Interview, Fireside Chat
AI's Dropout Problem: Why the Best Agents Never Stop Learning | Decagon | RAISE Summit 2026
- Decagon is developing enterprise-grade systems to enable hundreds of employees to collaborate on agent development, featuring test suites that simulate interactions against the last million conversations to validate functionality before deployment to millions of consumers.
- The "Duat Autopilot" system aims to automatically analyze human agent escalations to identify patterns and pre-draft fixes, with automated changes staged for human admin review prior to production deployment as enterprise workflows shift toward daily morning reviews of overnight improvements.
- Future technology gains over the next 12 months are expected to focus on the automatic improvement of agents without enterprise effort, relying on learning from data generated by millions of daily interactions.
- Enterprises are projected to invest in agent technology to increase customer interaction volumes, solve problems proactively, generate revenue, and reduce churn, utilizing reduced costs to expand touchpoints rather than cut headcount.
- The business model is expected to transition from seat-based pricing to outcome-based pricing tied to human agent costs, with specific AI use cases like PR reviews and sales briefs evaluated against ROI metrics to justify expenditures.
- Market strategy is predicted to shift from using frontier models for new tasks to employing smaller, tuned open-source models for known tasks to achieve higher accuracy, speed, and lower costs within customer infrastructure, alongside the continued development of custom models for specific use cases.
- Competitive advantage is anticipated to rely on building surrounding infrastructure including collaboration tools, testing suites, compliance, and QA, rather than solely on core agent capabilities, addressing the current gap in regulatory compliance and deep product depth for major enterprise customers.
- A significant gap remains between proof-of-concept capabilities and the ability to deploy agents creating millions of dollars of value, with market demand expected to remain thin for solutions handling hundreds of millions of people in complex, regulated environments.
- Post-deployment tooling is designed to monitor for behavioral deviations and catch issues immediately before they impact the end customer, while industry adoption is shifting focus from model capability to the construction of infrastructure, guardrails, and data ingestion tools required for enterprise deployment.
- Broad industry adoption is expected to move away from unverified AI rushes in favor of rigorous ROI analysis, with the core competitive moat driven by the construction of infrastructure and guardrails necessary for reliable large-scale enterprise integration.