How does local optimisation of team- and role-level tooling in knowledge work…

Research Question

How does local optimisation of team- and role-level tooling in knowledge work reduce organisation-level throughput, and which interdependencies determine when local gains become global losses?

Findings

(Populated from §6 Synthesis above.)

Executive Summary

In knowledge-work pipelines with shared dependencies, local tooling optimisation reliably degrades whole-system throughput when the shared constraint's capacity is not increased commensurately, a mechanism that the Theory of Constraints (TOC) has described since 1984 and that DevOps Research and Assessment (DORA) 2025 and Faros AI 2026 telemetry now confirm at scale in AI-assisted software delivery. Faros AI telemetry across 22,000 developers in 2026 documents the failure mode directly: individual task completion rose 33.7%, but pull request (PR) review time rose 441% and production incidents per PR rose 242.7%, confirming that individual acceleration floods shared review infrastructure when review capacity is not scaled. The interdependencies that most amplify the gap between local and global outcomes are shared approval and review queues, tightly-coupled architectures that force coordination at every change, and change-management gates that cannot be bypassed without regulatory or architectural redesign. Interventions that restore whole-system throughput operate on two levers simultaneously: raising shared constraint capacity (platform engineering, review automation, architectural decoupling) and limiting the rate at which local speed floods that constraint (Work in Progress (WIP) limits, small-batch discipline, PR size controls).

Key Findings

1. Local tooling optimisation in a knowledge-work pipeline does not improve total throughput unless the shared constraint's capacity is increased, because throughput is bounded by the constraint and not by the sum of individual resource outputs, as TOC established in 1984.

2. Faros AI telemetry across 22,000 developers in 2026 measured individual task completion up 33.7% and epics completed up 66.2%, while PR review time rose 441%, bugs per developer rose 54%, and production incidents per PR rose 242.7%; this pattern is consistent with queue overflow at shared review infrastructure from local tooling gains.

3. The DORA 2025 report, based on nearly 5,000 technology professionals, found that Artificial Intelligence (AI) acts as an amplifier of existing organisational conditions rather than a universal productivity booster: teams with mature platform engineering and loose coupling convert AI gains into system-level improvement, while teams with fragmented tooling experience increased instability.

4. Shared approval and review queues are the primary interdependency surface that converts local speed gains into global throughput losses, because they represent a constraint whose capacity is fixed by personnel, policy, or regulatory mandate rather than by technical infrastructure alone.

5. Tightly-coupled architectures amplify local-to-global throughput losses by forcing coordination at every dependency boundary: a speed increase in one team generates cascading coordination demand from adjacent teams, compressing the constraint's capacity for productive work.

6. Platform engineering is the single most evidence-backed intervention for converting local gains into global gains, by providing shared, standardised infrastructure through which increased local output can flow without accumulating in ad hoc queues; DORA 2025 found 90% of organisations now have platform engineering capabilities but quality varies.

7. WIP limits and small-batch discipline are the structural mechanism by which non-constraint resources are subordinated to the constraint, preventing the queue-flooding dynamic, and DORA 2025 identifies working in small batches as one of seven capabilities required for AI gains to translate to organisational performance.

8. In regulated environments such as banking, local department automation that raises technical throughput predictably shifts the visible bottleneck to compliance, validation, and incident-response queues that cannot be removed without regulatory redesign, making the local-to-global throughput loss more durable than in unregulated contexts.

9. The "stay busy" norm identified by Forte Labs acts as a cultural amplifier of the local-optima failure mode: when managers equate individual utilisation with value creation, they structurally resist the TOC subordination prescription that non-constraint resources should idle rather than flood the constraint with new WIP.

10. Shadow Information Technology (IT) and local workaround automation exhibit the same structural failure as local tooling optimisation in software delivery: each local workaround reduces friction for one team while adding governance, maintenance, and approval cost to shared constraint surfaces, producing a global cost that exceeds the local saving.

Assumptions

• Knowledge-work pipelines with shared downstream dependencies follow the same constraint-throughput logic as manufacturing pipelines in TOC, because both involve discrete work items flowing through sequential stages with a rate-limiting shared resource; TOC has been applied to software delivery, project management, and healthcare in peer-reviewed literature cited in the prior backpressure-TOC repository item, and the Forte Labs series explicitly extends the analysis to knowledge work.

• Faros AI's developer telemetry represents a sufficient cross-section of knowledge-work organisations to support a directional inference about the local-to-global gap, even though the population over-represents GitHub-hosted engineering teams; Faros reports the population size (22,000 developers, 4,000+ teams) and methodology, and the direction is consistent with DORA survey data and TOC theory from independent sources.

• AI-assisted coding is a valid proxy for local tooling optimisation in the knowledge-work sense: it raises an individual contributor's delivery rate without automatically raising the capacity of shared downstream stages; DORA 2025 and Faros both use AI adoption as the test case and find the amplifier pattern consistent with this assumption.

Analysis

The evidence converges on a single structural pattern: local tooling gains become global losses at the shared constraint.

TOC provides the theoretical frame: throughput is bounded by the constraint, so non-constraint gains cannot raise global throughput; they can only accumulate as WIP in front of the constraint, reducing its effective capacity via coordination overhead.

The Faros telemetry provides the most direct empirical test in this body of evidence, comparing individual output against shared-infrastructure outcomes simultaneously. The 441% increase in PR review time against a 33.7% increase in task completion is inconsistent with local gains translating into global gains; instead, the shared review gate is absorbing the throughput increase as queue depth rather than producing faster delivery.

DORA 2025 contextualises this pattern: the seven capabilities required for AI gains to translate to organisational performance are all concerned with reducing the coupling density that amplifies the gap. Platform quality reduces queue accumulation. Small-batch discipline limits WIP at the constraint. Loose coupling reduces coordination-overhead tax on the constraint's capacity.

The dominant rival explanation is that the Faros data reflects a transition period and quality and stability will improve once organisations adapt their review processes to the new throughput level. This is consistent with TOC step 4 (elevate the constraint), but requires active intervention, not passive adaptation. The DORA 2025 finding that even organisations with strong engineering foundations see downstream quality pressure from AI adoption suggests that explicit constraint-elevation investment is required regardless of baseline maturity.

Risks, Gaps, and Uncertainties

• No peer-reviewed controlled experiment directly tests the local-tooling-to-global-loss hypothesis in knowledge work with a randomised design. The evidence is correlational (Faros telemetry, DORA survey) and theoretical (TOC). The direction is consistent across multiple independent sources, supporting the medium-confidence assignment.
• The Faros developer population over-represents GitHub-hosted engineering teams and may not generalise to non-software knowledge work (legal review, financial analysis, strategic planning). No equivalent telemetry study for those knowledge-work types was found in this search.
• The relative weight of the three amplifying interdependencies (approval queues, architectural coupling, change-management gates) is not directly measurable from the available evidence. The ranking is an inference from the Faros and DORA pattern, not from a direct comparison experiment.
• TOC's five-step model assumes a stable, identifiable constraint; in software delivery the constraint migrates after each improvement cycle (from coding to review, from review to deployment, from deployment to incident response). This migration is described in TOC step 5 but is not directly measured in the DORA or Faros data.

Open Questions

• What is the measurable coupling-density threshold below which local tooling gains stop reducing global throughput? This is a candidate for a new research backlog item in quantitative form.
• How does the local-global throughput gap behave in non-engineering knowledge work (legal review, financial modelling, regulatory analysis) where AI tooling adoption is accelerating but shared constraint surfaces differ from software delivery?
• Does architectural decoupling eliminate the local-global gap, or does it displace the constraint to a different shared surface (Application Programming Interface (API) governance, integration testing, data contract management)?

sources

• [x] Forte Labs (2019) Theory of Constraints 102: Local Optima - practitioner explanation of local versus global optimisation in knowledge work.
• [x] Forte Labs (2019) Theory of Constraints 101: Applying the Principles of Flow to Knowledge Work - foundational TOC principles applied to knowledge-work settings.
• [x] Forte Labs (2019) Theory of Constraints 103: The Four Fundamental Principles of Flow - flow principles including Work in Progress (WIP) limits and constraint exploitation.
• [x] Wikipedia: Theory of Constraints - overview of TOC five focusing steps, throughput accounting, and Drum-Buffer-Rope (DBR).
• [x] DORA (2025) State of AI-assisted Software Development - dora.dev summary - official summary: AI acts as amplifier of existing org conditions, not universal productivity booster.
• [x] InfoQ (2026) AI Is Amplifying Software Engineering Performance, Says the 2025 DORA Report - detailed analysis of DORA 2025 findings on fragmented tooling, platform maturity, and system-level throughput.
• [x] Faros AI (2025/2026) Key Takeaways from the DORA Report 2025 and AI Engineering Report 2026 - telemetry-backed analysis: "Acceleration Whiplash" showing individual gains with system-level quality deterioration.
• [x] DORA Capabilities: Loosely Coupled Teams - DORA capability definition linking architectural decoupling to high software delivery performance.
• [x] Mitchell (2026) Backpressure Infrastructure and the Theory of Constraints - prior repository synthesis on backpressure and the Theory of Constraints.
• [x] Mitchell (2026) How large can the Information Technology (IT) throughput constraint become before debt accumulation becomes discontinuous? - prior repository work on throughput constraints and debt accumulation.
• [x] Mitchell (2026) Why does do-mode demand keep displacing build-mode work in Information Technology (IT), and what operating model changes stop that displacement? - prior repository work on queue displacement and capacity trade-offs.
• [x] Mitchell (2026) How should banks govern department-level agent sprawl and bottleneck shifts across divisions? - domain case linking local automation to shifted organisational bottlenecks across shared review queues.