When do AI productivity gains become economic returns?

Task performance, complementary investment and the accounting of released capacity

This retrospective review considers source versions available by 29 June 2026. The cutoff defines the evidence included; preparation and source checking took place on 21 October 2026.

Keywords: productivity; economic returns; AI; complementary investment; measurement

1. Why task savings and economic returns diverge

A reduction in drafting time is an operational observation. A reduction in payroll expenditure is a financial outcome. An increase in incremental profit is a further outcome involving demand, prices and costs. All can improve after an AI intervention, but one does not establish the others. The economic question concerns how a change in the production process alters valued output and resource use relative to an alternative.

This review examines what the productivity literature establishes about that translation. Its focus is the accounting and inference between task performance, workflow performance and firm outcomes. Marketing is a useful setting because producing work, distributing it and generating customer response are separate processes. Faster production can release resources while the distribution budget and commercial result remain unchanged.

The research question has two parts. First, how should experimental task gains be interpreted when other studies find smaller or negligible changes in wider economic outcomes? Second, what must be measured to distinguish released time, realised cost savings and additional value? These questions can be answered using established distinctions in productivity and technology-adoption research without asserting a universal return on AI investment.

2. Review method and version control

Targeted searches on 21 October 2026 located 18 primary studies and working papers on generative-AI task performance, firm productivity, complementary investment, adoption and measurement. Author or publisher abstracts, research summaries and publication records were consulted. Selection was purposive, combining recent task studies with older information-technology research that clarifies the economic mechanism. The review does not provide a systematic inventory, independent extraction or a meta-analysis.

Different versions of the same research can report different samples and estimates. Noy and Zhang's journal article and its earlier working paper, for example, should not be combined without identifying the version used. The discussion below uses the journal summary for that study [1]. The customer-support study is cited through its working-paper summary, with its subsequent journal publication noted [2]. Historical estimates remain attached to the models and deployment conditions actually studied.

Survey statements about perceived productivity, experimentally timed work, administrative earnings records and firm output measures are treated as different evidence types. Their units, denominators and identification strategies differ. A survey percentage describing adoption cannot be subtracted from a laboratory time-saving percentage to derive an implementation gap. The absence of a common scale is part of the interpretation rather than a numerical inconvenience.

3. Task experiments demonstrate conditional gains

Noy and Zhang's published study randomly exposed part of a sample of 453 professionals to ChatGPT for occupation-related writing tasks. The journal summary reports faster completion and higher assessed quality [1]. This establishes a result for the specified tasks and assessment conditions. It does not by itself show a change in annual marketing expenditure, customer demand or firm profitability.

Brynjolfsson, Li and Raymond study the introduction of a conversational assistant in customer support and report productivity gains that differ substantially across workers [2]. The setting contains repeated service interactions and a defined operational output. Its relevance is that AI can improve performance inside an existing workflow, with heterogeneous benefits. The findings do not supply a universal multiplier for every form of knowledge work.

The jagged-frontier experiment with 758 consultants reports benefits on selected tasks and worse performance on a task outside the assessed capability frontier [3]. METR's early-2025 trial, involving 16 experienced developers and 246 tasks in familiar repositories, instead found that allowing AI increased completion time [4]. Together these reports establish variation across interventions and task contexts. They cannot isolate which difference between studies causes that variation.

Task selection therefore matters twice: it determines where a tool is used and what an evaluation observes. A study made up of short, self-contained writing assignments may estimate a different treatment from one requiring extensive existing-project knowledge. The analyst needs to identify the task distribution before treating a mean result as representative of a function's workload.

4. Measuring the complete workflow

A production task has boundaries. Measuring only the interval between entering a prompt and receiving a draft excludes preparation, source checking, corrections and later rework. An estimate of task speed can be correct while overstating the gain in completed work. The relevant denominator depends on the question: generation efficiency, accepted deliverables per hour or total cost of producing a deliverable.

Quality should be held sufficiently comparable for a productivity contrast to be interpretable. Producing twice as much work of a lower standard changes the output as well as the resource requirement. The comparison may still be economically relevant if the lower standard is acceptable, but it should be described as a change in the quality-price bundle. Treating every item as an identical unit hides that trade-off.

Handoffs complicate measurement. Time released from a writer can reappear as review time for a manager or verification time for a specialist. A department-level measure that observes only the first person will miss the transfer. Whole-workflow measurement needs enough scope to include the people and systems whose workload changes because of the intervention. Otherwise the measured saving depends on where the analyst draws the organisational boundary.

Elapsed time and labour time also differ. A job can wait for approval while requiring little active effort. Reducing waiting may improve delivery speed without reducing paid hours, while reducing active effort may release capacity without changing a fixed approval deadline. Both can matter, but each implies a different mechanism. A productivity report should identify which was observed and avoid assigning a monetary benefit before the use of the released resource is known.

5. Why released time does not equal a lower bill

A salaried worker who completes an assignment sooner may perform another task, take on additional demand, improve the original assignment or leave capacity unused. The original salary may remain unchanged. Multiplying all released hours by a wage rate estimates the value of capacity under an assumption about its use; it does not demonstrate an equal reduction in cash expenditure.

Humlum and Vestergaard link Danish adoption surveys with administrative labour-market records. Their working-paper research finds limited early changes in measured earnings and hours alongside widespread use and reported benefits [6]. This is compatible with several adjustment processes. It does not prove that no useful work occurred, nor does it establish that every reported saving eventually becomes revenue. The observation window and measured outcomes limit both claims.

Capacity can have value even when it does not reduce expenditure. It may allow work that would otherwise be delayed or purchased externally. That value depends on a credible alternative: what would have happened without the released time? If an organisation would not have bought the additional work or if demand is insufficient, assigning the full external price to every extra deliverable can overstate the benefit.

Double counting is a particular risk. If a report values released hours at their labour cost and also counts the full value of the extra work produced with those same hours, the two lines may represent overlapping benefits. A coherent calculation explains whether it measures avoided cost, additional contribution or the opportunity value of a resource. The categories can coexist, but their overlap must be resolved before they are added.

6. Complementary investment and adjustment time

The productivity J-curve research models how general-purpose technologies require complementary intangible investments and how the timing of those investments affects measured productivity [5]. The relevant complements can include redesigned processes and human capital. Its argument concerns a mechanism and measurement problem; it is not proof that every unprofitable AI deployment is merely waiting for a future payoff.

Brynjolfsson and Hitt's firm-level computerisation research studies productivity and output growth over different horizons [11]. The historical setting predates generative AI, so the findings cannot determine the speed or scale of current returns. They support examining organisational adjustment and observation periods rather than assuming that installation and productive use occur simultaneously.

The implication for a marketing workflow is that the intervention includes more than a licence. Preparing approved source material, integrating access, changing review procedures and training users consume resources. Some are incurred once; others recur as the business and model change. A short trial can therefore combine unusually high setup costs with unusually close support. Extrapolating either cost or performance requires specifying which elements persist.

Longer observation can reveal returns that a short trial misses, but it also admits more competing changes. A later improvement may reflect staff learning, new models, changed demand or unrelated process reform. A time series alone cannot assign the improvement to one component. Adjustment-time explanations become stronger when the study observes the proposed complement and compares outcomes under a design capable of separating it from other changes.

7. Adoption rates and firm-level outcomes

Bick, Blandin and Deming document rapid generative-AI adoption using worker surveys [8]. Adoption measures indicate that a technology is reaching people and tasks. They do not establish the intensity, quality or economic effect of every use. An employee trying a tool once and a team relying on it daily can both be classified as users under a broad adoption definition.

Firm Data on AI reports executive survey evidence on adoption and business effects across countries [9]. The inspected institutional summaries describe widespread use alongside many respondents reporting no measured productivity effect. These are survey responses at a particular point in diffusion. They are informative about the reported business picture, but they cannot replace a causal estimate from actual output and resource records.

Firm productivity can change through several channels simultaneously. A new tool may improve task efficiency while demand falls, input costs rise or the firm changes product mix. A stable sales-per-worker measure can therefore coexist with a useful local intervention. Conversely, rising sales per worker during an AI rollout does not establish that AI caused the increase. The aggregate measure and the local mechanism need an identifying comparison.

Aggregation also weights activities. A large improvement in a task occupying a small share of total work can have a modest effect on whole-firm productivity. A small improvement in a frequent bottleneck can matter more. Acemoglu's macroeconomic analysis emphasises the share of tasks affected and task-level cost savings in translating microeconomic gains to broader effects [7]. The logic explains why a striking demonstration need not imply a similarly large aggregate change.

8. Data investment, organisational practice and selection

Bajari and colleagues investigate the relationship between data and firm performance [10]. Their work belongs to a broader literature in which having more information and using it effectively are distinct economic questions. A marketing organisation can accumulate extensive records while failing to make decisions that exploit them. The relevant treatment includes how the information enters work, rather than the volume stored alone.

Strength in Numbers examines data-driven decision-making and firm performance [12]. It is cited here as a working-paper contribution to the organisational-practice literature, without importing an unverified effect size. Comparing firms with different management practices also raises selection questions: firms choosing a practice may differ in other productive capabilities. The interpretation depends on how the original study addresses those differences.

The same issue affects comparisons between AI adopters and non-adopters. Adopters may have better data, stronger management or more suitable work before the technology arrives. A cross-sectional performance difference combines the intervention with those pre-existing characteristics unless the design separates them. Controlling for observed firm size or industry does not automatically remove differences that were never measured.

Complementarity is therefore both a substantive mechanism and an identification problem. If the tool and a process redesign are introduced together, their combined effect can be estimated under an appropriate design. Identifying their separate contributions requires additional variation. A report should not describe the entire package effect as a property of the model, nor assume that buying the same model reproduces the same organisation.

9. Cost boundaries and scenario assumptions

A cost account must state its comparison. The appropriate alternative might be the existing internal workflow, an external supplier, a simpler automation or no activity. Choosing the most expensive possible alternative can make almost any new system appear attractive. The counterfactual should represent what the organisation would realistically do, with the quality and volume differences visible.

Relevant costs include implementation and integration, recurring subscriptions or usage, human operation, verification and maintenance. Their classification depends on the decision horizon. A one-off setup expense may be material to a short campaign and spread across many future uses in a continuing service. Spreading it assumes those uses occur. A scenario model should expose the utilisation assumption rather than treating forecast volume as observed demand.

Unit costs can fall while total spending rises because usage increases. The arithmetic is straightforward: total variable expenditure depends on both the price per unit and the number of units consumed. More extensive reasoning, repeated generation and automated checking can change the number of units required for one accepted output. Cost per generated item and cost per accepted assignment therefore need separate denominators.

A scenario is useful when it makes uncertainty explicit. It can vary adoption, task volume, acceptance rates, review effort and the value of additional output. Its results remain conditional on those inputs. Presenting the central scenario as measured ROI removes the distinction between an estimate based on assumptions and an observation of realised performance. Sensitivity analysis should identify which assumptions drive the investment case.

10. Heterogeneity and the interpretation of averages

An average gain can conceal different responses by task or worker. A tool that helps less-experienced workers more may narrow performance differences in one setting [2]. Another workflow may impose extra verification on experienced specialists [4]. These observations concern particular interventions. They do not justify sorting all workers into a fixed hierarchy of who benefits from AI.

For a portfolio of marketing tasks, the relevant weighting reflects actual workload and resource use. Averaging a percentage saving across tasks gives each task equal weight even when one occurs daily and another once a year. A workload-weighted estimate answers a different question from an unweighted experimental mean. Reporting the weights is necessary for translating the trial into a forecast of operational impact.

The distribution of outcomes also matters. Mean speed can improve while a small set of assignments incurs substantial rework or a costly failure. A commercial account may need both the typical gain and the frequency and consequence of those failures. The evidence required depends on their rarity and severity. A trial with no observed incident cannot establish that the incident probability is zero.

Benefits can vary over time as users learn and model versions change. A repeated measurement should distinguish these changes from a stable treatment being observed again. Combining periods without recording the configuration can obscure both progress and regression. This is particularly important when an investment case assumes that future capability improvements arrive without additional integration or review costs.

10.1 The denominator in an economic comparison

Time reductions and throughput gains use different denominators. If an otherwise unchanged assignment requires less time, output per hour rises by the reciprocal of the time requirement. A percentage fall in time is consequently not the same numerical percentage as the rise in throughput. This arithmetic matters when a published task result is used in an investment model. The model should retain the original outcome scale or show the conversion explicitly, including the assumption that quality and all other resources are unchanged.

A further denominator is adoption intensity. A tool available to every employee may be used on only part of the eligible workload. Multiplying a task-level gain by total staff expenditure assumes both that the gain transfers to every task and that everyone uses it at the required proficiency. Neither follows from availability. A forecast should instead identify the eligible work, expected use and accepted-output requirement. Those quantities can be estimated, but they should remain visible as assumptions until observed.

The cost of an evaluation is also part of the investment decision, although it should not be confused with the operating cost of the resulting system. A small pilot may be designed to learn whether a larger commitment is worthwhile. Its immediate return can be negative while the information it produces prevents a poor rollout. That informational value depends on whether the evidence can actually change the decision. A trial conducted after an irreversible commitment supplies a different option from one conducted while alternatives remain open.

The reporting period should also match the cost period. A setup expense recorded in one month and benefits forecast over a year form a prospective investment case, while a monthly operating margin is a different account. Reporting both can be useful, provided the horizon and treatment of setup costs are explicit. Otherwise apparent payback can change simply because costs and benefits were assigned to different windows, with no change in the underlying workflow.

Vaccaro and colleagues distinguish gains over humans alone from gains over the better standalone performer in human–AI studies. [13] The GitHub Copilot experiment concerns completion of a specified programming task; its endpoint differs from an organisation’s realised financial return. [14] The Cybernetic Teammate working paper examines professional teamwork and expertise, extending the unit of analysis beyond an isolated worker. [15]

The FinOps Framework identifies unit economics within business-value measurement; the relevant unit remains a substantive accounting choice. [16] Hidden Technical Debt describes system obligations around an ML model that are absent from a model-only cost comparison. [17] Amershi and colleagues examine software-engineering work around machine learning. Integration and continuing operation belong in the resource boundary when productivity is translated into returns. [18]

11. Conclusions and limitations

The selected literature establishes useful but conditional productivity gains in several settings and weaker or negative results in others. Wider labour and firm outcomes measure different quantities and evolve through additional organisational and market processes. The apparent disagreement between task experiments and aggregate measures can therefore reflect differences in scope, time and identification as well as genuine variation in effects.

The evidence reviewed is selective and summary-based. It does not audit company accounts, reproduce the cited studies or identify a causal return for marketing as a whole. Historical information-technology research supplies mechanisms and measurement distinctions, not a forecast of the current AI cycle. Working papers and survey findings retain their provisional and observational status.

A defensible return claim links a defined intervention to comparable accepted work, measures all resource changes and distinguishes cash savings from capacity. It then establishes what additional value that capacity or output produces. This accounting preserves the practical value of task gains while avoiding the inference that every minute released is automatically an equal financial return.

Source access and appraisal register

The additional comparative sources in references 13–18 are used for the specific distinctions cited in the discussion. Access is limited to previously inspected primary-source abstracts, publication records or manuscripts. Their inclusion does not imply a fresh full-text risk-of-bias appraisal or direct validation in the application discussed here.

The 12 references below were located through targeted title and topic searches on 21 October 2026. Author or publisher abstracts, official research summaries and publication records were consulted. The discussion uses the claims available in those materials; full manuscripts, supplementary analyses and underlying data were not independently appraised. Institutional guidance is treated as guidance, rather than a controlled estimate of effectiveness. Where a working paper is cited, its provisional status remains part of the interpretation. The evidence map is selective and does not estimate the prevalence of results in the complete literature.

References