The AI Productivity Paradox: Efficiency Gains Are Not Enough

Generative artificial intelligence (GenAI) has demonstrated measurable productivity gains in specific workplace tasks, but these improvements do not necessarily translate into stronger enterprise-level performance. Evidence is particularly strong for GenAI applications that support writing, coding, customer service, and other knowledge-intensive tasks. In a study of 5,172 customer support agents, Brynjolfsson, Li, and Raymond (2025) found that a GPT-based conversational assistant increased productivity by 15% on average, measured by customer issues resolved per hour. The findings demonstrate how GenAI can improve performance within specific workplace activities. However, the International Labour Organization (2026) observes that substantial productivity gains at the task and worker levels have not consistently translated into measurable improvements at the firm, sectoral, or macroeconomic levels.

This divergence reflects a broader AI productivity paradox: productivity improvements observed in individual tasks and workflows have not consistently translated into measurable gains at the organizational level.  This article examines evidence of GenAI-related performance gains, the challenges of translating these improvements into enterprise-level outcomes, and the economic implications for enterprise customers and GenAI providers.

1. GenAI Performance Gains Vary Across Applications and Metrics

Generative AI has demonstrated measurable performance improvements in specific applications, although the nature and magnitude of these effects vary across tasks, users, and performance metrics. Evidence from software development and online retail illustrates improvements in task efficiency and commercial outcomes, while other studies document cases in which AI assistance increases the time required to complete work (Reuters, 2025; Financial Times, 2025).

In software development, GenAI can improve task efficiency. According to Reuters (2025), JPMorgan Chase reported that its internal AI coding assistant helped engineers complete work 10% to 20% faster. JPMorgan identified opportunities to redirect the time saved toward additional engineering projects. The reported gains indicate faster execution of engineering tasks, rather than a measured increase in firm-wide productivity or financial performance.

In customer-facing applications, field experiments have also documented improvements in specific commercial outcomes. The Financial Times (2025) reported on field experiments in online retail in which a GenAI presale assistant increased sales by 16.3% and conversion rates by 21.7%. A hybrid AI–human customer service system also increased sales by 11.5% compared with human agents alone. These findings demonstrate improvements in sales performance within the tested applications, but do not establish corresponding improvements in labour productivity or profitability.

However, these performance gains are not universal. In a randomized study involving 16 experienced open-source developers, METR found that access to early-2025 AI tools increased task completion time by 19% on average.  Taken together, these findings show that GenAI's effects vary across applications, user groups, and performance metrics. While some deployments improve task efficiency or commercial outcomes, others deliver limited or negative results. Importantly, improvements within specific tasks or applications do not, by themselves, establish enterprise-wide productivity gains.

2. From GenAI Adoption to Enterprise Performance: Three Gaps in Value Realization

Generative AI adoption is expanding, but measurable benefits remain uneven across enterprises. Research highlights three important distinctions: adoption does not necessarily translate into measurable efficiency improvements; task-level gains do not automatically translate into firm-wide productivity; and higher productivity does not, by itself, establish stronger financial performance.

First, widespread GenAI adoption does not necessarily result in measurable efficiency improvements. According to Reuters (2026), a Bpifrance survey of 534 French mid-sized business leaders found that 77% reported using GenAI, while only 17% of adopters reported time savings. These findings suggest that GenAI adoption can advance faster than measurable efficiency benefits, although the reported time savings are based on respondents' assessments rather than independently measured productivity outcomes.

GenAI Adoption and Reported Time Savings Among French Mid-Sized Businesses, 2026 (%) – Source: Bpifrance, via Reuters (2026)

Second, task-level efficiency gains do not automatically translate into broader organizational productivity. A Gartner (2025) survey of supply chain organizations found that GenAI saved desk-based employees an average of 4.11 hours per week, with improvements associated with higher output and work quality. However, at the team level, reported time savings declined to 1.5 hours per member per week, with no corresponding correlation to improved output or quality. These findings illustrate how individual efficiency gains may diminish when assessed across broader organizational activities.

Third, higher productivity does not, by itself, establish stronger financial performance. A company may increase output per employee without realizing corresponding improvements in profitability if additional capacity does not generate incremental revenue or cost savings, or if these benefits are offset by AI-related expenses. Evidence from broader AI adoption also illustrates the distinction between operational and financial outcomes. According to McKinsey & Company (2025), organizations reported cost reductions and revenue increases from individual AI use cases, yet only 39% of respondents reported any enterprise-level EBIT impact attributable to AI. Although the survey does not directly measure the conversion of productivity gains into EBIT, it illustrates why use-case benefits should not be treated as evidence of firm-wide profitability improvements.

3. The Economics of GenAI: Customer Value and Provider Profitability

Demonstrated productivity gains provide evidence of GenAI's operational effectiveness, but economic value depends on how these improvements affect both enterprise customers and technology providers. Two questions are central: whether customers realize financial benefits after deployment and operating costs, and whether providers can generate sustainable gross profit from delivering those benefits. 

For enterprise customers, the relevant measure is net economic benefit rather than efficiency alone. Faster task completion or increased output creates financial value when it leads to realized cost savings, additional revenue, or other measurable benefits that exceed the costs of adopting and operating AI. These costs extend beyond software subscriptions to include integration, training, oversight, and ongoing usage. As the The Times (2026) reports, usage-based AI costs can become difficult to predict, particularly as agentic applications increase computing consumption. Consequently, lower costs per AI interaction do not necessarily guarantee lower total expenditure or positive returns for enterprise customers. 

For GenAI providers, revenue growth must be evaluated against the cost of delivering AI services. Unlike traditional software applications, GenAI services can incur substantial variable inference costs with each interaction. According to Vista Equity Partners (2026), these costs can materially influence how much revenue from AI applications translates into gross profit. Provider economics therefore depend not only on recurring customer demand and pricing, but also on inference efficiency, infrastructure costs, and the ability to maintain gross margins as usage scales. 

Importantly, economic value created for customers is not necessarily captured by providers. Competition, declining token prices, and alternative models can allow customers to retain a larger share of productivity-related benefits, while limiting the prices providers can charge. The Financial Times (2026) reports that declining token prices reflect a combination of technical improvements, increased competition, and shifts toward lower-cost models, with different implications for provider margins. As a result, strong customer returns do not automatically imply equally strong provider profitability. 

Together, these considerations distinguish customer value realization from provider value capture. For investors evaluating enterprise GenAI businesses, these distinctions provide a useful starting point for assessing commercial viability, although they do not, on their own, establish investment attractiveness. 

References

Financial Times (2026). What’s really going on with AI token price deflation? https://www.ft.com/content/96184ef4-05be-467c-9d9c-745e21dad31a

Gartner (2025). Gartner survey shows supply chain GenAI productivity gains at individual level, while creating new complications for organizations. https://www.gartner.com/en/newsroom/press-releases/2025-02-05-gartner-survey-supply-chain-genai-productivity-gains-at-individual-level-while-creating-new-complications-for-organizations

McKinsey & Company. (2025). The state of AI in 2025: Agents, innovation, and transformation. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2025

Reuters (2026). French mid-sized firms adopt AI but see few gains, survey shows.https://www.reuters.com/technology/french-mid-sized-firms-adopt-ai-see-few-gains-survey-shows-2026-06-23/

The Times (2026). Businesses are hooked on AI. Do they know how much it will cost?https://www.thetimes.com/business/entrepreneurs/article/ai-business-costs-enterprise-network-db0tc080g

Vista Equity Partners (2026). Understanding inference and the economics of enterprise AI.https://www.vistaequitypartners.com/insights/inference-economics-enterprise-ai/