Why we watch him. Levie is the sharpest, most data-grounded public voice on the exact theme Dominic wants to major on: why AI coding has taken off but enterprise knowledge work has not, and where the durable human value sits (judgement and taste). As Box CEO he speaks from first-party enterprise data across 68% of the Fortune 500, posts almost daily, and is analytical rather than hype-driven. High signal per post, and a useful tripwire for the wider enterprise-AI debate.
Current headline view: coding agents have reached escape velocity; enterprise knowledge work is the hard frontier, and the real value sits in judgement and taste.
Latest 1-page summary (as at 5 Jul 2026)
His thinking is stable and consistent across X and long-form interviews. The ten core theses we track him against:
Tale of two cities. Coding races, knowledge work crawls. The gap is explained by four deltas: medium (code is text, models trained on it), user skill (engineers can steer agents), verifiability (code passes a regression test, a contract must "experience reality"), and data access (engineers already hold the whole codebase; enterprises are fragmented permissions). Dominic's headline theme.
The human moat is judgement and taste. Humans stay in the loop because agent taste is off and ROI is a squishy judgement problem.
Context is the binding constraint. Right context at the right time with the right guardrails, never maximum context.
The applied AI layer is the value layer. Four-part playbook: bridge features, model routing, FDE change-management, domain go-to-market. Cursor's ~$60B is the first proof.
The forward-deployed engineer is the key new role, internal and external.
Data and IT architecture is the real unlock. "Show me your IT stack, I'll show you what you'll get from agents."
Model routing is inevitable. Frontier for planning and review, cheaper or open models for the bulk.
The token subsidy is over. Real costs now; the line of business, not IT, must own the AI budget. Value maxing over token maxing.
Regulate the use of AI, not the model layer (the Fable export-control saga was a bad, event-driven precedent).
AI adoption grows headcount, not cuts it (Box's own survey plus Ramp, JOLTs, Apollo data).
Latest scan (5 Jul): no new X posts since 3 Jul. Replies add nothing (we scan Posts only). His latest interview is the CXOTalk "Advice for CIOs on AI Agents" (15 Jun), fully transcribed and analysed.
Analysis backlog (newest first)
Deep source docs are held locally in the workspace at expert-watch/, not on this page.
5 Jul 2026 · Eight-project sprint: replies scan (nil additive), May scan, June-vs-May deviation report, session discovery, Google and LinkedIn method.
5 Jul 2026 · CXOTalk interview, "Advice for CIOs on AI Agents": full analysis of the coding-vs-knowledge-work thesis, with the four-part T-chart, plus transcript.
5 Jul 2026 · a16z Show, "AI Inside the Enterprise" (24 Apr): full audio transcript.
5 Jul 2026 · Base profile and June deep analysis: full-text read of all 55 X posts (1 Jun to 3 Jul), the ten core theses, and the deviation log.