In-House Training vs Fine-Tuning vs Vendor Licensing for a Generative AI Feature

A product team is planning a new feature that runs on a generative AI model. Which factors decide whether you should train a model in-house, adapt an existing model through fine-tuning, or pay for a vendor's model? ```hint Find what has to be yours Ask which part of this feature a competitor could not copy by calling the same vendor: the model itself, your data, or the workflow around it. The answer changes how much each path is worth. ``` ### Clarifying Questions - What will the model do, and is that capability a core differentiator for the product or a common one such as summarizing or drafting text? - Does the feature rely on proprietary or personal data, and are we allowed to send that data to an outside vendor or use it to train a model? - What request volume, latency budget and launch date should the decision assume? - Does the team already have ML engineers, GPU capacity and labeled data, or would those need to be built up? ### What a Strong Answer Covers - Strategic fit: how much differentiation each path creates and how quickly each one reaches users - Quality measured on examples from this task rather than general benchmarks, including whether prompting and retrieval close the gap without any training - Data rights, privacy, security and regulatory limits on each path - Total cost of ownership as usage grows: per-call fees against fixed team, compute and maintenance costs - Dependency risk (vendor pricing, model retirement, lock-in) and the checkpoints at which the decision is revisited ### Follow-up Questions - Six months after launch the vendor doubles its price. How would your original decision have prepared you for that? - How would you show leadership, with numbers, the usage level at which fine-tuning becomes cheaper than paying per call? - Legal rules that customer data may not leave the company. Which options remain, and what do they cost you?