Key takeaways
- Private means where inference runs, not whether the vendor “doesn’t train on your data.”
- Hosted CRM AI adds a subprocessor and often a usage meter.
- Local models are strong at retrieval and weak at essays — which matches what a CRM assistant should do.
- Qivity’s Lilly runs on Qivity servers when you are hosted, or on yours when you self-host. Either way she is an optional AI add-on on paid plans — not on the Free trial.
A private AI CRM is CRM software whose assistant runs on language models inside your own deployment. When a rep asks “how much is in proposal in the south?”, the question is classified locally and answered from your database. No customer record is packaged into a prompt and sent to OpenAI, Anthropic, Google or the vendor’s hosted cluster.
That is the whole definition. Marketing pages that say “AI-powered” without naming the infrastructure are describing a chatbot bolted onto someone else’s API.
Private AI CRM vs hosted-model CRM
There are two architectures. They look similar in a demo and they are not.
Hosted-model CRM AI copies relevant records into a prompt and sends that prompt over the internet. The vendor will usually say the data is not used for training. That can be contractually true and still be the wrong property. The prompt left the building, crossed a border, and created a subprocessor your legal team has to name.
Private AI CRM keeps inference next to the database. In Qivity that assistant is Lilly: small local models, answers executed as real queries, documentation quoted rather than paraphrased.
What a private AI CRM should actually do
A sales assistant does not need to write a strategy memo. It needs to understand what a busy person meant and fetch the right thing:
- Counts and values — “how many leads in Riyadh?”, “pipeline at proposal or later.”
- Filters inside the sentence — and a statement of which filter was applied.
- Reminders and meetings on the real calendar, confirmed before they land.
- Record search by meaning, not exact wording.
- A download of the matched set, not a paragraph that might have invented a row.
None of that requires a frontier model. It requires classification, retrieval, and the discipline to say “I don’t know.” Hallucinating a pipeline number is worse than silence. More on the privacy argument is in where the model runs.
Questions that separate the two
- Which companies’ servers process an AI request? Names, not categories.
- Is that named in the DPA as a subprocessor?
- Can I pin inference to a region — or turn egress off entirely?
- Is AI metered separately from the seat licence?
- Can a junior rep’s assistant see the VP’s pipeline?
If the demo cannot be run as a junior user, assume the permission layer was skipped.
Who a private AI CRM is for
Regulated teams, anyone whose board will ask about AI vendors, and anyone tired of a usage meter that punishes the people you are trying to get onto the tool. It is a weaker fit if you want the assistant to draft long-form content; that is not the job Qivity hired Lilly to do.
See Lilly on a sandbox with egress blocked, or read the security page for the one-sentence version legal actually wants.