Is Unstructured Data The Key To Successful AI Deployments?

TL;DR

• 1

  Most AI pilots fail because they are performative, not intentional: Highland CEO Jitesh Guy says board-driven pilots often exist to cure FOMO, while successful deployments start with a specific business problem, budget, and a plan to combine models, compute, and proprietary enterprise data.

• 2

  Enterprise data, especially unstructured content, is the real moat: Guy cites the common view that models are converging and compute is widely available, so the durable advantage is a company's own contracts, emails, notes, conversations, images, and documents that frontier labs do not have.

• 3

  Brute-force vectorizing petabytes creates a new 'data swamp': Highland argues companies need industry ontologies, a graph of related content, and governed access controls for agents, otherwise hallucinations and messy retrieval show up fast at Fortune 500 scale.

• 4

  Healthcare's big win is automating document-heavy workflows around care: Mike Campbell points to referrals, scheduling, billing, and patient onboarding, where pulling the right subset of records can shrink waits that currently take weeks and free doctors and nurses from paperwork.

• 5

  Insurance gets sharper when AI listens for what humans miss in the moment: Erie Insurance CIO Partha Srinivasa describes agentic copilots that monitor claims calls in real time, prompt adjusters to collect police reports or witness info, and improve subrogation, fraud detection, and customer follow-up.

• 6

  Token spend needs the same discipline companies once applied to cloud costs: Srinivasa says Erie is tracking AI usage through an AI center of excellence because wasted tokens are real, and he expects token costs could eventually exceed compute if teams use them without a business case.