What this covers
- What actually needs a model
- The privacy argument is stronger than the cost one
- What a sensible split looks like
- Questions worth asking a vendor
What actually needs a model
Writing needs a model. Drafting sequence copy, rewriting for tone, summarising a long reply thread, producing account research — these genuinely benefit from generative AI, and the output quality justifies the cost.
Counting does not need a model. Pipeline value, stalled deals, win rate by source, forecasting, duplicate detection, data hygiene — these are deterministic operations over your own data. Routing them through a language model adds latency, cost and a third-party data transfer, and returns an answer that is occasionally wrong in ways arithmetic never is.
The privacy argument is stronger than the cost one
The cost case is real: credit-based enrichment and token metering both turn routine work into a variable bill that grows with usage. But the procurement case is what wins enterprise deals.
When a security reviewer asks which third parties receive customer records, there is a meaningful difference between "all of them, briefly" and "none, for the large majority of operations". Counting your pipeline should not require sending your pipeline anywhere.
What a sensible split looks like
In Quotarider, 73 of 106 assistant commands run locally with no AI key at all — instant, free, and with no data leaving your server. The remaining 33 cover writing and strategy, and they run on your own model provider key.
That means you pay your provider directly, there is no markup, and every AI feature can be left switched off entirely without the product becoming useless.
Questions worth asking a vendor
Which operations require an AI call? What happens to my data during that call? Is there a markup on model usage, or do I bring my own key? Can I use the product with AI disabled? Does routine reporting consume credits?
The answers vary far more than the marketing suggests.