| Route | Useful when | Main cost structure | Key responsibility retained by buyer | Contract question |
|---|---|---|---|---|
| Internal build | Team can own data, experiments and service operations | Staff, compute, tools and opportunity cost | Almost all lifecycle responsibility | Who covers absence and incidents? |
| Managed adaptation platform | ML lead can drive experiments but wants infrastructure support | Usage plus capacity/enterprise terms | Task design, labels, integration and acceptance | Are artefacts exportable and which support is included? |
| Engineering consultancy | External delivery capacity is the missing resource | Discovery, milestones and support | Product decisions and independent acceptance | Which named people deliver each stage? |
| Domain studio | Specialised review expertise and workflow integration matter | Scoped work plus recurring support | Confirm domain evidence and transferability | Is the domain claim supported by a comparable reference? |
| Independent evaluator | Buyer needs a credible challenge to internal or supplier claims | Assessment and repeat testing | Business thresholds and action on findings | Who funds it, and is the test under buyer control? |
| Specialist API | A ready-made capability can meet the task | Usage, minimum commitment or enterprise quote | Integration, oversight and service dependence | What happens to data and functionality at exit? |
Together AI and Fireworks describe adaptation/inference platforms; Winder.AI describes consulting and development services. These establish different offer categories, not comparable fixed scopes. A tooling subscription should not be priced as if it also includes domain labelling and production ownership.