Why AI Productivity Gains Do Not Reach the P&L

Business leaders are increasingly asking what margin improvements they should expect from artificial intelligence. The answers vary sharply depending on where they come from. Vendors point to rapid efficiency gains and reduced marginal cost. Operational leaders, by contrast, often struggle to see those gains reflected in financial performance.

A (Solow*) paradox is emerging. Leaders of AI initiatives that appear compelling on paper need to think differently to deliver predictable returns once they are deployed at scale. This is not because the technology under performs it is because the standard business assumptions used to justify AI investments need to change.

Anekanta®’s article examines why AI efficiency gains struggle to translate into economic value, why this is a structural rather than technical problem, and what operating conditions must be in place for AI investments to deliver sustained return.

The Productivity Mirage: When Time Saved Is Not Value Created

In enterprise environments, personal productivity tools can be justified on the basis of “minutes saved per day”. On the surface, recent UK Government trials appear to support this logic. The Department for Work and Pensions (DWP) reported average time savings of 19 minutes per day from the use of generative AI tools, while a broader cross-government trial reported savings of 26 minutes.

By contrast, another evaluation by the Department for Business and Trade (DBT) found no evidence of productivity gains, despite high levels of user satisfaction. In several task categories, time spent actually increased, and AI-generated outputs were often discarded. Arguably, the DBT trial may have been conducted using earlier less mature Gen AI models and this reflected in the results. Nevertheless, the discrepancy is not explained by differences in technology. It is explained by what happens after an output is generated.

Gen AI produces probabilistic outputs which are the most likely answers to the questions posed rather than verified facts. In operational contexts, those results must be verified, interpreted and potentially reworked before they can be relied upon for business decisions. Where decision rights are unclear, guidance is absent, or contextual understanding is weak, the time spent validating outputs, a verification tax, can exceed the time saved in generation.

In these conditions, AI improves the experience of work without improving the economics of work. The productivity benefit exists in theory, but this does not reflect through into the organisation’s financial statements.

Crucially, this is not a user problem. It is an operating model problem.

Why AI Breaks Traditional ROI Logic

Traditional enterprise software is largely deterministic. Given the same inputs, it produces the same outputs. Risk typically materialises outside the system though various sources which could include poor training, insecure design, or misuse. AI systems, particularly those which contain Gen AI may violate this assumption. The same question may generate different answers. Systems may change behaviour over time as data, context and usage evolve. Increasingly, AI systems can act autonomously, with or without explicit human decision points before the next autonomous action – in the case of agentic AI driven workflows.

As a result, risk is both internal and external to the system. It arises not only from how AI is used, but from how it behaves. This distinction matters economically. With traditional software, cost and risk can be mostly addressed before deployment. With AI, a material portion of cost, risk and potential liability emerges during operation. Drift, misalignment, inappropriate inference and loss of control are not necessarily build-time defects, they are lifecycle effects.

When organisations apply static software economics to variable systems, the know path to ROI predictability is broken.

What Actually Determines AI ROI

If AI is a business system defined by variability, then the factors that determine its financial return are managerial, not necessarily technical. Across organisations that successfully scale AI, four operating conditions consistently determine outcomes. They are economic necessities.

Senior Decision Rights (Strategy)

The use of AI systems must be governed by clear, senior-level decisions about purpose, scope and acceptable use. Without this, AI outputs may be embedded into workflows without clarity over the provenance of outputs, reliability or consequence. The result could be silent erosion of value – rework, withdrawal, reputational exposure and, ultimately, stranded investment. The critical leadership questions are simple:

What AI use cases are we authorising in this organisation – and what do we explicitly not authorise?

Risk Intelligence (Context, Not Code)

It is clear that AI risk does not arise solely from the models alone, but from how outputs interact with real-world decisions. Effective organisations move risk analysis upstream in the decision chain and focus on context to avoid value creation turning into value destruction. They treat drift, autonomy and inference as commercial variables that must be observed and constrained over time. This requires visibility of where AI is used, including AI embedded in vendor software, and the ability to test, measure and intervene when behaviour moves outside acceptable bounds.

Risk intelligence, applied early, reduces lifecycle cost. Applied late, it amplifies lifecycle cost.

Human Authority (Literacy + Expertise)

AI literacy is necessary but alone it is insufficient. Understanding how AI works does not guarantee correct use. What protects value is human experience, subject-matter expertise combined with decision rights. Vendor training can explain features. It cannot determine whether an output is correct, appropriate or safe to rely on in context.

Where humans lack the authority to challenge or override AI outputs, monitoring becomes symbolic and risk accumulates invisibly. Where they do, AI becomes a controlled asset rather than an uncontrolled influence.

Governance (Lifecycle Control)

Governance is not an administrative overlay. It is the mechanism through which AI value is preserved over time. Clear ownership, auditability, review cycles and escalation paths allow organisations to intervene when systems drift, constrain use when assumptions break, and scale confidently when performance holds.

Without governance, AI value leaks, margins erode, pilots stall, and intervention becomes reactive rather than deliberate. With it, AI moves from experimentation to enterprise value.

The Cost Question – Why AI Requires Lifecycle Provisions

The unpredictability of AI ROI is often framed as an adoption problem. In reality, it is a cost-allocation problem. Traditional software economics assume that most risk is addressed before deployment, with ongoing costs confined largely to maintenance and cybersecurity threat management. AI systems invalidate this assumption. Because behaviour evolves, cost must be provisioned across the lifecycle.

Strategy, risk intelligence, human oversight and governance are not discretionary add-ons. They are the mechanisms through which variability is managed. When benchmarked against established enterprise software practices, the resulting cost profile is not excessive – but it is redistributed. Less is spent correcting failure late in the lifecycle, and reasonable budgets are allocated early and to managing it continuously.

For boards, CFOs and CTOs, the implication is not that AI is more expensive. It is that AI must be costed and resourced differently if returns are to be predictable.

A Brief Case Reflection: Isolation vs Integration

The Dutch childcare benefits scandal remains a stark illustration of what happens when automated decision-making is deployed without effective human challenge or proportional controls. A system operating as a ‘black box’ generated systemic financial and political fallout, not because automation existed, but because authority and oversight did not.

By contrast, organisations that integrate strategy, risk, human authority and governance into the AI lifecycle follow a different trajectory. They treat AI governance as organisational and IT infrastructure, not intervention.

As a result, AI scales as a managed capability rather than an unbudgeted shock arising at an unexpected time.

AI is Not Cheaper Software – Many Miss This Point

AI ROI is not uncertain because organisations lack ambition or capability. It is because systems defined by variability are being managed as if they were static. Strategy, risk intelligence, human authority and governance are not overheads imposed on AI. They are the operating conditions under which AI becomes economically viable at scale. Organisations that treat these elements as optional will continue to experience volatile returns, stalled deployments and unplanned intervention. They need to look beyond the individual licence costs and treat AI adoption as a organisational culture change driven from the top down and supported from the bottom up through appropriate training.

Those that embed the right operating conditions into senior decision-making replace experimentation with control and uncertainty with predictability. The productivity dividend from AI becomes real.

Practical Steps Towards Operational Integration

Taking the steps which translate high-level concepts into practical management solutions is not as daunting as it was 3 years ago. Frameworks and entire governance systems embodied in regulation alongside certifiable standards have emerged since Anekanta® first developed its 12 principle AI Governance Framework in 2020* (applied within this article). The ecosystem which supports implementation is growing in numbers, strength and capability, as are the maturity of the standards and regulations.

To find out more, explore Anekanta®’s resource centres:

Anekanta® EU AI Act Resource Centre

Anekanta® ISO/IEC 42001 Resource Centre

Anekanta® AI Governance Framework for Boards**

Please do get in touch if you need our help.

*Solow, Robert. (1987). “We’d Better Watch Out,” New York Times Book Review, July 12, 1987, p. 36 Available at: https://www.semanticscholar.org/paper/We%E2%80%99d-better-watch-out-Solow/cef149b3dbdaa85f74b114c2c7832982f23bcbf0?sort=pub-date

** Copyright Anekanta® 2020-2026. Available for use under a creative commons licence with permission from Anekanta®. Enquire for details.

Anekanta®AI and Anekanta®Consulting
AI Strategy | Risk | Literacy | Governance
Contact us | Explore our services | Subscribe to our newsletter | Follow us on LinkedIn

Intellectual Property: © 2016–2026 Anekanta®. All rights reserved. Unless otherwise expressly stated, all materials published on this website, including the Anekanta® AI Governance Framework for Boards, the 12 Principles, and all AI risk and impact evaluation methodologies, software, models, diagrams, text and materials, are proprietary intellectual property of Anekanta®. No reproduction, adaptation, distribution, or commercial exploitation is permitted without prior written authorisation. No rights are granted other than those expressly stated. The Anekanta® AI Governance Framework and 12 Principles are developed, maintained and continuously enhanced as part of Anekanta®’s proprietary governance architecture.
Professional Disclaimer: The information provided on this website is for general informational purposes only and does not constitute legal, regulatory, financial or professional advice. Any reliance placed on the information is strictly at the user’s own risk. Professional advice should be sought in relation to specific circumstances through a formal engagement with Anekanta®.
Use of Generative AI: Generative AI tools may be utilised in research and drafting processes. All published materials are subject to substantive human review, professional judgment and oversight prior to release.