Prompt insights: what the models are asked to do, and by whom
Prompt insights in Cureman show what people and services ask the models to do. The module classifies the requests that pass through the AI gateway by task, such as coding, extraction or translation, and attributes tokens and cost to the user and the tag the caller reported. Where the figures support it, it recommends a better-suited tool, for example a seat for people whose coding through the API costs more than the seat would. No prompt text is stored unless the organisation switches prompt samples on.
Token counts say how much was used. They do not say what for. Prompt insights label every request that passes through the gateway with the kind of task it was, attribute it to the person and the feature that sent it, and compare what that work costs with the tools the organisation could use instead.
What it reads
The module has no data source of its own. It works on the records the AI gateway keeps for each request: task type, tokens, cost, and the user and tag the caller reported. The task type is set once, at the moment a request passes through and is classified, either by fixed rules or by the small model a routing policy can use. What is stored is the label. Nothing on the insights pages reads prompt text.
| Task type | Typical requests |
|---|---|
| Coding | Writing, explaining or fixing code |
| Extraction and classification | Pulling fields out of text, labelling, yes-or-no questions |
| Summarisation | Condensing a longer text |
| Translation | From one language into another |
| Writing and editing | Drafting and rewording texts |
| Analysis and reasoning | Comparing, evaluating, explaining causes |
| Conversation | Short back-and-forth without a clear task |
| Other | Everything the rules cannot place |
Attribution to people and features
A request counts for a person when the caller says who it is for: through the header x-gateway-user, the user field of the OpenAI formats, the user id in an Anthropic request's metadata, or the signed-in user forwarded by the chat front end of the AI workspace. It counts for a feature or a job through the tag in the header x-gateway-tag. The page shows which share of requests carries a user at all, so gaps in attribution stay visible.
Usage per person is personal data. Names appear only for roles that hold the permission to view people, and cost figures only for roles that may see spend. How to set up keys, tags and users is described in Allocating AI costs to teams.
What it shows
- Task mix: requests, tokens and cost per task type for the chosen period.
- People: the users with the highest cost, the task each spent most on, and their cost per 30 days.
- Tags: the tags with the highest cost, usually a service or a project.
- Long conversations: requests with many messages and their share of all input tokens.
What it recommends
| Recommendation | How it is computed |
|---|---|
| A seat for heavy use through the API | People whose metered cost for the tasks a service's playbook names, coding for example, exceeds the price of that service's seat by a set margin: their token cost per 30 days minus the seat price, times twelve |
| The context optimiser for long conversations | Raised when conversations with many messages carry a large share of input tokens and the optimiser is not booked. It sizes the question and states no saving |
The seat recommendation says whether the seat's list price has been verified, and states its assumption: that the seat replaces this traffic, while work beyond the seat's usage limits stays metered. Two neighbouring findings come from other modules and appear in the same list. Repeated prompts that could be cached are found by the gateway. Holders of a chat seat whose measured use would cost less by token in the AI workspace are found by licence optimisation, where that module is booked.
From a recommendation to done
Every recommendation is a finding with a calculation and an action plan. For the seat recommendation the plan is a manual step: a person buys and assigns the seats and tells the people concerned. For long conversations the plan holds one prepared step, booking the context optimiser, which starts in observe mode and changes nothing. A prepared step takes effect only after a person approves it in the list of decisions.
What it needs and what it measures
Prompt insights require the AI gateway, because every figure comes from gateway traffic. The module meters nothing for the monthly statement.
Related reading
Questions and answers
Do prompt insights store what people type?
- No. The module works on what the gateway records for each request, which is the task type, token counts, cost, hashes and the user and tag the caller reported. Prompt and response text is not stored. The one exception is a setting the organisation has to switch on itself, prompt samples, which keeps the beginning of each request's system prompt.
How is the task type of a request determined?
- It is set when the request passes through the gateway and is classified. By default, fixed rules look at the wording of the latest message and at features of the request. Where the routing policy uses a small model as classifier, that model assigns the label. In both cases only the label is kept, and the page shows how many requests each method classified.
How does a request get attributed to a person?
- The caller has to say who the request is for. The gateway accepts a request header for the user, the user field of the OpenAI formats, the user id in the metadata of an Anthropic request, and the signed-in user that the chat front end of the AI workspace forwards. Requests without any of these are counted as not attributed.
Who can see usage per person?
- Only people whose role holds the permission to view people. Everyone else sees how many people were active and no names. Cost figures are likewise shown only to roles that may see spend.