Grounding vs. training
Grounding supplies selected client data to a model at request time — retrieved, placed in the prompt, used for the answer — without using that data to update the model's weights. Training or fine-tuning changes model weights using a dataset. Grounding does not by itself mean zero retention or that data never leaves the client's infrastructure: with an external provider, grounded data is still processed outside, and provider terms, processing location, caching, logging, and deletion controls still matter.
Why it matters
For a dealer evaluating AI-assisted software, "will our prices train someone else's model?" has a checkable answer: a grounded system under no-training terms does not feed your data into weights. The follow-up questions are just as checkable and often more consequential — where inference runs, what the provider retains and for how long, and how logs are controlled.
Common confusion
"Not used for training" is not the same as "zero retention." They are different controls — one governs weights, the other governs storage — and provider documentation distinguishes them explicitly: standard API logs may be retained for a period even where training is excluded. A careful vendor names both controls in the contract, not the pitch.
Sources
Where we use it
The full mechanism, including provider data-control comparisons, is Where an AI quoting system gets its prices; our own commitments — including runtime processing controls — are in the AI Policy.