Why Modular Organizations Struggle to Recombine AI Capabilities

Organizations spent decades perfecting modular design—breaking themselves into agile teams, independent business units, and platform architectures to gain flexibility and speed. Yet a new constraint has emerged: Companies can decompose work far more easily than they can recombine those pieces when conditions change.

Researchers Wei Wei, Katherine Xin, George S. Yip, and Mark J. Greeven identify this as the "recombination bottleneck." Their research across firms in Asia, Latin America, Europe, and North America reveals that integrating a new AI capability now takes quarters rather than weeks. Moving a fraud-detection model or demand-forecasting engine from one business unit to another requires data governance alignment, regulatory review, model-risk approval, and incentive structures that encourage sharing rather than hoarding.

The bottleneck has shifted from organizational design to operational execution.

Why it matters

Three forces make this bottleneck increasingly costly. AI creates recombination opportunities faster than most firms can act on them. Ecosystems are becoming the competitive unit, requiring companies to plug in new partners in days rather than months. Strategic windows are compressing as AI diffusion and geopolitical volatility shrink response times. A modular firm that needs six months to reconfigure may as well not be modular at all.

Four strategies for composable integration

The researchers propose "composable integration"—the operational capability to recombine existing modules at speed. Four moves define this approach:

Make recombination a managerial skill. Most firms treat integration as a one-off event rather than a continuous competence. JPMorgan Chase exemplifies the alternative: The bank released LLM Suite, a unified generative AI platform, to eligible employees in summer 2024. Within eight months, 200,000 employees were onboarded. Instead of allowing each unit to build bespoke AI tools in isolation, the bank created a secure common platform that teams could use and extend for different tasks.

Standardize the handshake, not just the interface. The fastest firms define explicit rules for how components connect—which data can flow between units, what governance applies when a new partner plugs in, which compliance standards must be met. Amazon Web Services' 2002 API mandate requiring all internal teams to expose services through standardized interfaces was a governance decision that created connective tissue for reuse at scale. Mastercard's "gold standard" scorecard for developer experience set a measurable quality bar that made partner integration repeatable.

Turn shared infrastructure into reusable capabilities. India's Reliance Jio built a programmable, all-IP network from the outset with standardized interfaces across deployment, customer onboarding, and analytics. Those interfaces allowed the company to extend the platform from connectivity into content, payments, commerce, and cloud services without redesigning the organization each time.

Push recombination authority to the edge. If every integration decision requires senior escalation, recombination will always be slow. Siemens Xcelerator, an open digital business platform, enables plant teams to combine energy data, machine performance, simulation tools, and maintenance workflows without waiting for central redesign. Headquarters shifts from approving every connection to designing the rules within which others can connect.

The new competitive frontier

The competitive advantage has shifted from who has the best-designed components to who can recombine them fastest under pressure. For most organizations, composable integration remains the capability they have not yet built.

These findings were first reported by Wei Wei, Katherine Xin, George S. Yip, and Mark J. Greeven in Harvard Business Review.