Key takeaways
- AI CRM data privacy is decided by where inference runs, not by the model’s brand name.
- Hosted-model assistants send your records to a third-party provider as a subprocessor.
- Local-model assistants run inside your own deployment. Qivity’s Lilly does that when you self-host; hosted plans run Lilly on Qivity’s servers.
- Ask any vendor which companies process an AI request — names, not categories.
Every CRM vendor shipped an AI assistant in the last two years. Ask one of them a straight question — when a rep types "summarise this account", which company's servers process our customer list? — and the answer is usually three paragraphs long.
It shouldn't be. AI CRM data privacy reduces to exactly one architectural fact, and there are only two possible answers. The difference decides whether your CRM is a compliance question or not.
The two architectures
Hosted-model AI is what almost every mainstream CRM ships. Your records are packaged into a prompt and sent over the internet to a foundation-model provider — OpenAI, Anthropic, Google, or the vendor's own hosted cluster. The answer comes back. The vendor will tell you the data isn't used for training, and in most cases that is contractually true. It still left your building, crossed a border, and passed through a subprocessor you did not choose.
Local-model AI runs the model inside your own deployment. The prompt never touches a network the vendor doesn't control — in a self-hosted install, never touches a network you don't control. Qivity offers both: hosted Lilly runs on Qivity’s servers; self-host and she runs on yours.
For a marketing team brainstorming subject lines, the first is fine. For a pipeline containing named individuals at named companies with deal values attached, it is a question your legal team is going to ask. Hosted Qivity is Qivity’s servers, not OpenAI. Self-host if the requirement is that it never leaves your network.
What AI CRM data privacy actually changes for a buyer
Three things.
Your subprocessor list grows. Under GDPR Article 28 and most equivalent regimes, a hosted-model CRM adds a subprocessor to the chain that touches personal data. That is disclosable, and in regulated sectors it is negotiable — which means procurement cycles measured in months.
Data residency stops being simple. You can pick an EU region for your CRM database and still have inference happen elsewhere. The two are separate decisions, and buyers routinely assume they are one.
Your AI budget becomes usage-based. Hosted models are metered. Adoption of the assistant is a variable cost, which is a strange incentive to hand a sales team you are trying to get to use the tool.
The trade-off nobody admits
Local models are smaller. A 3-billion-parameter model on your own CPU will not write you a strategy memo, and any vendor claiming otherwise is selling you something.
What it does do well is the job a CRM assistant actually has:
- Understand what a person meant — "how many leads do I have in Riyadh?", "what meetings do I have Thursday?", "remind me to call Acme next Tuesday at 3".
- Route that to a real query against real records.
- Quote the product documentation accurately instead of paraphrasing it wrongly.
Notice that none of those need creative generation. They need intent classification and retrieval. That is precisely the workload small local models are good at — and it is also the workload where a hallucination is most damaging. A confidently invented pipeline number is worse than no answer at all.
This is why Qivity's assistant is built to say "I don't know" rather than guess, and why answers about your records are executed as queries under the asking user's own permissions rather than generated as prose. A manager asking about their team gets their team. Someone asking about a colleague's pipeline gets a polite explanation of why not.
Five questions to ask any CRM vendor
Take these into your next demo. The hesitation is as informative as the answer.
- When I use the AI assistant, which companies' infrastructure processes the request? Ask for names, not categories.
- Is that in the DPA as a named subprocessor? If not, why not.
- Where does inference physically happen, and can I pin it to a region?
- Is AI usage metered separately from my seat licence?
- Can the assistant answer about data the asking user is not allowed to see? You would be surprised. Permission-scoping an AI layer is genuinely hard and frequently skipped.
Where this is heading
Local models are improving faster than hosted ones are getting cheaper. The capability gap that made hosted inference the obvious default in 2023 has narrowed considerably for retrieval-and-routing workloads, which is most of what business software needs.
Our bet is that within a few years AI CRM data privacy — specifically "where does the model run" — will be a standard line in every enterprise software RFP, the way "where is the data stored" became one after 2018. We built for that assumption from the start.
If AI CRM data privacy is the blocker in your evaluation and you want to see what a local-model assistant actually feels like to use, book a demo — we will run it against a sandbox with no internet egress and let you watch it work anyway. There is more on the architecture on the AI CRM page and the security page.