AMA: Scaling AI Applications into the Enterprise

Enterprise AI adoption hinges on two things: proving measurable ROI fast enough to win internal buy-in, and building systems with guardrails that can evolve as models change. Decagon and Clay describe approaches that treat AI deployment less like a one-time “pilot” and more like an ongoing product cycle—complete with evaluations, controlled rollouts, and governance that non-technical teams can manage.

Clay’s co-founder, Verun, frames the company’s origin around scaling human capability—giving teams an order-of-magnitude boost in what they can do with data and automation. Clay started with data enrichment for cold email marketing agencies, then shifted into a go-to-market platform built on intent signals and automated actions. AI changed the economics and workflow of go-to-market: usage-based pricing reduced the need to charge by seat or rep, and AI made sales less “one-to-one” and more “one-to-many,” enabling systems that scale growth efforts. In enterprise deployments, Clay emphasizes that pilots must translate into quantifiable outcomes—ticket resolution, cost reduction, revenue impact, and customer satisfaction—otherwise the business case stalls among competing priorities.

Decagon’s co-founder, Jesse, ties the company’s founding to customer discovery rather than a preconceived product. Customer support kept surfacing as a problem where GenAI agents can deliver both value and measurable impact. Decagon’s agent design aims to match what human support agents do while adding proactive capabilities. When new models arrive, Decagon relies on evaluation suites for each use-case “surface area,” including checks for off-rails behavior. Crucially, customers can define their own evals and test sets—effectively unit tests—so performance can be validated against each enterprise’s real workflows. If results look good, rollouts happen via A/B testing (e.g., starting with a small audience share and expanding).

Both companies stress guardrails as a prerequisite for enterprise trust, especially in customer-facing contexts where incorrect outputs can carry reputational and contractual risk. Decagon highlights “agent operating procedures” (AOPs)—a structured way for non-technical stakeholders to set and customize constraints—shifting responsibility from a central AI team to the business users who understand policy and risk. The risk conversation is also evolving: instead of obsessing over rare catastrophic mistakes, enterprises increasingly look at error rates and how they compare to human performance.

On deployment success, the panel pushes back on the idea that enterprise AI fails because of the models alone. The biggest failure mode is organizational: traditional go-to-market teams operate in silos, while top-performing teams run growth like product development. That means building reusable systems, starting with small betas, using appropriate data sources, and keeping a human-in-the-loop for escalation. Decagon and Clay also differentiate by choosing wedges that create numerical proof—Clay starts with data quality and enrichment tests, while Decagon focuses on agent performance and measurable customer support outcomes.

Finally, they address market strategy tradeoffs: whether to go deep in one vertical or stay horizontal. Clay leans toward horizontal because conversation automation needs are broadly similar across industries, while Decagon’s enterprise focus still allows it to serve multiple verticals without losing product fit. Resource allocation advice is pragmatic: early-stage companies should be go-to-market driven, but scaling go-to-market too aggressively can mask product bottlenecks. Their shared takeaway for founders is to avoid overfitting to others’ playbooks and instead build from personal strengths and curiosity—then iterate step by step.