Workflows Over Fine-Tuning at Copy.ai
Chris Lu, co-founder of Copy.ai, on generative AI in the enterprise
This reveals that Copy.ai is trying to win at the workflow layer, not the model layer. For broad enterprise tasks like account research, email drafting, and earnings call analysis, the bottleneck is less the base model itself and more how quickly a company can swap in a cheaper or better model, connect it to CRM and web data, and turn it into a repeatable business process. That is why Copy.ai stays multi model for general work, while only fine tuning narrow jobs with exact labels and clear pass fail criteria.
-
Copy.ai had already used many fine tuned models in its earlier prosumer product, but by 2024 it had shifted toward enterprise workflows where model vendors were improving fast enough that retraining a general model could become obsolete in weeks. The product advantage moved from custom weights to orchestration, integrations, and fast model switching.
-
The exception is narrow classification work. In the interview, a customer needed websites mapped into exact CMS category IDs. That is a task with stable labels, a clear right answer, and high repetition, so a small custom model can pay off quickly and then be reused across many workflow steps inside the business.
-
This mirrors the broader stack. Open source and API providers keep cutting cost and adding tuning options, while infrastructure for model training and data preparation keeps getting easier. That makes bespoke fine tuning more valuable at the edges, for tightly scoped high volume tasks, and less valuable as a moat for horizontal application companies.
The next phase is likely a split architecture. Frontier models will handle messy reasoning and long context jobs, while companies like Copy.ai will distill repeated tasks into smaller task specific models once enough clean workflow data accumulates. Over time, the product that owns the workflow and the labeled feedback loop will be in the best position to decide when fine tuning finally becomes worth it.