Build vs Buy Generative AI Consulting

Buy Copilot or ChatGPT Enterprise when that product already completes the workflow under a quality bar you can sample. Build when you need permissioned retrieval, tool actions, or eval-gated behavior the seat cannot provide. ReinforcedX makes that call in week one. If the answer is build, the standard path is four weeks in your cloud. Book /demo with the workflow, not with a preference.

It is a decision plus, when required, an implementation. We do not default to custom. You own weights, datasets, eval suites, and runbooks on anything we build. Work runs in your cloud with zero-retention defaults. Pricing is a platform subscription plus a fixed-scope fee quoted before week one. Inference is billed by you to the provider; no token markup.

When the job is drafting, summarizing, or Q&A on data the vendor’s controls already cover, and you can live without write actions against core systems. It is not enough when answers must cite permissioned sources, when the agent must change an order or a ticket, or when promotion needs CI gates. Those cases are a custom system, not a better prompt pack.

When the workflow requires tools, live entitlements, reconstructable traces, or an eval suite that blocks deploy. That is production software. A well-bounded custom workflow is four weeks: discovery, environments, shadow-mode, handover. Financial-services governed builds typically take 8–12 weeks; healthcare 10–14; ecommerce 6–10. /how-to is the self-serve path if your engineers will build it.

Not if it is built as we build it. The stack is model-agnostic: Anthropic, OpenAI, Google, Mistral, and client fine-tunes. Switching models is an eval-gated config change. Buying a copilot does lock you to that product’s roadmap, which is acceptable when the product already does the job. The mistake is a custom agent hard-wired to one API with no evals.

Not on a workflow that a seat cannot close. Shelfware copilots plus a still-manual process are more expensive than a four-week custom system with a metric. We quote the implementation fee before week one and you pay inference to the provider. We will not invent a TCO multiple; we will put both lanes on the same golden set and compare.

Yes, if one process owner holds the map and both lanes have a quality bar. Hybrid without evals is how sprawl starts. Shared tracing and a shared golden-set habit matter more than a shared model. Patterns for the custom lane live under /ai-systems. The weekly 45-minute review is where the mix is kept honest.

You do: weights, datasets, eval suites, and runbooks. Data is not used to train shared models. Thirty days on-call follow handover. There is no proprietary runtime you must keep renting. Those terms are the same as /faq and are in the quote before week one. If you only buy seats, IP is the vendor’s product — which is the point of buying.