Enterprise AI resilience, not speed, is the talent challenge, Crafting CEO says
Crafting CEO Sumeet Vaidya says AI teams should cut provider lock-in and build systems that can swap models as costs and reliability shift.
By Marcus Adeyemi · Startups Editor
· 3 min read
Crafting CEO and co-founder Sumeet Vaidya says enterprise AI resilience has become a higher-priority talent and infrastructure problem than speed, as companies build more of their work on paid models from frontier labs and hyperscalers. His argument: engineering leaders need architectures that can change providers, control agent access to real systems and keep costs from being dictated by a single model vendor.
Vaidya pointed to an Anthropic policy reversal and a security incident involving Hugging Face and OpenAI as signs that the companies supplying major AI infrastructure do not offer the stability many enterprise teams need. He did not argue that frontier labs are failing to innovate. He argued that reliability is becoming harder to assume when core AI operations depend on outside model providers.
The cost side is also changing, according to Vaidya. He said open-source groups such as OpenClaw and DeepSeek are offering free models with quality close enough to put pressure on paid alternatives. He also said the gaps that kept some enterprises away from open-source options, including utility, safety and accessibility, are narrowing.
What does enterprise AI resilience mean for engineering teams?
Enterprise AI resilience means designing systems so teams can replace or add models without rebuilding the workflows, permissions and in-house logic around them. In Vaidya’s framing, the goal is to keep reliability high and costs controlled even when a hyperscaler changes pricing, releases a stronger model or creates operational risk through an outage or policy shift.
That requirement is different from chasing short-term output gains. Vaidya criticized “tokenmaxxing,” which he described as a trend that has produced unsustainable spending and burnout. He said some Big Tech and enterprise teams are moving back toward more measured AI use, including Meta, which he said has shifted toward reinvesting in engineering team culture after publicly pushing companywide AI use.
Vaidya’s prescription is not to avoid frontier models. He said there will be cases where a new hyperscaler model is the right choice, and other cases where open-source models paired with new harnesses make more sense. The operating principle is optionality: AI infrastructure should preserve the ability to adopt new models without losing custom internal work.
How should AI agents work with real systems?
Vaidya argued that agents should not be confined to toy tasks or synthetic test environments. He said companies should let agents interact with real infrastructure, realistic data and business workflows, while applying the same guardrails used for human engineers.
His recommended controls include limiting credentials and permissions to the circumstances in which they are needed, maintaining visibility into agent actions and making activity auditable when something fails. That approach treats agents as participants in production-adjacent engineering work rather than as separate experiments run outside normal governance.
The strategic point is vendor risk. Vaidya said companies should avoid trapping custom workflows, automation and operational knowledge inside one long-term provider relationship. For CTOs, CIOs and engineering leaders, the practical takeaway is to invest in the layer between teams, agents and models, because that is where provider switching, access control and accountability will be handled.
Vaidya was previously an early engineering leader at Meta, Uber and Discord. Crafting says it builds enterprise-grade infrastructure for autonomous agents and engineers.
This story draws on original reporting from Crunchbase News.
