The applied AI layer: work between model and enterprise workflow

Summary

Levie maps six areas of work in the applied AI layer—between a raw AI model and production enterprise workflows. Case studies show more value lives in this integration work than in raw model capability. The areas are: workflow representation (how agents surface to users), data access and context (domain-specific approaches for different industries), change management (vertical-specific implementation), model tuning (using multiple models for cost/performance tradeoffs), evaluation (domain-specific evals create room for large ongoing gains), and industry-specific pricing. Read the full thread.

Key Claims

• The applied layer between AI model and end-user workflow contains more value than raw model capability in enterprise settings.
• Agent design must vary by business process: sometimes a chat interface, sometimes a deterministic background workflow, each needing domain-specific harnesses and product surfaces.
• Enterprise workflows connect to entirely different systems and data types (life sciences, financial, legal); handling this requires contextual data approaches and UX tuned to each domain.
• Domain-specific evaluation frameworks are critical to AI utility; no single frontier model is tuned for all tasks, so custom evals for specific workflows unlock lots of further gains across the economy.
• Multi-model capability allows firms to tune workflows for different cost and performance levels, and to post-train models for specific tasks beyond frontier model capability.
• Vertical-specific change management is essential; technology rollout at a bank differs fundamentally from a law firm, and dedicated domain talent is more effective than generic implementation teams.

Quotes

• "Model capability is obviously doing a lot of heavy lifting in agentic products, but there's still a lot more work to diffuse AI into the enterprise."
• "AI is basically not useful if it can't be evaluated. Domain-specific evals that let you dramatically improve the performance of your harness for specific workflows just has a crazy long tail given how many tasks there are in the economy."