Systems Capability Debt Thesis Evidence Assessment

Synthesis Question

Across the completed research items mapped to the authoritative systems capability debt thesis, which thesis claims are supported, qualified, contradicted, or still unresolved, and what is the strongest version of the thesis that the current evidence can sustain?

Cross-Item Findings

Claim 1 — "Organisations accumulate systems capability debt ... arising from integration failures, missing functionality, inaccessible or untrustworthy data, and workflow failures."

Evidence position: Supported
Confidence: High

The completed items support systems capability debt as a real organisational condition, but they support it through observed failure mechanisms rather than through a single corpus-wide direct measurement construct. Integration gaps, backlog fragmentation, domain handoffs, absent accountability, and unmanaged workaround growth recur across the source set.

Strengthening finding: the research base sharpens "systems capability debt" into at least three empirically distinct debt-creation channels: governance defects, delivery-topology defects, and standards-and-lifecycle defects.

Claim 2 — "This debt is not passive. It generates active demand ... humans acting as connective tissue between systems ... That demand flows into IT departments as change requests."

Evidence position: Supported with qualification
Confidence: Medium

The source set strongly supports the claim that unmet capability gaps generate active workaround demand, including shadow Information Technology (IT), citizen development, local automations, and requests routed into central teams. It is weaker on direct public measurement of the manual "human glue" workload in hours or Full-Time Equivalent (FTE) terms.

Qualification: the evidence supports the mechanism directionally, but not a universal conversion rate from unmet debt to manual synchronisation and translation labour.

Claim 3 — "IT departments are structurally capacity-constrained and cannot clear the backlog."

Evidence position: Supported with qualification
Confidence: Medium

The research supports structural capacity constraint, especially where annual project budgeting approves demand in large batches, where product teams absorb project-coded work, and where customer demand must traverse multiple domain queues. Public evidence also supports only a bounded central absorption range of existing workaround demand over three years.

Qualification: the evidence base does not support a pure staffing-or-throughput reading. Dependency topology, coupling, review burden, and governance drag are part of the constraint itself.

Claim 4 — "The party naming the risk does not control delivery spend ... The party bearing the cost of the backlog does not capture the benefit of closing it ... The backlog therefore persists ... because the governance structure actively prevents anyone from acting on it."

Evidence position: Supported, but overstated in its strongest form
Confidence: High for mechanism; Medium for absoluteness

The split-accountability item directly supports the missing-integrator mechanism: separated risk, operating cost, and benefits ownership repeatedly produce weak investment cases, cost shifting, persistent backlogs, ownerless risk items, and slow correction.

The stronger thesis wording, that governance "prevents anyone from acting on it," is partly qualified by later evidence showing that separated structures can still act when an explicit integrator has budget leverage, benefits authority, convening power, and escalation rights.

Strongest supported version: split accountability reliably prevents timely action unless the organisation substitutes co-location with explicit closure rights.

Claim 5 — "These two mechanisms are additive. Throughput constraint limits how fast gaps can be closed ... Split accountability prevents the investment decision from being made ..."

Evidence position: Supported
Confidence: Medium

The completed items converge on two distinct but interacting mechanisms. Project-to-product mismatch and cohort-to-domain topology explain queue growth and slow closure even when work is approved. Split accountability explains why some closure investments never achieve decisive approval in the first place.

Strengthening finding: the research suggests a third amplifier that the thesis text currently underplays: vendor-mediated standards and lifecycle failure can preserve backlog even after apparent implementation spend.

Claim 6 — "The workaround taxonomy has expanded: bespoke spreadsheets, VBA macros, Access databases, Power Automate flows, and now AI agents ... Citizen development in this thesis means non-engineers deploying agents or automation tooling to fill gaps IT has not closed."

Evidence position: Supported
Confidence: Medium

The LCNC rollout item and the workaround-persistence item support the taxonomy expansion from local business-built solutions toward governed or shadow automation estates. The evidence is strongest for LCNC and automation-programme rollout patterns, and indirect but consistent for Artificial Intelligence (AI) agents as the current generation of workaround.

Qualification: the completed items do not establish one empirical sequence that every organisation follows from spreadsheet to bot to agent. They support the broader class claim that local workaround capability keeps moving outward from central engineering.

Claim 7 — "AI can be deployed in two modes ... Do mode automates the workaround ... Build mode closes the gap ... These modes are not equivalent."

Evidence position: Supported
Confidence: High

The evidence strongly supports the non-equivalence. Build mode is more controllable because its dominant control point is release gating before external effect; do mode is a live runtime control problem with materially lower baseline task reliability and more post-effect failure visibility.

The cost evidence also supports the economic non-equivalence: recurring agent operation is usually less attractive than gap closure when the underlying capability can be built within a bounded planning horizon.

Claim 8 — "Every investment in do-mode tooling is a decision not to invest in build-mode delivery."

Evidence position: Qualified support
Confidence: Medium

The completed items support a real displacement risk. Temporary automation gives users a faster local path, can absorb attention and budget, and can weaken urgency for deeper remediation.

But the research also shows conditional complementarity: governed temporary automation can surface demand, process knowledge, and target-state requirements that later improve build-mode closure.

Strongest supported version: do-mode investment creates displacement risk by default, and becomes complementary only when explicitly governed as a temporary bridge with a path back into core delivery.

Claim 9 — "Do mode also carries a structural cost that is systematically underestimated ... Production AI agents have the largest external dependency surface of any system type ... variance ... is harder to bound ... automated workarounds are rarely decommissioned ..."

Evidence position: Mostly supported, with one unresolved superlative
Confidence: Medium

The dependency-surface item strongly supports the claim that production LLM agents have a broad multi-class external dependency surface spanning providers, lifecycle changes, harnesses, tools, identity, and runtime context. The variance-control item supports that do-mode variance is harder to bound than build-mode variance because many failures become visible only after live action. The decommission item supports that temporary bridges do not self-retire and need explicit lifecycle controls.

Unresolved sub-claim: the phrase "largest external dependency surface of any system type" is not fully proven by the completed items. The research supports "unusually large and operationally heterogeneous" much better than the absolute superlative "largest of any system type."

Unresolved sub-claim: the research supports persistence risk and the need for retirement controls, but does not provide a robust cross-organisational post-gap-closure persistence rate.

Claim 10 — "Vendor-supplied do-mode automation compounds this ... the organisation loses visibility of and control over the automation surface entirely."

Evidence position: Partially supported
Confidence: Low to Medium

Vendor standards and lifecycle evidence supports the broader mechanism that externally mediated delivery weakens convergence, exit control, compliance evidence, and decommission safety. Dependency-surface evidence also supports provider and harness drift as real operational exposures.

What remains unresolved is the stronger thesis wording that vendor-supplied do-mode automation causes organisations to lose visibility and control entirely. The completed items support degraded visibility and reduced control, not total loss as a general rule.

Claim 11 — "Prioritise AI investment in build mode ... Use do-mode agents only where software delivery cannot close a gap within an acceptable horizon ..."

Evidence position: Supported with bounded conditions
Confidence: Medium

The source set supports build-mode priority as the strongest evidence-backed default. Cost, controllability, and lifecycle evidence all point in that direction for stable, repetitive, compliance-relevant gaps.

The bounded exception is also supported: do mode is economically and operationally defensible as a bridge when the target-state capability is too slow, too uncertain, or too unstable to close within the planning horizon.

Contradictions and Tensions

Tension	Items	Resolution
The thesis implies split accountability can block action outright; the governance-structures item shows separated structures can still act if they create an explicit integrator with budget and escalation rights.	2026-05-14-org-failure-modes-split-risk-cost-benefits-accountability, 2026-05-16-governance-structures-build-mode-without-full-accountability-colocation	resolved
The thesis treats do-mode investment as displaced build-mode investment; the workaround-persistence and LCNC items show that temporary local automation can also complement later build-mode closure when governed as a bridge.	2026-05-16-do-mode-demand-persistence-and-build-mode-displacement, 2026-05-14-citizen-development-rollout-empirical-evidence, 2026-05-16-decommission-trigger-design-for-do-mode-agents	resolved
The thesis foregrounds throughput constraint; the delivery-topology items show that coupling and boundary design are part of the constraint rather than secondary implementation details.	2026-05-14-org-failure-modes-project-demand-product-it, 2026-05-14-org-failure-modes-cohort-demand-domain-it, 2026-05-16-it-throughput-constraint-magnitude-and-debt-accumulation-rate	resolved
The thesis states that production agents have the largest external dependency surface of any system type; the dependency-surface item proves a very large heterogeneous surface but does not prove the absolute superlative against every other system class.	2026-05-16-external-dependency-surface-taxonomy-for-production-llm-agents	open
The thesis implies automated workarounds are rarely decommissioned; the completed research proves persistence risk and the need for explicit retirement controls, but not a generalisable persistence percentage after gap closure.	2026-05-16-do-mode-demand-persistence-and-build-mode-displacement, 2026-05-16-decommission-trigger-design-for-do-mode-agents	open

Perspectives Considered

• Organisational-governance perspective: accountability gaps, split risk/cost/benefits ownership, and decision-rights design converge on governance as a first-order mechanism, not merely a background condition. Represented by 2026-05-14-org-failure-modes-accountability-gaps, 2026-05-14-org-failure-modes-split-risk-cost-benefits-accountability, and 2026-05-16-governance-structures-build-mode-without-full-accountability-colocation.
• Delivery-topology perspective: project funding, cohort-to-domain routing, dependency topology, and queueing show how demand becomes backlog even before model capability enters the picture. Represented by 2026-05-14-org-failure-modes-project-demand-product-it, 2026-05-14-org-failure-modes-cohort-demand-domain-it, and 2026-05-16-it-throughput-constraint-magnitude-and-debt-accumulation-rate.
• Citizen-development and workaround perspective: LCNC rollout evidence converges with workaround-persistence evidence, but qualifies the thesis by showing that governed local automation can also complement later gap closure. Represented by 2026-05-14-citizen-development-rollout-empirical-evidence and 2026-05-16-do-mode-demand-persistence-and-build-mode-displacement.
• Economic perspective: recurring agent-operation cost versus software-gap-closure cost supports the build-mode preference, while leaving precise breakeven thresholds assumption-sensitive. Represented by 2026-05-16-agent-operational-cost-vs-gap-closure-cost.
• Operational-risk perspective: dependency-surface, variance-control, decommission, and vendor-governance items converge on do mode as a live operational system with broader runtime exposure than gated software delivery. Represented by 2026-05-16-external-dependency-surface-taxonomy-for-production-llm-agents, 2026-05-16-variance-control-comparison-across-delivery-modes, 2026-05-16-decommission-trigger-design-for-do-mode-agents, and 2026-05-14-org-failure-modes-vendor-standards-gaps.

For acronym discipline, this synthesis uses Information Technology (IT), Artificial Intelligence (AI), Low-Code and No-Code (LCNC), Large Language Model (LLM), and Visual Basic for Applications (VBA).

Confidence Map

Finding	Confidence	Limiting factors
Systems capability debt is an empirically real organisational condition expressed through governance, delivery-topology, and standards failures.	high	Supported through convergent mechanisms rather than one common quantitative measure.
Systems capability debt generates active workaround demand that later loads central delivery.	medium	Strong directional support; weak direct measurement of manual glue labour.
Central delivery is structurally capacity-constrained and cannot absorb the whole queue quickly.	medium	Constraint is real, but measured partly through proxy and bounded-range estimates.
Split accountability is a second independent mechanism, but it can be partially overcome by explicit integrator rights.	medium	Mechanism is strong; substitute-governance design is still medium-confidence synthesis.
Throughput constraint and split-accountability mechanisms are additive.	medium	Supported across multiple items, but not from one unified quantitative model.
Workaround evolution into LCNC and AI agents is a real pattern.	medium	Strong on LCNC and local automation; weaker on universal sequence claims.
Build mode and do mode are not equivalent in controllability, reliability, or economics.	high	Multiple convergent items; no direct contradiction located.
Do-mode investment creates default displacement risk but can be complementary when explicitly governed as a bridge.	medium	Complementarity evidence is conditional rather than universal.
Do mode carries structurally underestimated operational cost, runtime variance, and lifecycle burden.	medium	Strong directional support; some superlative wording remains unproven.
Vendor-supplied do-mode automation reduces visibility and control.	low	Mechanism supported, but "loses visibility and control entirely" remains too strong for the current evidence.
The strongest supported investment thesis is build-mode-first with tightly bounded do-mode exceptions.	medium	Supported by cost, control, and lifecycle items; precise horizon thresholds remain assumption-sensitive.

Open Questions

• What public or internal datasets could measure the actual rate at which manual workaround demand is converted into formal central Information Technology (IT) backlog items?
• What longitudinal evidence exists on post-gap-closure persistence rates for low-code apps, bots, and agents in live enterprise estates?
• What common denominator would allow a direct matched comparison between post-pipeline build-mode failure rates and production do-mode incident rates for equivalent workflow classes?
• Under what governance designs do explicit integrator rights fully substitute for structural co-location of risk, cost, and benefits, and when do they fail?
• How often does vendor-supplied do-mode automation create materially worse visibility and exit outcomes than internally governed do-mode automation?

The current evidence establishes a strong but qualified thesis: systems capability debt is real, it generates workaround demand, throughput mismatch and split accountability are additive mechanisms, and build mode is materially more controllable and usually more economical than do mode for stable gaps. The evidence also strengthens the thesis by showing that delivery topology and vendor-standards failure are important amplifiers, but it leaves several stronger formulations unproven: there is no robust cross-organisational persistence rate for temporary automations after gap closure, no direct matched distribution comparing build-mode and do-mode failure in production, and no proof that agent systems have the single largest dependency surface of every system type. The conclusion would change materially if future evidence showed either that governed do-mode programmes routinely retire cleanly without explicit lifecycle controls, or that separated governance structures without integrator rights can still close investment trade-offs reliably at scale.

sources

• 2026-05-14-org-failure-modes-accountability-gaps — establishes recurring strategic and delivery-layer failure modes from overlapping or absent accountability.
• 2026-05-14-org-failure-modes-split-risk-cost-benefits-accountability — tests the missing-integrator mechanism directly.
• 2026-05-14-org-failure-modes-project-demand-product-it — supplies evidence for demand-batch mismatch against stable delivery capacity.
• 2026-05-14-org-failure-modes-cohort-demand-domain-it — shows how organisational topology turns demand into queues, handoffs, and long waits.
• 2026-05-14-org-failure-modes-vendor-standards-gaps — adds evidence that vendor-mediated delivery weakens convergence, exit, and control.
• 2026-05-14-citizen-development-rollout-empirical-evidence — gives the empirical rollout pattern for organisation-wide Low-Code and No-Code (LCNC) workarounds.
• 2026-05-16-agent-operational-cost-vs-gap-closure-cost — compares recurring do-mode cost with closing the underlying gap in software.
• 2026-05-16-external-dependency-surface-taxonomy-for-production-llm-agents — maps the operational dependency surface of production Large Language Model (LLM) agents.
• 2026-05-16-do-mode-demand-persistence-and-build-mode-displacement — tests workaround persistence, displacement, and complementarity claims.
• 2026-05-16-it-throughput-constraint-magnitude-and-debt-accumulation-rate — estimates how much workaround demand central delivery can realistically absorb.
• 2026-05-16-variance-control-comparison-across-delivery-modes — compares pre-effect detectability and runtime variance across build mode and do mode.
• 2026-05-16-decommission-trigger-design-for-do-mode-agents — identifies the strongest supported trigger design for retiring temporary bridge agents.
• 2026-05-16-governance-structures-build-mode-without-full-accountability-colocation — qualifies the split-accountability thesis by testing whether explicit integrator rights can substitute for full co-location.