barry mccardel

AI expands data access along three vectors — breadth (more people asking questions), depth (multi-turn follow-up analysis), and speed (proactive insights before questions asked). The percentage of data-informed decisions will rise from near-zero as friction collapses.

The step-change from one-shot LLM outputs to models that can reason, iterate, loop, call tools, and manage sub-agents over dozens of turns unlocked Hex's notebook agent and Threads. This same capability wave powers AI coding (Cursor) and customer success (Decagon, Sierra).

The differentiator is not UI or model choice but context engineering — semantic models, governance, online eval systems (confuse meter, thinking-token inspection), and human-in-the-loop feedback loops that compound over time to make answers trustworthy, not just syntactically correct.

A separate AI team creates silos and bolted-on features. Embedding AI fluency in every PM/designer/engineer — with shared AI platform infrastructure — compounds value as teams natively consider agent interaction in every feature, even non-AI ones.

Public benchmarks are misleading; rigorous evals on real customer workloads (1,700 companies) show material differences in how models handle analytics context (huge result sets, chart encoding), tool-calling patterns, and sub-agent delegation. Anthropic's coding lead stems from Cursor/Claude Code token flywheel.

Charging for AI as an add-on (like cloud in 2014) warps incentives and is short-lived. Hex includes AI in base seats (expanding TAM as more roles become data users) and will add consumption overages for high-volume agent workloads, aligning price with value.

Palantir's model put real engineers in the field (war zones, factories) to discover problems and build solutions that fed the core platform — a Darwinian search function. Most companies calling sales engineers 'forward-deployed' miss the builder-in-the-field essence and the organizational chaos tolerance required.