Scenarios rarely hand you one job. They hand you three features in one release, and the filters point at a different model for each, which is worked through at length in the business-side weighing and again, at design time, in the per-feature selection. This card is the list itself, in the order that throws candidates out fastest. It starts after the prior question has been answered: whether a generative model suits the job at all is settled in the read on where generative AI helps and where it hurts.
Flash card
Choosing a foundation model: filters run per feature, where model type, constraints and compliance eliminate, and capabilities, context window, latency and cost rank what survives.
- Model type splits the catalogue first. A text-generation model, an embedding model that turns input into a vector for similarity search, an image-generation model and a multimodal model that reads an image or audio and answers in text are different products. Prompting does not fix the wrong type.
- Capabilities and performance requirements are two questions: whether the model can do the task, and how well it has to do it before the feature ships. Output a person reviews first has a lower bar than output sent straight to a subscriber, and that lower bar often lets a smaller, cheaper model through.
- The context window is the text a model holds in one call, prompt and response together, counted in tokens. A two-hundred-page contract fed in one go needs a large window; a short chat turn does not. On-demand billing is per input and output token, so an unfilled window adds nothing to the bill, but requiring one cuts models that would have done the job.
- Latency and cost track model size: inference latency scales with the model, and the larger model costs more per token. That decides a live assistant with somebody watching a spinner, and barely registers for an overnight batch.
- Constraints and compliance eliminate before any ranking: which Regions offer the model, where inference may run (a geographic Bedrock inference profile keeps processing inside one geography, a global one does not), what the licence permits you to do with the output, and which languages it handles well.
Pick it when
Run the filters per feature, not once per product, eliminating ones first: model type and modality, then constraints and compliance, then the window the job needs. Capabilities against the performance requirements, then latency and cost, order what is left.
It's the wrong answer when
This list does not settle whether generative AI suits the job, nor how the model is customised: prompting, retrieval over your own documents and fine-tuning come after the shortlist exists. For a numeric prediction over your own tabular data, a traditional machine-learning model is usually better.