Information Technology (IT) throughput capacity as a constraint on unmet…

Information Technology (IT) throughput capacity as a constraint on unmet operational capability demand accumulation: empirical evidence and Artificial Intelligence (AI)-assisted delivery absorption modelling

2026-05-16 · agentic-ai tools-infrastructure cost-performance workforce-skills organisational-design benchmarks-eval · medium · source → · wiki →

key claims

Public evidence consistently shows that unmet central IT delivery speed or functional fit is a primary driver of shadow IT, business-managed IT, and citizen-development demand, so throughput shortfall is strongly associated with the accumulation of unmet operational capability needs even though the literature does not provide a universal elasticity coefficientFürstenau et al. (2020)Raković et al. (2020) Shadow IT (2020)Mitchell (2026)
Throughput loss is driven as much by dependency topology as by raw staffing levels, because tightly coupled architectures and domain-queue handoffs convert customer or operational demand into long waits, release orchestration, and queue growth that central teams cannot clear quicklyDevOps (n.d.)Conway (n.d.)Teamtopologies (n.d.)Topologies (2024)Mitchell (2026)
Only a bounded subset of workaround automation demand is tractable for centrally governed software delivery, because the strongest automation evidence repeatedly limits durable automation to stable, already-digital, repetitive, and governable work while the long-tail residue remains human, uneconomic, or governance-constrainedSpringer (n.d.)Fischer et al. (2022)Digital (2021)Viljoen et al. (2024)
The empirical literature on citizen-development rollout supports faster local solution creation and broader participation, but it does not provide a strong cross-firm effect size for backlog reduction, which means any three-year closure estimate must remain a bounded inference rather than a measured enterprise benchmarkAjimati et al. (2025)Binzer et al. (2024)Mitchell (2026)
AI-assisted software delivery evidence is too mixed to justify an aggressive planning multiplier, because bounded laboratory tasks show large gains while realistic repository work can still slow experienced developers down and DORA reports that throughput gains are offset by stability losses in weaker systemsPeng et al. (2023)Kalliamvakou (2022)Becker et al. (2025)Cloud (2025)
A realistic planning bracket for end-to-end AI-assisted throughput uplift is approximately 0% to 30%, where 0% reflects net slowdown or rework drag, 15% reflects modest system-level gain, and 30% reflects strong but platform-dependent improvement rather than unconstrained coding speedPeng et al. (2023)Becker et al. (2025)Cloud (2025)DeBellis et al. (2025)
Combining the demand-suitability evidence with the realistic throughput bracket supports a cautious three-year closure range of roughly 25% to 50% of current workaround automation demand, with the lower half of the range more plausible in estates that still have high data debt, boundary friction, and governance dragSpringer (n.d.)Fischer et al. (2022)Peng et al. (2023)Becker et al. (2025)Cloud (2025)DevOps (n.d.)
Results near the top of the range would likely require a workaround queue already dominated by tractable workflow automation plus strong internal platforms, dedicated platform teams, low dependency coupling, and fast feedback loops that let AI gains survive contact with production controlsCloud (2025)DevOps (n.d.)Teamtopologies (n.d.)Topologies (2024)

Research Question

What is the empirical relationship between Information Technology (IT) throughput capacity and the rate at which unmet operational capability needs accumulate across comparable organisations, and what proportion of workaround automation demand now handled outside central engineering can realistically be absorbed into centrally governed software delivery within three years under realistic Artificial Intelligence (AI)-assisted productivity assumptions?

Findings

Executive Summary

Central IT throughput is a strong directional constraint on the accumulation of unmet operational capability needs, but public evidence does not support a single universal coefficient for how quickly that backlog of unmet capability accumulates per unit of lost throughput.

A cautious synthesis answer is that centrally governed software delivery is likely to absorb only about 25% to 50% of current workaround automation demand under realistic AI-assisted productivity assumptions, not the whole queue.

That ceiling exists because only the stable, already-digital, rules-dominant middle of workaround demand is tractable for centrally governed software delivery, while the unstable tail remains resistant even if coding productivity improves.

Results near the top of the range would likely require strong internal platforms, loosely coupled delivery teams, and a queue already dominated by tractable workflow work rather than by data debt, policy ambiguity, or cross-team coordination friction.

Key Findings

Public evidence consistently shows that unmet central IT delivery speed or functional fit is a primary driver of shadow IT, business-managed IT, and citizen-development demand, so throughput shortfall is strongly associated with the accumulation of unmet operational capability needs even though the literature does not provide a universal elasticity coefficient.
Throughput loss is driven as much by dependency topology as by raw staffing levels, because tightly coupled architectures and domain-queue handoffs convert customer or operational demand into long waits, release orchestration, and queue growth that central teams cannot clear quickly.
Only a bounded subset of workaround automation demand is tractable for centrally governed software delivery, because the strongest automation evidence repeatedly limits durable automation to stable, already-digital, repetitive, and governable work while the long-tail residue remains human, uneconomic, or governance-constrained.
The empirical literature on citizen-development rollout supports faster local solution creation and broader participation, but it does not provide a strong cross-firm effect size for backlog reduction, which means any three-year closure estimate must remain a bounded inference rather than a measured enterprise benchmark.
AI-assisted software delivery evidence is too mixed to justify an aggressive planning multiplier, because bounded laboratory tasks show large gains while realistic repository work can still slow experienced developers down and DORA reports that throughput gains are offset by stability losses in weaker systems.
A realistic planning bracket for end-to-end AI-assisted throughput uplift is approximately 0% to 30%, where 0% reflects net slowdown or rework drag, 15% reflects modest system-level gain, and 30% reflects strong but platform-dependent improvement rather than unconstrained coding speed.
Combining the demand-suitability evidence with the realistic throughput bracket supports a cautious three-year closure range of roughly 25% to 50% of current workaround automation demand, with the lower half of the range more plausible in estates that still have high data debt, boundary friction, and governance drag.
Results near the top of the range would likely require a workaround queue already dominated by tractable workflow automation plus strong internal platforms, dedicated platform teams, low dependency coupling, and fast feedback loops that let AI gains survive contact with production controls.

Assumptions

Assumption: A 0% to 30% AI-assisted throughput uplift bracket is a realistic planning range for end-to-end delivery. Justification: the bounded-task Copilot experiment, realistic-task Model Evaluation & Threat Research (METR) trial, and DORA system-level evidence point to a wide but still bounded range rather than a single robust multiplier.
Assumption: The share of workaround demand that is tractable for centrally governed software delivery is materially below 100% in a typical large-enterprise queue. Justification: automation evidence repeatedly reserves a non-trivial tail for human handling, exception processing, or work that is not economically viable to automate.
Assumption: Demand without clear ownership, review paths, or environment controls remains resistant inside a three-year horizon even if it is technically scriptable. Justification: rollout evidence shows governance and expert support are part of what makes automation durable at scale.

Analysis

The evidence base supports a clear causal direction but only a bounded quantitative answer. Shadow-IT and citizen-development studies repeatedly show that people build or buy local solutions when official systems cannot meet the needed speed or functional fit. That makes throughput shortfall a credible driver of the accumulation of unmet operational capability needs.

The closure estimate cannot be derived from coding productivity alone because the throughput constraint is structural. DORA, Conway, Team Topologies, and the component-team case all indicate that queue growth depends on coupling, hand-offs, and platform quality, so a faster model in a still-coupled estate may increase change volume without proportionally reducing the workaround queue.

The one-quarter to one-half range reflects two compounding filters. First, not all workaround demand is tractable for centrally governed software delivery. Second, the tractable part is not all absorbable at the same speed because AI-assisted throughput gains are mixed and contingent. The result is a bounded rather than expansive three-year answer.

Plausible rival explanations remain. One rival view is that current field evidence understates future model gains. Another is that governance improvements to local workaround automation could outperform central absorption. The present evidence does not reject those possibilities, but it does show that platform quality, review structures, and dependency reduction are already prerequisites, which means the constraint is organisational as well as model-related.

Risks, Gaps, and Uncertainties

The cited public reviews and synthesis items do not report a stable cross-firm coefficient linking lost throughput to annual accumulation of unmet operational capability needs.
The cited citizen-development literature does not report a standardized enterprise effect size for backlog reduction from citizen-development or low-code rollout.
The 25% to 50% range is a low-confidence synthesis estimate and may shift with better internal demand-segmentation data or new field studies on AI-assisted delivery in high-control environments.
The GitHub Copilot and Model Evaluation & Threat Research (METR) studies bound the productivity question from opposite sides, but neither directly measures a regulated enterprise software-delivery organisation end to end.
Some workaround demand may disappear through policy clarification or process redesign rather than through central software delivery, so engineering closure is not the only valid remediation path.

Open Questions

What internal metric set best distinguishes stable workaround demand that is tractable for centrally governed software delivery from structurally resistant demand in a large-enterprise queue?
How much of current workaround demand is caused by missing platform capability versus missing decision rights, governance, or data quality?
What new field evidence will emerge on AI-assisted throughput in production engineering organisations with high compliance and release-quality bars?
Can a reliable early-warning metric for the accumulation of unmet operational capability needs be built from shadow-IT discovery, workaround inventory growth, and queue lead-time data?