Bring your own key

Three things, and all three are commercial before they are technical:

• Price. You pay the provider's rate rather than the provider's rate plus a margin, and you can move to a cheaper model the week it ships.
• Choice. Model quality moves monthly; a vendor-selected model is a bet on someone else's procurement.
• Data path. Your prompts go to a provider you have a contract with, under terms you agreed, which is the version of this question a security review can actually assess.

Pointing the same integration at a model running on your own machine or your own server changes the answer to where does the data go from a policy question to an architectural one — nothing leaves. For teams whose prospect notes and call summaries are commercially sensitive, that is often the deciding factor.

The trade is quality and speed against control. A small local model summarising a reply is fine; the same model writing a first draft usually is not. Being able to choose per task is the point.

The feature should degrade honestly. With no key configured, an assistive feature ought to fall back to something deterministic and say that it did, rather than erroring or quietly producing something that looks generated.

And the credential should be revocable by you, in one place, without a support ticket. A key you cannot rotate yourself is not really yours.

SalesShift's assistive features use the key you connect — OpenAI, Anthropic, or a local Ollama. With no key configured they fall back to a deterministic template and say so, rather than erroring or implying a model wrote it.