The real AI bottleneck isn't models—it's understanding your business

The talk argues that AI models have solved execution of knowledge work, but enterprise adoption fails because AI is slapped onto broken processes. The real bottleneck is understanding unique business context—how humans actually work, including edge cases and exceptions.

Forward deployed engineers (FDEs) map current workflows, re-engineer them around AI, and deploy agents on top of existing systems of record (SAP, Salesforce) to avoid costly migrations. They keep humans in the loop where risk is high, achieving 25-75% department-wide ROI.

Varick built an FD agent tool to assist FDEs: an engagement agent synthesizes documentation, a workflow agent catches missing edge cases, and a future autonomous agent handles small changes. They use dependency graphs, post-trained models, and RL for context extraction.

The claim that knowledge work is "almost entirely solved" is hype. The evidence is anecdotal, and the approach requires significant upfront consulting. Teams should focus on process redesign and governance, not just model improvements.