Skip to main content

Documentation Index

Fetch the complete documentation index at: https://concentrate.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

Anomalous spend detection monitors your daily usage and alerts you when spending deviates significantly from your historical baseline. It runs on a scheduled basis and evaluates spend across organizations, teams, and individual API keys.

How it works

Each evaluation cycle:
  1. Collects daily spend over the last 30 days
  2. Separates the baseline from the detection day — yesterday’s spend is the detection day, everything before it is the baseline
  3. Validates the baseline — the baseline must have enough data points and consistent usage (see Baseline requirements)
  4. Computes a z-score — measures how many standard deviations yesterday’s spend is from the baseline average
  5. Compares against your sensitivity threshold — if the z-score exceeds the threshold, an alert is triggered

Configuration

You can configure anomalous spend detection from the Alerts page in your dashboard.
SettingDescriptionOptions
EnabledToggle the alert on or offtrue / false
SensitivityHow sensitive the detection should behigh, medium, low
ScopeWhat entities to monitororganization, team, user, key

Sensitivity levels

Sensitivity controls the z-score threshold that must be exceeded to trigger an alert:
SensitivityZ-score thresholdWhat it means
High2.0Alerts on smaller deviations — more alerts, fewer missed anomalies
Medium2.5Balanced — catches significant spikes without excessive noise
Low3.0Only alerts on extreme deviations — fewer alerts, may miss smaller spikes
A z-score of 2.0 means yesterday’s spend was 2 standard deviations above the baseline average. For normally distributed data, this corresponds to roughly the top 2.3% of expected values.

Scope

Scope controls which entities you want to monitor for anomalous spend. This is a per-user setting — organizations and teams do not have their own alert configurations. You choose what to watch, and you receive the alerts. You can select multiple scopes simultaneously:
  • Organization — monitors total spend across your entire organization. All org admins, owners, and billing members are included as recipients.
  • Team — monitors spend per team that you are an owner or admin of within your organization. Each team is evaluated independently.
  • User — monitors aggregate spend across all your personal (non-organization) API keys. This covers your personal wallet only — spend billed to an organization wallet is covered by the Organization scope.
  • Key — monitors spend per individual API key that you own. Each key is evaluated independently.
If multiple users in the same organization both enable organization-scope alerts, the organization’s spend is only evaluated once — both users are added as recipients of the same alert.
The User and Key scopes both cover personal API key spend. User monitors the aggregate total, while Key monitors each key individually. Enabling both means you could receive an aggregate alert and per-key alerts for the same spend — this is intentional and mirrors the Organization/Team relationship.

Baseline requirements

To avoid false positives, the detection engine requires a meaningful spending baseline before it will fire an alert. Specifically:
  1. Minimum data points — at least 7 days of data in the 30-day lookback window
  2. Consistent usage — at least 50% of baseline days must have non-zero spend
  3. Grace period — accounts less than 14 days old are never alerted
If any condition is not met, no alert is triggered. This means:
  • New accounts are given 14 days before alerts activate
  • Accounts resuming after a long break (where most of the lookback window is zeros) are not alerted — there is no meaningful baseline to compare against
  • Sporadic, low-frequency users will not receive anomalous spend alerts until their usage becomes consistent enough to establish a pattern
Anomalous spend detection is designed for accounts with regular, consistent usage. If your usage is highly irregular (e.g., large batch jobs once a week with nothing in between), this alert type may not be a good fit — consider using low balance alerts instead.

Edge cases

Flat baseline (zero variance)

If every day in the baseline has the exact same spend (standard deviation is zero), the z-score cannot be computed normally. In this case, the engine falls back to a percentage-based check: if yesterday’s spend exceeds 150% of the baseline average, the alert fires.

Today’s spend is excluded

The detection day is always yesterday, not today. Today’s spend is excluded from both the baseline and the detection value, since the day is not yet complete and partial data would skew the results.

Notifications

When an anomalous spend alert fires, you are notified via your configured channels:
  • Email — includes yesterday’s spend, the baseline average, the percentage increase, and a link to your billing page
  • SMS — a shorter summary with the same key figures
Each entity (organization, team, or key) is subject to a 24-hour cooldown after an alert fires. During this window, no duplicate alerts are sent for the same entity.