Generative AI reshuffles the make-or-buy ledger but stops short of a SaaSocalypse: firms will increasingly build commodity utilities and AI-differentiating applications in-house, yet regulated and mission-critical systems are still likely to be purchased from vendors.

Main Finding

Agentic generative AI materially alters the economics of the enterprise make-or-buy decision, but it does not produce a wholesale “SaaSocalypse.” AI sharply lowers the cost and time barriers to in-house development for certain application classes (especially commodity utilities and bespoke differentiating apps), yet lifecycle operational costs, compliance burdens, and new dependencies on external AI infrastructure mean most regulated and mission‑critical systems remain more economically and institutionally suited to buying from established vendors. Crucially, the Make option itself transforms into a hybrid governance form: firms may own code but remain dependent on external AI platforms and compute, creating different trade-offs than pre-AI in-house development.

Key Points

• Conceptual approach: The paper re-evaluates seven canonical MoB determinants (cost/TCO, strategic differentiation, asset specificity, vendor lock‑in, time‑to‑market, quality & compliance, organizational capability) in light of current agentic AI capabilities and limits.
• Cost and TCO:
  • AI substantially reduces initial development (CAPEX) and time-to-market for well-specified, modular tasks.
  • Operations, maintenance, compliance, and monitoring still dominate lifecycle cost; AI may increase OPEX due to model governance, monitoring, and compute costs.
• Strategic differentiation:
  • The locus of strategic advantage shifts toward AI-orchestration capabilities (data, model‑management, and integrating AI into business processes) rather than routine coding skill.
• Asset specificity & customization:
  • AI makes customization cheaper for pattern-rich, well-documented domains but struggles with tacit knowledge, legacy entanglement, and highly coupled systems.
• Vendor lock‑in:
  • SaaS vendors can embed AI to deepen personalization and workflow integration, increasing switching costs; firms that “make” may trade vendor lock‑in for dependency on external AI infrastructure (cloud+model providers).
• Time-to-market & agility:
  • Agentic systems accelerate prototyping and feature delivery, favoring in‑house Make for short‑lived/fast‑changing features; vendors also adopt AI, narrowing this advantage over time.
• Quality, reliability & compliance:
  • AI-generated code is effective for standard functionality but raises concerns for high-stakes correctness, auditability, explainability, and regulatory compliance.
• Organizational capability:
  • New required capabilities center on AI orchestration, prompt/agent design, AI governance, data engineering, and vendor/infrastructure management—not just traditional software engineering.
• Application typology (sensitivity to AI-induced shifts):
• Commodity utilities: Highly susceptible to Make (cheap to build with AI).
• Differentiating custom applications: Often attractive to Make if AI‑orchestration is a core strategic capability.
• Regulated standard applications: Continue to favor Buy due to compliance and auditability.
• Mission‑critical systems of record: Largely remain Buy because of high reliability, certification, and long-tail operational complexity.
• Governance transformation:
  • The Make governance form evolves from Williamson’s pure hierarchy to a hybrid: code ownership coexists with reliance on external AI models, APIs, and cloud compute. This hybrid has distinct economics, risk profiles, and governance needs (e.g., contracts with model providers, redundancy strategies, data portability).
• Overall conclusion: The SaaSocalypse narrative is overstated; AI redraws boundaries and trade-offs but yields a mixed, application‑specific outcome rather than universal insourcing.

Data & Methods

• Methodological approach: Conceptual research (Jaakkola, 2020). Deductive synthesis combining transaction cost economics (TCE) and the resource‑based view (RBV) with a detailed assessment of emergent agentic AI capabilities and limitations.
• Evidence sources and inputs:
  • IS theory and empirical syntheses (Williamson; Barney; Lacity et al.; Mata et al.; pace‑layering literature).
  • Literature on generative AI and agentic coding systems, including model and tool descriptions (Anthropic, OpenAI, GitHub Copilot, Google Gemini, SWE‑agent, AutoGen, LangChain, etc.).
  • Empirical proxies from recent studies and industry signals: productivity studies for copilot tools (e.g., Peng et al., 2023), adoption analysis of agentic tooling in public GitHub projects (~129k projects; Robbes et al., 2026), SWE‑bench benchmark results, market reactions (S&P Software & Services index movements; TechCrunch/Bloomberg reporting).
  • Secondary sources on vendor strategies, AI infrastructure economics, and compliance concerns.
• Analysis approach: Factor‑level re‑evaluation of the seven canonical MoB determinants; development of an application typology and a revised binary decision framework; conceptual modeling of governance shift toward hybrid forms.
• Limitations acknowledged by the author:
  • No primary empirical study or large‑scale causal inference; field is nascent and fast‑moving, so findings are provisional and theory‑driven.
  • Outcomes depend on rapid evolutions in model capabilities, costs of compute, vendor strategies, and regulation.

Implications for AI Economics

• TCO and pricing dynamics:
  • Upfront development costs decline for many applications, but lifecycle OPEX (model inference costs, monitoring, data pipelines, compliance work) becomes a larger share of total cost. Economic advantage may therefore accrue to providers who can amortize OPEX at scale.
  • SaaS vendors retain economic defensibility via scale economies in operations, certification, and amortized R&D—AI does not automatically erase these advantages.
• Market structure and competition:
  • Two countervailing forces: easier in‑house building lowers barriers to entry for internal teams and startups in niche/commodity segments; conversely, platform and model providers (cloud + foundation model vendors) gain leverage because both buyers and builders depend on their compute/models.
  • Expect consolidation around model/infrastructure providers and differentiated SaaS incumbents that embed AI deeply into workflows, increasing platform lock‑in.
• Firms’ investment allocation:
  • Firms should reassess CAPEX/OPEX mixes: invest less in traditional large developer teams for routine apps, but allocate more to data engineering, AI governance, model ops, and vendor management.
  • Strategic investment in AI‑orchestration capability (data assets, model fine‑tuning, agent design, governance) becomes central to capturing differentiated value.
• Labor and skills economics:
  • Demand shifts from bulk coding labor toward higher‑value roles: AI prompts/agent engineering, systems integration, ML ops, compliance officers for models, and vendor/infrastructure contracting specialists.
• Procurement and contracting:
  • Contract design must evolve to address hybrid dependencies: SLAs for model availability, data portability clauses, model provenance, audit rights, and pricing for inference/throughput.
  • Firms buying SaaS should expect vendors to offer deeper AI‑driven integration and personalization—even if firms can build basic versions themselves.
• Policy and regulatory implications:
  • Regulatory regimes (financial, healthcare, telecom) will slow insourcing for regulated apps due to compliance/audit needs; regulators should consider standards for model provenance, testing, and certification.
  • Antitrust and data‑portability debates may intensify around model and compute provider dominance.
• Strategic takeaways for practitioners:
  • Use an application‑specific MoB calculus: favor Make for commodity and short‑lifecycle innovation projects; favor Buy for heavily regulated, mission‑critical, or highly coupled legacy systems.
  • If Make is chosen, treat it as hybrid governance: plan for external AI infra risk, contractual protections, redundancy, and strong AI governance.
  • Build AI‑orchestration capabilities as a distinct strategic asset—those capabilities, not raw coding, are likely to drive differential advantage.

Caveat: The conclusions are theory‑driven and contingent on rapid advances in model capabilities, costs of compute, vendor strategies, and regulatory responses. Empirical validation as practices mature is necessary to refine the framework.