Most AI tool integrations follow the same pattern: fetch raw data, paste it into the prompt, ask the model to analyze it. A CSV with 500 rows. A database dump. An API response with nested JSON. The LLM reads every byte, burns tokens parsing structure it cannot see efficiently, and produces a summary that a three-line script could have generated.
This approach is expensive, slow, and lossy:
- Token cost scales with data volume. 50 Reddit posts at ~500 tokens each = 25,000 input tokens just to score and rank them. At Opus pricing, that is $0.42 per query.
- Context windows are finite. A 200K context window sounds large until you load a real dataset, a system prompt, conversation history, and tool definitions. The data competes with reasoning space.