Interview
Giving New Life to Unstructured Data with LLMs and Agents
- AI is expected to fully absorb RPA and drive automation by solving previously unstructured data problems, shifting enterprise acceptance criteria from absolute perfection to predictable error rates with clear escalation paths.
- The future execution model will be decentralized and federated, allowing thousands of agents to dynamically discover, communicate, and execute tasks across siloed systems without a central controller.
- Workflows will be constructed by using AI agents at "compile time" to draft logic and reason about system operations, while "runtime" execution remains deterministic, auditable, and defined by human validation to ensure safety.
- Companies plan to build codeless solutions that generate complex, explainable, and guaranteed accurate workflows, utilizing identity pass-through to allow agents to act on user behalf with capped permissions.
- User experiences in sectors like lending, immigration, and insurance will transform into conversational interfaces with significantly reduced latency, moving processes from weeks to seconds.
- Human oversight will evolve from reviewing every step to handling only the 20% of cases where AI confidence is low, with humans editing AI-generated drafts to produce deterministic artifacts.
- A major strategic bet involves using AI agents to manage end-to-end workflows and multi-agent communication, though there is a recognized risk that free-form agentic systems may not succeed in the short term.
- Enterprises must invest in systems ensuring reliability, predictability, and auditability to meet compliance, shifting from "black box" models to transparent systems where explainability is as critical as accuracy.
- As AI reliability improves, processes will shift from sequential document handoffs to real-time, interactive collaboration, reducing boilerplate work and allowing humans to focus on high-value decision-making.
- Over the next two years or more, the industry is predicted to move toward hybrid approaches for unstructured data and federated execution, potentially replacing traditional centralized models with dynamic agent collaboration.