What the AI in our CRM actually costs to run

Every AI feature in a product is, underneath, a request to a language model. Providers charge for that by token, separately for input and output, at a rate that depends on the model. A quick classification task can run on a fast, low-cost model. A task that needs more reasoning, such as drafting a careful reply or summarizing a complex record, benefits from a slower, higher-quality model billed at a very different rate. Two AI features in the same product, doing different jobs, do not cost the same to run. A pricing page that says "AI included" does not change that.

A flat number is not dishonest by itself. But it hides information a customer might reasonably want: which features are cheap enough to use freely, and which ones drive the real cost. Without that, a customer cannot reason about their own usage or predict what happens if they lean on one feature heavily.

SalesCrew routes every AI request, from any feature, through one internal gateway called ai-complete. Each request is logged as an ai_run. The run references an ai_rates table that defines the real cost for that model and task type. So for every AI action taken anywhere in the product, the system has a record of which model handled it and what it cost. Not an aggregate number nobody can break apart afterward.

The model for each feature is chosen by task class, which SalesCrew's architecture calls fast versus quality tiers. Model cost and capability are matched to what the task needs. No feature defaults to the most expensive model when the task does not benefit from it. That keeps average cost lower across the product. It does not make any request free. The metering exists so that reality stays visible rather than smoothed into one marketing number.

A vendor offering uncapped AI at a flat price is making a bet. Either customer usage stays low enough that the flat price still covers the provider's usage-based cost, or the vendor absorbs a cost that grows with usage whatever the customer pays. Neither is a stable basis for pricing at scale. So SalesCrew publishes a cap instead. Each tier carries a fair-use allowance measured in model calls: roughly 5,000 a month on Core, 25,000 on Outbound, and 60,000 on Agency. There is a soft alert at 80% and a hard stop at 100%. The cap applies to AI usage only. Every manual, non-AI path in the product keeps working past it.

A customer who wants to remove the allowance question entirely can use BYOK. They connect their own AI provider key, that provider bills them directly, and they pay SalesCrew a flat add-on fee instead of drawing on the shared allowance. Both paths rest on the same principle. The real, variable cost of AI should be visible and traceable, per request if needed, not hidden behind one number that cannot be checked against what happened.