---
lang: en
template: insight-detail
---

# The productivity gap
> Jadey Insight | 2 October 2026 | Jörg Hubacher

![Abstract illustration: Offset fields of white light filaments meet a continuous orange path across dark gaps.](/_assets/img/insights/die-produktivitaetsluecke-abstract.webp)

**Why the economic value of agentic AI depends on how complete services are organised.**

This analytical article develops a testable organisational hypothesis. It distinguishes individual empirical findings, theoretical mechanisms and the author's own conclusions. A separate section considers the implications for Jadey after the main analysis.

## Abstract {#abstract}

AI can measurably improve the performance of individual employees. This does not, however, imply any particular productivity gain for the company as a whole. Between a faster task and an economically usable outcome lie further work, dependencies, decisions and checks. The effort they require can amplify, limit or consume local time savings.

This article calls this gap in translating local improvements into wider outcomes the productivity gap. Its hypothesis is that, for knowledge-intensive services with many dependencies, the organisation of collaboration becomes a decisive condition for technological progress to deliver economic value. Agentic AI expands the possibilities for changing that collaboration too. It can also create new coordination and verification costs.

The implication is a different unit of analysis: the complete, usable service, including preparation, coordination, verification and rework. Whether an agentic organisation delivers it more cheaply, quickly or reliably must be demonstrated against suitable alternatives.

*Keywords: productivity; agentic AI; coordination; operating model; accountability for outcomes; organisational design.*

## The problem: work gets faster, but the outcome is still pending {#problemstellung-die-arbeit-wird-schneller-das-ergebnis-lasst-auf-sich-warten}

A manufacturer uses AI to help respond to customer enquiries. Draft quotations are produced faster, technical information is easier to compile, and cost calculations are prepared more quickly. Yet the delivery commitment remains unresolved: materials and capacity have not yet been jointly confirmed, and a decision on an economically viable alternative is still missing.

This hypothetical example illustrates a possible limit to local acceleration. Several participants may save time while the customer receives a reliable answer hardly any sooner. The saving remains real. Its contribution to overall performance depends on what happens next.

In “Produktivität. Neu gedacht.”, McKinsey focuses on portfolio, innovation, scaling and transformation. For agentic AI, the publication recommends redesigning entire functions and the interaction between people and agents, with economic impact at the centre. This is a strategic argument for organisational design; the publication's historical company data do not provide causal evidence for the effects of today's agentic operating models. [McKinsey, 2026](https://www.mckinsey.de/publikationen/produktivitaet-neu-gedacht)

The question that follows is: under what conditions does faster work translate into a more capable organisation?

## Clarifying the terms: three levels of productivity {#begriffliche-klarung-drei-ebenen-der-produktivitat}

**Task productivity** describes the relationship between a defined work output and the effort required to produce it. An equivalent draft produced in less time represents progress at this level.

**Productivity across a complete service** takes account of the entire path to a usable outcome. A customer quotation, for example, includes clarifying technical requirements, calculating costs, reaching a viable delivery decision and making any necessary corrections. A faster draft can shorten this path. It can also trigger additional checks.

**Company productivity** considers value creation in relation to the resources used. It also depends on whether there is demand for the outputs, whether capacity can be put to use and whether released resources are deployed effectively. An available hour of work can enable additional output, reduce workload or initially remain unused.

The productivity gap described here is therefore not an already measured macroeconomic indicator. It refers to a possible gap between a local improvement and its contribution to the wider outcome. Shorter lead times and lower labour input must also be considered separately: a case can be completed faster while consuming more resources.

## The evidence: effects depend on the task, experience and working environment {#evidenzlage-wirkung-hangt-von-aufgabe-erfahrung-und-arbeitsumgebung-ab}

### Productivity gains from assistance are documented {#produktivitatsgewinne-durch-assistenz-sind-belegt}

In the journal version of “Generative AI at Work”, published in 2025, Brynjolfsson, Li and Raymond examine the staggered introduction of an AI assistant among 5,172 customer support employees at one company. The number of issues resolved per hour increases by an average of 15 per cent. Less experienced and lower-skilled employees benefit more. The effects differ substantially between employee groups. [Brynjolfsson, Li and Raymond, 2025](https://doi.org/10.1093/qje/qjae044)

This shows that assistance can improve economically relevant performance in a specific working context. The study examines neither a fully autonomous company nor an organisation-wide redesign. It contradicts a blanket claim that AI can only become productive after comprehensive reorganisation.

### A capable tool can also slow work down in a particular setting {#ein-leistungsfahiges-werkzeug-kann-im-konkreten-einsatz-auch-bremsen}

In a randomised trial in 2025, METR studied 16 experienced open-source developers working on 246 tasks in projects they knew well. With the AI tools available at the time, they took an average of 19 per cent longer to complete their work. The result concerns these tasks, participants and tools from the first half of 2025. [Becker et al., 2025](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

The timing matters: in February 2026, METR reported a follow-up study with indications of faster work. However, substantial selection and measurement problems led the researchers to regard its data as an unreliable signal of the current effect size. The earlier contrary finding is therefore not a general statement about developer productivity today. [Becker et al., 2026](https://metr.org/blog/2026-02-24-uplift-update/)

The customer support and developer studies concern different activities and working environments. Their percentages cannot therefore be compared directly. Together, they show why evaluation must start with the actual setting in which the tools are used.

### Organisational complements can initially require additional effort {#organisatorische-erganzungen-konnen-zunachst-aufwand-erzeugen}

In “The Productivity J-Curve”, Brynjolfsson, Rock and Syverson describe how new general-purpose technologies require complementary intangible investments, and how incomplete measurement of those investments can distort measured productivity trends. Processes, knowledge and organisational capabilities help explain why technological possibilities and visible economic outcomes may emerge at different times. [Brynjolfsson, Rock and Syverson, 2021](https://www.aeaweb.org/articles?id=10.1257/mac.20180386)

This offers an explanation for transitional effort. It guarantees neither future returns from a particular AI project nor the superiority of a particular operating model.

## Three mechanisms can limit local gains {#drei-mechanismen-konnen-lokale-gewinne-begrenzen}

The following mechanisms are the author's own organisational analysis. The cited sources explain individual relationships; their combined effect in an agentic company has not thereby been established empirically.

### 1. Dependencies remain when the activities involved become faster {#1-abhangigkeiten-bleiben-bestehen-wenn-ihre-beitrage-schneller-werden}

Malone and Crowston describe coordination as managing dependencies between activities. These include shared resources, required inputs and timing. This perspective focuses on how work fits together. [Malone and Crowston, 1994](https://crowston.syr.edu/sites/default/files/acmcs94.pdf)

For our quotation example, this means that a faster cost calculation confirms neither material availability nor priority over another order. Until these relationships are resolved, the delivery commitment remains open.

Acceleration can even increase the need for coordination. If sales produces more variants for review while technical approval capacity remains unchanged, the backlog may grow. The organisation only gains more usable outcomes when this bottleneck is addressed as well. Whether different prioritisation, additional capacity or a simpler approval process is appropriate depends on the specific cause.

### 2. Verification and rework can increase elsewhere {#2-prufung-und-nacharbeit-konnen-an-anderer-stelle-anwachsen}

As the effort required to produce initial drafts falls, their number may rise. Recipients must assess which outputs are usable, which conditions apply and which errors need correcting. If only production is measured, this work remains outside the assessment of success.

Multiple agents do not automatically solve this problem either. In their study of selected multi-agent systems, Cemri and co-authors identify failures in specification, alignment between agents, and task verification and termination, among others. The first preprint version, from March 2025, is used here. It identifies possible failure patterns; it does not quantify general operational failure costs. [Cemri et al., 2025, v1](https://arxiv.org/html/2503.13657v1)

This creates a question for organisational design: what additional verification is actually needed, and does it improve reliability sufficiently? Two reviewers using the same incorrect information can confirm the same error. What matters is suitable evidence and the effort required to reach an accepted service outcome.

### 3. Local objectives and decision rights can diverge {#3-lokale-ziele-und-entscheidungsrechte-konnen-auseinanderlaufen}

A function can improve its own metric while creating additional work in another. A high number of quotations is of little help if their delivery terms regularly have to be renegotiated. Low-cost purchasing can increase scheduling risks. Such conflicts between objectives must be addressed as part of the joint service.

Here, productivity intersects with the problem of delegation. An assignment may be clearly worded while knowledge, resources and decision rights remain distributed across different parts of the organisation. Whoever is responsible for bringing the outcome together needs either the appropriate authority or reliable access to the necessary decisions.

Agentic AI can identify such conflicts, prepare alternatives and execute decisions within a mandate. The company must establish the applicable priorities and limits. More capable models do not create this authority by themselves.

## Agentic work also changes coordination {#agentische-arbeit-verandert-auch-die-koordination}

The distinctive opportunity offered by agentic systems is that, alongside specialist contributions, they can also obtain, coordinate and carry those contributions forward. In the quotation example, an agent could gather missing information, identify conflicting variants, prepare a decision and then continue processing the case. This requires suitable capabilities and tools, as well as an appropriate assignment.

The design of complete services thus becomes practically relevant. The organisation must define the required outcome, who leads its delivery, which decisions are permitted and how fulfilment can be recognised. These arrangements can connect human, agentic and joint work.

An outcome agreement is one possible way to structure this. It records the contribution owed and the conditions attached to it. Whether such a design makes economic sense also depends on its reuse, maintenance effort and the number of additional handovers. Simply increasing the number of agents or services is not a measure of progress.

## Objections and alternative explanations {#einwande-und-alternative-erklarungen}

### Simple automation may already be sufficient {#einfache-automatisierung-kann-bereits-ausreichen}

This objection is valid. If a task is clearly defined, access to data is reliable and the result is easy to verify, an assistant or a predefined workflow may be sufficient. The customer support study shows that local assistance can deliver meaningful gains.

The organisational hypothesis mainly concerns services whose outcomes depend on several contributions and decisions. It does not justify comprehensive redesign for every task. An agentic approach must be assessed against the simplest suitable method.

### Better models could close the gap {#bessere-modelle-konnten-die-lucke-schliessen}

Model advances can reduce errors and make previously demanding tasks manageable. Some difficulties currently addressed through organisational measures will become smaller as a result. A weak system should therefore not be compensated for with ever more coordination if a more suitable tool solves the problem more simply.

Shared resources and conflicting priorities nevertheless remain, even when specialist capabilities are strong. The hypothesis would be weakened if model improvements alone regularly achieved the same overall effect in comparable cases as targeted changes to these dependencies.

### Additional agents can create new bureaucracy {#zusatzliche-agenten-konnen-neue-burokratie-erzeugen}

This objection also identifies a possible adverse outcome. Specialisation creates handovers; verification roles create verification work; dividing work into fine-grained parts can require more coordination than handling it as a whole.

The economic question is therefore whether the effort avoided exceeds the effort created. Centralised handling will be more suitable for some services. For others, reusable contributions from several participants may be effective. Both possibilities belong in the comparison.

### The observed gap could be a temporary implementation effect {#die-beobachtete-lucke-konnte-ein-vorubergehender-einfuhrungseffekt-sein}

Implementation and learning take time. An early measurement may underestimate later benefits. However, future learning gains cannot indefinitely justify persistent additional costs.

The comparison must distinguish implementation effort, ongoing operations and observed improvement. Review dates and exit conditions set in advance help distinguish a ramp-up period from a persistently disadvantageous division of work.

## The hypothesis can be tested against the complete outcome {#die-these-wird-am-vollstandigen-ergebnis-prufbar}

A sound comparison starts with a defined service for which demand is known. It includes the existing way of working, suitable simple AI assistance and, where appropriate, a different form of agentic delivery. Outcome requirements and case mix must remain comparable; changes in demand, staffing, models and access must also be recorded.

| Observation | What it reveals |
| --- | --- |
| Accepted outcomes against the same quality requirements | Whether more genuinely usable services are delivered |
| Effort including preparation, coordination, verification and rework | Whether time and costs fall overall or are merely shifted elsewhere |
| Lead times, waiting times and backlogs | Whether local acceleration stalls elsewhere |
| Errors, abandoned cases and reopened cases | Whether greater speed is achieved at the cost of incomplete handling |
| Human involvement across the entire service delivery process | Which work remains necessary and how it changes |
| Use of the outcomes | Whether additional capacity leads to services for which there is demand or to another intended effect |

Implementation and operating costs belong within the same economic assessment period. Difficult or rejected cases must not disappear from the evaluation. Workload and the need to maintain skills must also be considered if they change as a result of the new division of work.

**The hypothesis is supported** if, under comparable conditions, targeted changes to handovers, decision paths and verification generate additional overall gains beyond those achieved by accelerating tasks alone.

**The hypothesis is weakened** if these differences do not appear, if the new coordination effort consumes the benefits or if the same outcomes are achieved with a simpler solution. A growing number of automated steps alone does not resolve this question.

## Conclusion {#schlussfolgerung}

The existing research documents specific productivity gains and specific contrary findings. It does not provide a uniform factor for converting technological capability into company productivity.

The hypothesis developed here focuses on the conditions under which that translation occurs. Where work depends on many contributions, coordination, decision rights and verification are part of the productive service itself. Agentic AI opens up possibilities for distributing this work differently. Its value arises under the conditions of the particular operation and must be assessed there.

For company leadership, this creates a clear evaluation task: **Is a complete, needed service being delivered with less total effort or greater value — and which organisational change contributes to that result?**

## Implications for Jadey {#einordnung-fur-jadey}

*The following section describes the connection to Jadey's organisational model. It is separate from the analytical article and its empirical evidence.*

Jadey's service-oriented organisational model starts with complete services and the relationships between them. A service describes the agreed outcome. An accountable service agent leads delivery, including the contributions required, their verification and any necessary rework. Accountability for its own outcome remains in place when it delegates further. The model thus proposes a specific design approach to the productivity gap examined here.

There is a deliberate difference from the earlier Insight [“Management without self-interest”](/en/insights/management-ohne-eigeninteresse-betriebsmodell). In that article, accountability for outcomes explicitly remains with people; agents take on coordination. The organisational model discussed here goes further: it assigns service agents the operational obligation to lead the services they have accepted through to a verifiable outcome. Management services can also be included within appropriate mandates.

This operational allocation does not transfer legal liability or decision rights reserved for people. It is a further-reaching organisational assumption whose practical viability requires evidence of its own. The studies cited in this Insight establish neither the superiority of this model nor any particular economic performance of Jadey.

[Explore Jadey's service-oriented operating model](/en/operating-model)

## Source review and limits of the claims {#quellenprufung-und-aussagegrenzen}

The sources support different parts of the argument. The customer support study examines a specific operational deployment. The METR studies concern defined developer groups and particular generations of tools. The J-Curve paper explains complementary investments and measurement problems; coordination research provides concepts for dependencies. The multi-agent preprint identifies possible failure patterns. McKinsey provides the strategic starting point.

The claims used were checked against abstracts, introductions, and summaries of methods and results in the linked primary publications. No independent replication was conducted. Their application to complete company services is identified as the author's own hypothesis. Study results are neither added together nor extrapolated to an agentic company. The bibliography identifies the version used in each case.

## References and sources {#literatur-und-quellen}

1. **McKinsey & Company (2026):** [Produktivität. Neu gedacht.](https://www.mckinsey.de/publikationen/produktivitaet-neu-gedacht), 10 June 2026. Used: methodology and the section on transformation with agentic AI.
2. **Brynjolfsson, E.; Li, D.; Raymond, L. R. (2025):** [Generative AI at Work](https://doi.org/10.1093/qje/qjae044). *The Quarterly Journal of Economics*, 140(2), 889–942. Used: abstract and introduction of the journal version.
3. **Becker, J.; Rush, N.; Barnes, B.; Rein, D. (2025):** [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/). METR, 10 July 2025. Used: research report covering experimental design, results and limits to generalisation.
4. **Becker, J.; Rush, N.; Cunningham, T.; Rein, D.; Mahamud, K. (2026):** [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/). METR, 24 February 2026. Used: follow-up findings and explicitly identified selection and measurement problems.
5. **Brynjolfsson, E.; Rock, D.; Syverson, C. (2021):** [The Productivity J-Curve: How Intangibles Complement General Purpose Technologies](https://www.aeaweb.org/articles?id=10.1257/mac.20180386). *American Economic Journal: Macroeconomics*, 13(1), 333–372. Used: abstract of the journal version, supplemented by the [summary from the authors' institution](https://digitaleconomy.stanford.edu/publication/the-productivity-j-curve-how-intangibles-complement-general-purpose-technologies/).
6. **Malone, T. W.; Crowston, K. (1994):** [The Interdisciplinary Study of Coordination](https://crowston.syr.edu/sites/default/files/acmcs94.pdf). *ACM Computing Surveys*, 26(1), 87–119. Used: abstract and introduction on coordinating dependencies.
7. **Cemri, M. et al. (2025):** [Why Do Multi-Agent LLM Systems Fail?](https://arxiv.org/html/2503.13657v1). arXiv:2503.13657, **version 1 of 17 March 2025**. Used: abstract and introduction on the failure taxonomy; a preprint, not evidence of economic performance in operations.