Example (Zeeguu): Topic classification of articles is currently performed by pasting the abstract into an LLM. This works but is expensive. Once we are satisfied with the topic taxonomy and have accumulated enough labeled data, we will switch to a dedicated topic detection framework.
Forces: LLMs are general-purpose machines that can attempt almost any text-based task. Building dedicated, efficient solutions requires upfront investment and training data that may not yet exist.
Solution: Use the LLM to perform a task in production while you build a more efficient replacement. The LLM serves as the “wizard behind the curtain” — convincing to users, allowing early beta-testing and feedback, but intended to be temporary.
Variant — Bootstrapping: In a more subtle variant, the LLM generates the training data for its own replacement. For example, Zeeguu uses an LLM to estimate text difficulty levels. These LLM-generated difficulty labels are being accumulated as training data for a classical classifier that will eventually take over the task.
Note: This pattern has an interesting lifecycle relationship with pattern LLM as Fallback a system may start with the LLM as primary (Wizard of Oz), migrate to a specialized tool as primary, and then keep the LLM as fallback — completing a full cycle from LLM-first to LLM-as-backup.