑aware decision support.

AI creators match human engagement by volume despite consumers preferring human content, mediated by recommender algorithms Tianhao Shi, Yang Zhang, Xiaoyan Zhao, Fengbin Zhu, Chenyi Lei, Han Li, Wenwu Ou, Yang Song, Yongdong Zhang, Fuli Feng (correlational, medium evidence)

Platform‑level data indicate AI‑generated content achieves comparable aggregate engagement by sheer volume even when users prefer human posts, pointing to the governance role of recommender systems.

Role-separated, validator-gated agent architecture prevents irreversible errors in environmental data curation Boyuan Guan, Jason Liu, Yanzhao Wu, Kiavash Bahreini (descriptive, medium evidence)

Deterministic validators and role separation restore “fail‑stop” behavior, reducing the risk of corrupting datasets—an architecture worth emulating for critical data pipelines.

Production-derived benchmark finds foundation models solve 53–72% of fail-to-pass coding tasks; running tests helps Smriti Jha, Matteo Paltenghi, Chandra Maddila, Vijayaraghavan Murali, Shubham Ugare, Satish Chandra (descriptive, high quality)

Real‑world prompts from seven languages show mid‑to‑high solve rates and that agents executing tests perform better, reinforcing the value of iterative verification.

AI innovation in Chinese firms links to lower carbon intensity via governance and green investment shifts Xingxing Lu, Lianying Liao, Xiaojuan Luo, Bing Zhao (correlational, medium evidence)

Panel evidence associates AI innovation with reduced emissions intensity through governance improvements and green reallocation, conditional on executive and government attention.

Workplace design determines whether AI automates or augments—WADI instrument proposed to measure 'human-centricity' Cristian Espinal Maya (theoretical, medium evidence)

A framework ties management practices to realized augmentation and offers a diagnostic tool (WADI) to measure readiness, reinforcing that design choices drive returns.

Autonomous coding agents increase activity but generate code with higher churn and lower survival than humans Razvan Mihai Popescu, David Gros, Andrei Botocan, Rahul Pandita, Prem Devanbu, Maliheh Izadi (correlational, medium evidence)

A 110k PR dataset shows rising agent activity but lower long‑run code survival, flagging maintainability and governance costs of autonomy.

AI adoption initially raises firms' electricity-output growth more than output, but effect fades after ~3 years Guoyao Wu, Zhiqiang Lan, Yang Xu, Ye Guo (quasi-experimental, medium evidence)

Firm panels show a temporary increase in electricity intensity post‑adoption that normalizes over time, informing energy planning during diffusion.

The short run is about execution. Causal evidence shows that brief, structured adoption efforts—mapping workshops, diagnostics—convert potential into revenue and customers. Complementary papers link policy nudges and management design to measurable resilience and reorganization, implying adoption is a managerial technology as much as a digital one. Energy and emissions effects are heterogeneous and path‑dependent, with temporary intensity spikes offset by governance‑driven green shifts. Editorially, the throughline is clear: processes, training, and operating models are the lever arm on AI returns. Human–AI collaboration and partial automation

Cost curves favor keeping humans in the loop because pushing AI to near‑perfect accuracy is disproportionately expensive. Task structure matters: automation tends to arrive in adjacent chains, creating threshold effects even when the aggregate equilibrium is “partial.” In practice, developers are already co‑specifying and delegating diagnostics, and autonomous code shows higher churn—evidence that verification workloads are the complement. As capabilities rise broadly, displacement is likely to be localized along automatable chains while aggregate redesign sustains human roles. Benchmarking, evaluation quality, and methods

Reality checks are getting sharper. Production‑derived and industrial benchmarks reveal respectable but incomplete success rates, with systematic gaps in tool orchestration and transformation. Audits of popular benchmarks show that evaluation flaws can materially understate capabilities, so procurement and regulation should not rely on single, unaudited scores. Meanwhile, conformal recalibration and batched contextual training offer pragmatic gains in uncertainty reliability and token efficiency, pointing to a more engineering‑mature evaluation ecosystem. Macro growth, risk, and distributional consequences

At scale, bottlenecks dominate. A calibrated weak‑link growth model cautions that aggregate acceleration will lag until essential tasks are automated. Distributional work indicates AI amplifies returns to augmentable cognitive skills in formal sectors and produces episodic, gendered transitions elsewhere, while theory in finance shows participation and alignment risks can raise or lower the equity premium. Expert forecasts still lean upbeat under fast‑progress scenarios, but the identification of bottlenecks and participation dynamics argues for humility on timing. Governance, energy, and externalities

Deployment quality shapes externalities. Temporary energy intensity spikes appear common during adoption, yet governance and green investment can deliver lower emissions intensity over time. Operational controls—payment gating, validator‑gated workflows, ontology grounding—are maturing to manage spend, safety, and irreversibility risk in agent systems. Editorially, the governance layer is no longer optional infrastructure; it is part of the production function.

Training clears the last-mile adoption barrier established

A randomized field experiment shows a 90‑minute mapping workshop substantially raises AI use, customer acquisition, and revenue. Implication: fund and mandate low‑cost onboarding and mapping programs before large capex on bespoke tools.

Partial beats full automation on cost curves framework

A calibrated model finds convex accuracy costs make partial human‑AI collaboration the optimal choice in many tasks. Implication: prioritize verification tools, workflow redesign, and reskilling over all‑in autonomy bets.

Bottlenecks cap near-term GDP acceleration framework

A weak‑link growth model indicates aggregate gains are throttled by essential tasks until they are automated. Implication: target R&D and standards at bottleneck tasks and enabling complements (data, interfaces, regulation).

Evaluation quality changes capability estimates suggestive

Benchmark audits reveal that errors can materially understate agent performance, while production‑derived tests still expose real gaps. Implication: require audited, domain‑grounded benchmarks and uncertainty calibration in procurement and regulation.

Adoption briefly raises energy intensity suggestive

Firm panels associate AI adoption with short‑run increases in electricity intensity that fade after about three years. Implication: pair diffusion programs with time‑limited efficiency incentives and grid planning.

Randomized field experiment in accelerator mapping AI to production

Mapping AI into Production: A Field Experiment on Firm Performance

A large RCT at startup scale provides rare causal evidence on an adoption intervention that moves revenue, offering a template for policy and corporate rollouts.

Auditor–Corrector benchmark audit with human validation

ELT-Bench-Verified: Benchmark Quality Issues Underestimate AI Agent Capabilities

A repeatable auditing pipeline that diagnoses benchmark errors and recalibrates ground truth improves evaluation reliability for procurement and research.

Large-scale worker-evaluation panel on O*NET-like tasks

Crashing Waves vs. Rising Tides: Preliminary Findings on AI Automation from Thousands of Worker Evaluations of Labor Market Tasks

Standardized human assessments across thousands of tasks create a broad baseline for tracking capability diffusion and informing task-level policy.