Source note

Comparative Advantage in Software

AI Software ProductsCoding AgentsLLM ApplicationsProduct StrategyAgent Workflows

The essay argues that AI-generated code makes software valuable only when a product reduces token cost or raises task correctness better than a customer can do with a general model. It is a business and product strategy piece, with Kintow restaurant operations examples rather than experimental research.

  • Coding agents lower the cost of building software, so buyers need a clearer reason to pay for products they might recreate with Claude or another LLM.
  • Thin AI wrappers can lose value when they add little beyond model calls.
  • Complex operational software still matters when it maintains state, enforces workflows, and prevents errors that would cost users time or money.
  • The method is a cost model: compare a product with direct LLM use by asking how many tokens the task takes and how correct those tokens are.
  • Products create value by reducing the amount of model reasoning needed through workflows, domain models, typed data, and deterministic steps.
  • Products also create value by improving correctness through context engineering, evaluations, persistent state, and error-correcting workflows.
  • Kintow applies this to restaurant ERP tasks such as inventory, waste, recipes, labor, automated procurement, and invoice parsing.
  • The excerpt reports no experiments, datasets, baselines, or accuracy metrics.
  • It claims 2 product levers: lower the number of tokens needed for a task and increase the correctness of each token produced.
  • Automated procurement needs precise state and approval controls because purchasing can affect about 30% of a restaurant’s revenue.
  • Invoice parsing works on a single document with Claude, but repeated runs can vary and fail to sync with a catalog; the claimed fix is evaluations, context engineering, and correction workflows.
  • The strongest concrete claim is that customers with spreadsheets, incumbent tools, or home-built systems may still buy specialized software when maintenance and reliability costs exceed the cost of the product.