Leaders debate build-vs-buy for enterprise AI agents

At the Ai4 2026 conference in Las Vegas, technology executives mapped out the increasingly complex decision of whether to build proprietary AI agents or purchase third-party solutions. While early generative AI adoption favored buying models from providers like OpenAI or Anthropic, the rise of agentic AI has forced organizations to weigh their core competencies, data security, and infrastructure before committing.

For companies where customer experience is a primary differentiator, building in-house is often essential. Ningyu Chen, senior vice president of technology at Home Depot, emphasized that the retailer would never outsource its core customer experience, even as it partners with Google, Microsoft, Anthropic, and OpenAI. Chen advocates using an abstraction layer to remain flexible. Santhi Ramesh, CEO of Future Propel and former AI leader at The Hershey Company and Ferrero, noted that non-technical firms should buy unless they handle massive, highly sensitive datasets. Rachel Ibarra of Cardinal Group Companies, which uses an in-house agent named Stan, warned that buying external platforms can accumulate operational debt.

Infrastructure and model flexibility represent additional hurdles. Markus McKay-Fleisch of Smartsheet highlighted the need for strong hosting infrastructure and guardrails. In practice, companies must also swap models dynamically. Mahe Bayireddi, CEO of Phenom, noted that his firm utilizes open models from France, North America, and China, switching them on the fly. For smaller operations, building can offer precise cost controls. Steve Toy, creator of the nutrition app Just a Bite Better, built a custom model garden via APIs rather than relying on platforms from IBM, PayPal, Google, AWS, or Microsoft. Toy uses a hard cap on token usage to prevent waking up to a $50,000 bill.

For AI practitioners, this shifting landscape means the binary choice of building or buying has dissolved into a hybrid spectrum. Developers must design flexible architectures that abstract underlying models, allowing organizations to swap APIs dynamically as market dominance shifts. Furthermore, practitioners must implement strict cost-governance tools and robust hosting guardrails to protect their unique data moats without incurring runaway operational expenses.