Estimating what AI in your CRM will cost
Model spend is per call and per token, so an estimate is a multiplication. Here is the arithmetic and the two features that dominate the total.
- Per-feature formula
- Where the cost concentrates
- No email required
The short answer
- The formula per feature: monthly cost = number of records touched × calls per record × cost per call. Sum it across features. The hard part is the first number, and you probably already have it.
- Two features dominate: drafting replies and anything that runs over every record. A summary generated once per deal is cheap; a classification running on every inbound message all day is where the volume is.
- The structural protection matters more than the estimate. A soft alert at eighty percent of the allowance and a hard stop at a hundred, applying to AI only, means a bad month is capped rather than discovered on an invoice.
The method, not a widget
The formula
Monthly cost for a feature = records touched per month × model calls per record × cost per call
Total = the sum across every feature you turn on
Cost per call depends on the model and on how much text goes in and out. A short classification on a two-line email is a fraction of a long draft written against several knowledge-base sources, and the difference between them is often ten times or more.
The input you need is volume. How many emails arrive a month, how many deals move, how many contacts you score. Every one of those numbers is already in your inbox and your pipeline.
Where the volume actually is
Ranked by how much a typical team spends on each, highest first.
| Feature | What drives the volume | Relative cost |
|---|---|---|
| Reply drafting | One call per draft, with long input from the thread and the knowledge base | Highest per call, and it grows with team activity |
| Inbox classification | One call per message the rules did not match | Low per call, high volume. Rules first is what keeps this small |
| Lead scoring | One call per record scored, times how often you rescore | Moderate, and rescoring frequency is the lever |
| Record summary | One call per record summarised, on demand | Moderate per call, low volume |
| AI fields | One call per field per record when it fills | Depends entirely on how many fields you define |
| Sequence drafting | One call per sequence written | Low volume, so rarely the problem |
Ordering reflects where model calls concentrate in a sales workflow, not a price list. Actual cost depends on your model choice and your text lengths.
Four ways to cut it without losing the feature
- Rules before models. A regex that catches every out-of-office message costs nothing and is explainable. Send the model only the residue.
- Score fewer records, less often. Rescoring your whole database nightly is the most common source of a surprising bill.
- Pick the model class per feature. A fast model for classification and a stronger one for drafts. Using the expensive model everywhere is the default nobody chose deliberately.
- Bring your own key. Model spend goes on your own account at your own rates and does not consume an allowance.
How SalesCrew meters and caps it
Core includes a fair-use allowance of roughly five thousand model calls a month, Outbound is around twenty-five thousand and Agency around sixty thousand. Every call goes through one gateway that records what it cost and which feature spent it, so per-feature spend is a query rather than a guess.
The cap behaves in the way that matters: a soft alert at eighty percent, a hard stop at a hundred, and the hard stop applies to AI only. Cadences keep sending, the inbox keeps working and people keep using the CRM. Nothing manual stops because a model allowance ran out.
Top-ups are a thousand credits for ten dollars when you hit the cap, and bringing your own key at ninety-nine dollars a month takes model spend off the allowance entirely.
Estimate it, then cap it
A hard stop that never touches the manual path is worth more than a precise forecast.