Why the AI Bubble Must Deflate Before Real Business Value Can Emerge

The artificial intelligence industry stands at a precarious crossroads. After years of breathless hype, astronomical valuations, and promises of revolutionary transformation, the gap between expectation and reality has grown too wide to ignore. Thomas Davenport and Randy Bean, two seasoned observers of enterprise technology adoption, have issued a sobering but ultimately optimistic assessment: the AI bubble will deflate in 2026, and this correction represents not catastrophe but opportunity.

Their analysis, published in MIT Sloan Management Review, deserves serious attention from business leaders who have spent the past two years navigating conflicting signals about AI investment. The authors argue that while AI will remain transformative in the long term, we have succumbed to short-term overestimation, and a market correction would provide breathing room for organizations to catch up with technology that has outpaced their ability to absorb it.

This perspective challenges the prevailing narrative that more AI investment, faster deployment, and broader access automatically translate to business value. Instead, it suggests that strategic restraint, organizational maturity, and structural transformation matter more than technological capability. For executives wrestling with AI strategy, this reframing could not come at a more critical moment.

The Inevitability of Market Correction

The comparison between today's AI boom and the dot-com bubble of the late 1990s is uncomfortable but apt. Both periods featured sky-high valuations disconnected from current revenues, emphasis on growth metrics over profitability, breathless media coverage, and massive infrastructure buildouts predicated on future demand. The parallels are striking enough that dismissing bubble concerns requires willful blindness.

Yet the comparison also reveals important differences. Unlike many dot-com companies that had no viable business model, leading AI vendors are delivering real products to actual customers who are paying substantial fees. OpenAI reportedly generates billions in annual revenue. Microsoft, Google, and Amazon are integrating AI capabilities into products used by millions of businesses. The underlying technology works, which was not always true of dot-com ventures.

The problem lies not in whether AI works but in whether current valuations reflect realistic timelines for return on investment. Goldman Sachs Research published a provocative report in mid-2024 titled "Gen AI: Too Much Spend, Too Little Benefit?" that questioned whether the AI industry’s $1 trillion infrastructure investment could ever generate sufficient returns. The report noted that while AI capabilities were impressive, practical applications remained limited and the path to profitability uncertain.

This skepticism has proven prescient. Despite widespread AI adoption, most organizations struggle to demonstrate meaningful financial returns. Randy Bean's 2026 survey found that while 70% of respondents believe the Chief Data Officer role has been successful, only 39% have implemented AI in production at scale, up from just 24% the previous year. Progress is happening, but slowly, and the gap between technological capability and organizational readiness remains vast.

What could trigger the deflation? Davenport and Bean suggest several possibilities: a disappointing earnings report from a major AI vendor, emergence of cheaper competitive alternatives from unexpected sources like China, or spending pullbacks by enterprise customers struggling to justify continued investment. Any of these could puncture market confidence.

The January 2025 DeepSeek incident provides a preview. When the Chinese startup released an AI model that matched leading U.S. systems at a fraction of the cost, Nvidia’s stock briefly plummeted, wiping out hundreds of billions in market value. Though the stock recovered, the episode demonstrated how fragile AI valuations are when fundamental assumptions about competitive moats and economic models are challenged.

For business leaders, the key question is not whether deflation will occur but how to prepare for it. Organizations that have built AI strategies predicated on continued vendor capability improvements at predictable costs may find themselves exposed. Those that have focused on building internal capabilities, developing reusable infrastructure, and pursuing applications with clear ROI will weather the transition more successfully.

The Factory Model: Infrastructure as Competitive Advantage

Amid bubble concerns, a quieter but more consequential trend is emerging among leading organizations: the development of what Davenport and Bean call "AI factories." These are integrated platforms combining technology infrastructure, methodologies, data assets, and reusable algorithms that dramatically accelerate AI development and deployment.

The factory concept originated in banking, where institutions like BBVA and JPMorgan Chase built centralized platforms for developing analytical AI models. These banks faced hundreds of potential use cases across credit decisioning, fraud detection, customer service, and risk management. Rather than treat each application as a standalone project, they created shared infrastructure that made it faster and cheaper to build new models.

The approach has since expanded beyond banking and beyond analytical AI. Procter & Gamble and Intuit have built similar platforms incorporating generative and agentic AI capabilities. Intuit calls its system GenOS, positioning it as an operating system for AI-driven business processes.

The factory model represents a fundamental shift in how organizations think about AI. Rather than viewing it as a collection of point solutions, factory builders see AI as infrastructure requiring platform investment. This shift has profound implications.

First, it creates competitive moats. Organizations with mature AI factories can develop and deploy new applications in weeks rather than months, experiment more freely, and accumulate learning curves that compound over time. Competitors without similar infrastructure must replicate this work for each new initiative, putting them at permanent disadvantage.

Second, it changes the economics of AI investment. The factory model requires substantial upfront investment in platforms, data infrastructure, and talent. But once established, marginal costs for new applications drop dramatically. This makes AI investment look more like traditional infrastructure capital expenditure, with clearer ROI timelines, than the speculative R&D spending that has characterized early AI adoption.

Third, it demands different organizational capabilities. Building and operating AI factories requires platform engineering skills, data engineering expertise, and governance frameworks that many organizations lack. The factory is not just technology but an operating model encompassing processes, roles, and cultural norms.

McKinsey research supports the factory approach. Their analysis of generative AI's economic potential found that the highest-value applications, generating $2.6 trillion to $4.4 trillion annually, come from enterprise use cases in customer operations, software development, and R&D rather than individual productivity enhancement. These enterprise applications require exactly the kind of systematic infrastructure that factories provide.

Yet adoption remains limited. Most organizations still approach AI through isolated pilots or by providing tools to individuals without supporting infrastructure. This approach almost guarantees disappointment. Individual data scientists waste time recreating basic capabilities, struggle to access quality data, and face obstacles deploying models into production. The factory model solves these problems but requires executive commitment and sustained investment.

For leaders, the implication is clear: treating AI as a collection of tactical initiatives rather than strategic infrastructure investment is a recipe for suboptimal returns. The organizations that will thrive in the post-bubble environment are those building systematic capabilities now.

The Enterprise Shift: Moving Beyond Individual Productivity

Perhaps the most important prediction Davenport and Bean offer is that 2026 will mark the shift from generative AI as an individual productivity tool to an enterprise strategic resource. This transition addresses the technology’s most persistent challenge: demonstrating meaningful business value.

When ChatGPT launched in late 2022, organizations responded by making generative AI broadly available to employees. The logic was compelling: the technology was easy to use, required minimal training, and offered potential productivity benefits across virtually every role. Microsoft’s Copilot became ubiquitous, helping employees write emails, create presentations, and analyze spreadsheets.

The problem is that these individual applications generate only incremental, largely unmeasurable productivity gains. Yes, an employee might save 30 minutes drafting a memo or 15 minutes formatting a spreadsheet. But what do they do with that saved time? Without strategic direction, time savings dissipate into other low-value activities rather than accumulating into meaningful business impact.

Johnson & Johnson’s experience illustrates this challenge. The company initially pursued 900 different AI use cases across its organization, most focused on individual or departmental productivity enhancement. After finding this approach delivered minimal measurable value, J&J pivoted to focusing on a handful of strategic, enterprise-level applications with clearer ROI.

This mirrors broader research findings. A Harvard Business Review study found that only 10% of companies achieved significant financial returns from AI investments. The primary barriers were not technological but organizational: unclear strategy, inadequate change management, and failure to redesign work processes around new capabilities.

The enterprise approach addresses these limitations by focusing AI on strategic processes with measurable business outcomes. Rather than helping individual salespeople write better emails, enterprise applications might optimize inventory across global supply chains, accelerate drug discovery through systematic literature analysis, or personalize customer experiences at scale.

These applications are harder to build. They require cross-functional collaboration, significant data preparation, process redesign, and change management. But when successful, they generate quantifiable business value: reduced inventory costs, faster time to market, increased conversion rates.

Sanofi’s approach provides a useful middle ground. The pharmaceutical company created a competition where frontline employees pitch AI project ideas, with winning proposals receiving funding as enterprise-level initiatives. This preserves bottom-up innovation while ensuring projects have strategic scope and executive sponsorship.

The shift to enterprise applications also addresses the measurement problem. Individual productivity gains are notoriously difficult to quantify. Did that employee who used ChatGPT to draft an email really save time, or did they simply spend longer on other tasks? Enterprise applications, by contrast, generate measurable outcomes: units produced, errors reduced, revenue increased.

Yet this transition creates tension. Employees have come to view AI access as a workplace benefit, important for satisfaction and retention. Organizations that pull back from broad availability may face pushback. The solution is not to eliminate individual access but to be realistic about its value and strategic about where to invest for maximum return.

For executives, this means rethinking AI governance. Rather than treating AI primarily as an IT enablement question, it becomes a strategic priority requiring C-suite attention. Which processes offer the highest potential value? What organizational changes are needed to capture that value? How should success be measured? These questions demand business leadership, not just technical expertise.

The Agentic AI Reality Check

If generative AI represented 2023’s hype cycle, agentic AI claimed that distinction for 2024 and into 2025. The promise was compelling: AI agents that could autonomously complete complex multi-step tasks, make decisions, and interact with various systems to accomplish business objectives. Vendors promoted visions of AI agents handling customer service, managing supply chains, and executing business processes with minimal human oversight.

Davenport and Bean provide a necessary corrective: agentic AI is not ready for prime-time business applications. Research from Anthropic, a leading AI laboratory, found that agents made errors in 30 to 40% of complex tasks. Carnegie Mellon experiments revealed similar limitations. When money or reputation is at stake, these error rates are unacceptable.

Beyond accuracy problems, agents face cybersecurity vulnerabilities, particularly prompt injection attacks where malicious actors manipulate agent behavior through crafted inputs. More troubling are findings that agents can become deceptive or misaligned with stated objectives when pursuing goals, raising fundamental questions about reliability.

These limitations suggest that agentic AI has fallen into what Gartner calls the "trough of disillusionment," the phase following initial hype when practical limitations become apparent. This is a normal part of technology adoption, but organizations that invested heavily based on vendor promises may face uncomfortable reckonings.

Yet Davenport and Bean remain optimistic about agentic AI's long-term potential, predicting that agents will handle most transactions in many business processes within five years. This timeline is more conservative than some vendor projections but more optimistic than skeptics like OpenAI cofounder Andrej Karpathy, who estimates a decade before agents work reliably.

The key is treating agentic AI as an emerging capability requiring patient investment rather than an immediately deployable solution. Organizations should begin building trusted agents for low-stakes applications, developing internal capabilities for creating and testing agents, and piloting interorganizational agents with cooperative partners.

This measured approach contrasts sharply with the all-in betting that characterized early generative AI adoption. It reflects growing sophistication among business leaders about technology maturity and realistic deployment timelines. The organizations that succeed with agentic AI will be those that invest in capability building while maintaining realistic expectations about near-term returns.

The Leadership Structure Question

The final trend Davenport and Bean identify may be the most consequential for how organizations capture AI value: the ongoing confusion about who should manage AI and where that role should report. Their research found that 39% of large organizations have named Chief AI Officers, but these leaders report to widely varying functions: chief data officers (30%), business leadership (27%), technology leadership (34%), or transformation offices (9%).

This structural confusion is not merely an organizational chart issue. Reporting relationships shape priorities, resource allocation, and strategic integration. A CAIO reporting to the CTO may focus on technical implementation and infrastructure. One reporting to business leadership may emphasize use case delivery and revenue impact. One reporting to a CDO may prioritize data quality and governance.

Davenport and Bean argue, correctly, that AI leadership should report to chief data officers, creating integrated responsibility for data, analytics, and AI. This makes conceptual sense: AI depends on quality data, analytical capabilities inform AI applications, and governance frameworks must span the entire data-to-insight value chain.

Yet the diverse reporting structures reflect genuine strategic tensions. Some organizations view AI primarily as a business transformation tool, justifying business-side ownership. Others see it as technical infrastructure, supporting IT ownership. Still others position it as a standalone strategic capability requiring C-suite reporting.

The lack of consensus may actually reflect the technology’s immaturity rather than cause performance problems. Early in any major technology adoption cycle, organizational structures remain fluid as companies experiment with different approaches. Over time, best practices emerge and structures converge.

What seems clear is that splitting responsibility across multiple leaders creates coordination problems. When different executives own data strategy, AI development, and technology infrastructure, initiatives fall into gaps between functions. Integrated leadership under a Chief Data, Analytics, and AI Officer, as Davenport and Bean advocate, provides clear accountability and facilitates systematic capability building.

Yet this structure requires capabilities many organizations lack. It demands leaders who understand data engineering, analytical methods, AI technologies, business strategy, and change management. This rare combination explains why many organizations settle for more fragmented structures: they simply cannot find leaders capable of spanning all these domains.

The leadership question also intersects with the factory model and enterprise application shift. Building AI factories requires sustained platform investment that spans data, analytics, and AI. Deploying enterprise applications requires cross-functional coordination and business integration. Both are easier with integrated leadership than with fragmented accountability.

Preparing for the Post-Bubble Reality

If Davenport and Bean’s analysis proves correct, business leaders face a pivotal transition in 2026. The deflation of AI valuations, while potentially painful for financial markets, could create opportunity for organizations willing to focus on sustainable value creation over speculative positioning.

The prescription is clear: shift from reactive tool adoption to systematic capability building. This means investing in AI factory infrastructure, transitioning from individual productivity applications to enterprise strategic use cases, maintaining realistic expectations about emerging capabilities like agentic AI, and establishing integrated leadership for data and AI.

These recommendations may seem conservative compared to the revolutionary rhetoric that has dominated AI discourse. But conservatism may be exactly what this moment requires. The technology is real, the potential is enormous, but the path from capability to value runs through organizational transformation, strategic focus, and patient investment in fundamentals.

The organizations that thrive in the post-bubble environment will be those that resisted the temptation to chase every AI trend and instead built systematic capabilities aligned with strategic priorities. They will have invested in data infrastructure, developed reusable platforms, focused on measurable business outcomes, and established governance frameworks that balance innovation with risk management.

For executives, this analysis offers both warning and opportunity. The warning is that current AI investment levels and expectations are unsustainable and that market correction is likely. The opportunity is that this correction will separate organizations pursuing sustainable value from those chasing hype, creating competitive advantage for those who prepared.

The immediate question facing leaders is not whether to invest in AI but how to invest strategically. This requires moving beyond reactive tool adoption to systematic capability building, beyond individual productivity to enterprise transformation, and beyond technological enthusiasm to hardheaded assessment of organizational readiness and realistic timelines.

The AI revolution is real, but revolutions are measured in decades, not quarters. Organizations that align their strategies with this longer timeline, building capabilities systematically while maintaining realistic expectations, will emerge from the coming transition stronger and better positioned for sustainable competitive advantage. Those that bet on continued hype may find 2026 a very challenging year indeed.

For further exploration of AI trends and how these changes might impact your organization, check out this insightful article.