Your AI provider bills your own account for the tokens AI Engine uses. This article explains how AI Engine measures that usage, how the estimated cost in the Activity Monitor is calculated, and which settings have the biggest effect. It does not cover the price of AI Engine itself; for that, contact Cloud Connected.
How usage is measured#
For every request, the provider reports how many tokens it processed. AI Engine adds up every request that belongs to one action and writes the totals to the action's AI Activity Log record:
- Tokens Input and Tokens Output, as reported by the provider.
- Tokens Cache Read and Tokens Cache Write: input tokens the provider served from, or wrote to, its prompt cache.
- Web Searches: the number of searches the provider ran.
- Callouts: how many requests the action needed.
- Model and Models Used: which models were used, with a per-model breakdown when an action used both the main and the fast model.
The Activity Monitor summarises calls, tokens and estimated cost for the month, calls per day, calls and cost by feature, and tokens and cost per user, with a list of recent log entries. The cache, web-search, callout and per-model figures are on the AI Activity Log records themselves. See Activity Monitor.
How the estimated cost is calculated#
Estimated Cost USD is calculated per request, using the model that request actually used:
- Input and output tokens are priced with a built-in list of provider list prices per model.
- Cached input tokens are priced at a discount (10% of the input rate), and cache writes at 125% of the input rate, where the provider reports them.
- Each web search is added at a flat estimated amount.
- If AI Engine does not recognise the model name, for example a custom Azure deployment name, it estimates with a default rate. Including the model name in an Azure deployment name (for example
gpt-4o-prod) lets AI Engine match the right rate.
The figure is an estimate for monitoring and comparison. Your provider's own usage dashboard and invoice are the authoritative source for what you pay, and the two will not match exactly: provider prices, discounts and caching rules change, and the built-in list changes only with package versions.
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. Comparison questions can add up to 3 follow-up steps, and a query that fails gets one automatic repair attempt. A research question uses a web search request in place of running a query. |
| Next Steps, Draft Email, Summary, custom prompts | 1 |
| Web Check | 1, with web search |
| Company News | 2: a web search and one request to write the news cards |
| Analyze Email | 1 |
| Meet Prep | 1, plus 1 web search request when Meet Prep: include web search is on |
| AI Score | 1 per record, plus 1 web search per record when Verify Company On The Web is on |
| Magic Snap | 1 to read the photo. Research this company adds a web search and one request to assess the company; Find companies adds a web search. |
| Help mode and admin help | 1 |
| Show all, export, column changes | None |
What drives token usage#
- Agent Objects scope. Each chat question includes a description of the enabled objects and fields. More objects and fields mean more input tokens on every chat question, so trimming fields users don't need is the most effective saving.
- Record context. More enabled fields, child objects and activities make every record action larger.
- Conversation length. Recent messages are sent with each question.
- Prompt data sources. Rows from data sources are added to the prompt, up to Max Context Chars (Data Sources) (default 40,000 characters), a field on the AI Engine Profile record. See Limits for how to change it.
- Web search. Search results are added to the request as input, which makes web features much larger than other actions. Meet Prep web search adds roughly 30,000 input tokens per brief, which is why it is off by default.
- Output length and language. Longer answers cost more output tokens, and answers in languages other than English usually need more tokens for the same content.
- Scoring volume. Each scored record is a request. Automatic and scheduled scoring over many records adds up even though each request is small.
Keeping usage down#
- Enable only the fields users need in Agent Objects.
- Use a cheaper Fast model for web-search steps and Help mode. See Supported providers and models.
- Keep Meet Prep: include web search and Verify Company On The Web off unless they add real value.
- Use Filter — Which Records to Auto-Score in AI Scoring so that only the records you care about are scored, and schedule batches outside business hours.
- Keep data source row limits and Max Context Chars (Data Sources) as low as the use case allows.
- Set spending limits in your provider's console if it offers them.