Enterprise AI resilience matters more than speed, Crafting CEO says

Enterprise AI resilience is moving to the front of the AI planning debate as companies build more of their operations on outside models, according to Sumeet Vaidya, CEO and co-founder of Crafting. For investors watching the AI buildout, his argument is a reminder that adoption is not only about faster tools, it is also about cost control, vendor risk and reliability.

Vaidya pointed to recent uncertainty around major AI providers, citing an Anthropic policy reversal reported by NBC News and a security incident involving Hugging Face and OpenAI that OpenAI described in a company post. He argued that businesses using paid models from Anthropic, OpenAI and similar providers need those systems to be dependable because they may sit underneath real products and internal workflows.

At the same time, Vaidya said open-source groups such as OpenClaw and DeepSeek are offering no-cost models with quality that is becoming more competitive. In his view, the old reasons large companies avoided open-source AI, including gaps in usefulness, safety and access, are narrowing.

What is enterprise AI resilience?

Enterprise AI resilience means building AI systems that can keep working when model prices, policies, security conditions or vendor products change. In plain terms, it is the ability to swap models, adjust workflows and keep controls in place without rebuilding the whole system.

Vaidya said chief technology officers, chief information officers and engineering leaders are being pushed to solve a hard problem: keeping reliability high while limiting costs when large AI providers can change pricing or release new models on their own timelines. A hyperscaler is a large cloud or technology company that operates vast computing infrastructure, and those providers can shape what enterprise AI costs and how it performs.

His proposed answer is a flexible technical layer that lets teams change models and adjust how AI agents interact with people, data and tools. AI agents are software systems that can take steps toward a goal, rather than only returning a single answer to a prompt.

Why tokens became a cost issue

Vaidya also criticized “tokenmaxxing,” a term used for pushing heavy AI token use across teams. Tokens are the chunks of text that AI models process, and many paid AI systems charge based on how many tokens go in and out.

According to Vaidya, chasing token-heavy usage has created unsustainable spending and burnout for some engineering teams. He said companies should spend more effort modernizing infrastructure so teams can test and use AI tools at scale without creating new bottlenecks.

He cited Meta as an example of a company that publicly encouraged broad AI use and, according to The New York Times, has shifted toward reinvesting in engineering team culture. Vaidya framed that shift as part of a move away from token competition and toward morale and sustainable engineering practices.

How companies can reduce AI vendor lock-in

Vaidya said companies should give AI agents access to realistic environments and real business workflows, while applying the same controls used for human engineers. That includes limiting permissions to cases where they are needed, controlling credentials and keeping records of what agents do.

He argued that companies should avoid trapping custom workflows, automation and internal knowledge inside a single vendor relationship. The goal, in his view, is to preserve the option to use a leading paid model when it makes sense, or switch to an open-source model when cost, access or collaboration needs point that way.

For companies spending heavily on AI, that flexibility could shape future costs and execution risk. Vaidya’s central point is that the next phase of AI adoption will reward teams that can adapt their architecture as models change, rather than teams that only move quickly to adopt the newest tool.

This story draws on original reporting from Crunchbase News.