When AI in Salesforce runs on your own OpenAI, Anthropic Claude or Azure OpenAI account, the provider bills you directly for what you use. That is transparent, but it raises a fair question from finance: what will it actually cost to run?
This article explains where the tokens go in AI Engine, which settings have the biggest effect, and how to monitor usage. It covers the provider bill only, not the price of the package.
Tokens in one paragraph
AI providers charge per token, a small piece of text, roughly three quarters of an English word. Every request has input tokens (the prompt, including the CRM data sent as context) and output tokens (the answer). Output tokens are usually priced higher, but in CRM use input is typically the larger volume, because each request carries record data.
How many requests each action makes
| Action | Requests to the provider |
|---|---|
| Chat question | Usually 2: one to turn the question into a query, one to write the answer |
| Next Steps, Draft Email, Summary, custom prompts | 1 |
| Analyze Email | 1 |
| Meet Prep | 1, plus 1 web search request if web search is on |
| AI Score | 1 per record, plus 1 web search per record if company verification is on |
| Web Check | 1, with web search |
| Company News | 2 |
| Magic Snap | 1 to read the photo, plus research if used |
| Show all, export, column changes | None |
Most actions are one or two requests. The difference in cost comes from how big each request is.
What drives token usage
- The field allowlist. Each chat question includes a description of the objects and fields the AI may use. More enabled fields mean more input tokens on every question. Trimming fields users do not need is the single most effective saving.
- Record context. More fields, child objects and activities make every record action larger.
- Conversation length. Recent messages are sent with each follow-up question.
- Web search. Search results are added to the request as input, which makes web features much larger than other actions. Web search in meeting prep adds roughly 30,000 input tokens per brief, which is why it is off by default.
- Language. Answers in languages other than English usually need more tokens for the same content.
- Scoring volume. Each scored record is a request. Automatic or scheduled scoring over thousands of records adds up, even though each request is small.
Choosing models: main and fast
AI Engine uses two model settings. The main model handles chat, record actions, meeting prep, scoring, PDFs and Magic Snap. The fast model handles web-search steps and Help mode. Putting a cheaper model on the fast setting lowers the cost of web features without affecting the quality of main answers. Model prices differ a lot between tiers, so the main model choice sets the baseline for every request.
Monitoring what you spend
Every AI action writes an activity log record with:
- input and output tokens as reported by the provider,
- cached tokens, web searches and the number of requests,
- the model used, and an estimated cost based on a built-in list of provider prices.
The Activity Monitor in AI Engine summarises calls, tokens and estimated cost per month, per feature and per user. The estimate is for monitoring and comparison; your provider's dashboard and invoice are the authoritative numbers, and the two will not match exactly.
For Azure OpenAI, include the model name in the deployment name (for example gpt-4o-prod), so the estimate can match the right price.
Six ways to keep usage down
- Enable only the fields users actually need.
- Use a cheaper fast model for web-search steps and Help mode.
- Keep web search in meeting prep and company verification in scoring off unless they add real value.
- Filter which records are scored automatically, and schedule scoring batches outside business hours.
- Keep prompt data sources as small as the use case allows.
- Set spending limits in your provider's console if it offers them.
Why bring-your-own-key is easier to budget
With your own provider account, you see every token on your own invoice, you control rate limits and spending caps, and you can switch models or providers when prices change. There is no hidden margin on AI usage and no separate AI credit system to track.
Summary
- Most actions take one or two requests; size, not count, drives cost.
- The field allowlist, web search and scoring volume are the big levers.
- A cheaper fast model lowers the cost of web features.
- The Activity Monitor shows tokens and estimated cost per feature and per user; the provider invoice is the final number.
Want an estimate for your own org? Book a demo and we will look at your use cases together.
See AI Engine on your own data
Native Salesforce package. Your own OpenAI, Claude or Azure key. 5-day free trial, no credit card.
Book a 30-min demo