Skip to content

Getting Started

AI Monitor ingests CloudWatch metrics via Metric Streams, learns normal behavior per metric using statistical baselines, and alerts on deviations automatically. This guide covers the path from deployment to your first detections.

For the latest detailed information, see the Getting Started PDF guide in Guides & Downloads.

Prerequisites

  • A running Log Processor stack in the same AWS account and Region. AI Monitor deploys as a guest onto its shared VPC, subnets, and OpenSearch domain, so you need the Log Processor stack name before you deploy.
  • An AWS Marketplace subscription to AI Monitor.
  • Permission to launch CloudFormation in your account.

Deploy

  1. Find AI Monitor in AWS Marketplace and select your tier.
  2. Click Continue to Subscribe and accept terms.
  3. After a minute, click Set up your account.
  4. You are presented with a fulfillment page offering Deploy New or Update:
    • Deploy New — deploy one or more uniquely named stacks, for the tier purchased or any lower tier.
    • Update — previously deployed stacks appear below (with a green check); click one to populate the stack-name field.
  5. CloudFormation opens with the template pre-loaded.

Watch the fulfillment page

Deploy time

AI Monitor is a guest on the Log Processor stack's shared VPC and OpenSearch domain, so there is no cluster to initialize — a typical deployment is ready in about 10–15 minutes.

Configure the stack

Fill in the parameters below; you can typically leave the others at their defaults.

Parameter Required Description
ConfirmStackName Yes Copy and paste your stack name here to confirm it meets the stack-name criteria.
LogProcessorStackName Yes Name of your deployed Log Processor stack. AI Monitor resolves the shared VPC, subnets, and OpenSearch domain from it, so this must be an existing stack in the same account and Region.
AlertEmail Yes Comma-separated email address(es) (up to 5) for anomaly alert notifications. You will receive an "AWS Notification - Subscription Confirmation" email to accept.
EditorCertificateArn Yes ARN of an ACM certificate for HTTPS access to the Subscription Editor.
EditorDomain No Custom domain for the editor (e.g. monitor.example.com). Provides a stable URL that persists across stack updates. Leave empty to use the ALB-generated DNS name, which changes if the stack is recreated.
CrossAccountIds No Comma-separated AWS account IDs allowed to send metrics (advanced tier and above). Leave empty for single-account.
OrganizationId No AWS Organization ID (e.g. o-abc123xyz). If set, any account in the organization can deliver metrics without listing individual account IDs.
EditorMfa No MFA requirement for the Subscription Editor (ON, OPTIONAL, or OFF). Defaults to OPTIONAL.

Subscribe

Open the subscription editor (linked from your deployed application's landing page), browse CloudWatch namespaces, and select the metrics and dimensions you want to monitor. Saving a subscription updates the CloudWatch Metric Stream's include filters so only subscribed namespaces stream — adding the first subscription in a namespace starts streaming; removing the last one stops it.

A single subscription can monitor all matching resources independently by using regex wildcards in dimension values.

Baseline

After you subscribe, the detector builds a model per metric over the training period — 7 days on most tiers, 14 days on advanced and above. During this phase, alerts can still fire (a genuine spike should alert immediately), but low-confidence alerts are marked while the baseline is still training. To suppress alerts entirely during initial learning, set the baselineDays field on the subscription.

Detect

Once the baseline is ready, deviations trigger alerts that include an anomaly score (0–10) and any correlated metrics. Alerts are delivered via SNS and carry message attributes (severity, score, namespace, metricName, subscriptionId, description, accountId) you can use for filter-based routing to email, webhooks, Slack/Teams, PagerDuty, Lambda, SQS, and more.

AI analysis and model selection

On advanced tiers and above, AI Monitor uses Amazon Bedrock to add a plain-language root-cause explanation to anomalies and to power the AI Suggest and Tune features in the subscription editor. All inference runs inside your account and Region — no data leaves your environment.

A sensible Claude model ships as the default, so nothing is required to get AI analysis working. If you want to change it, open the editor's FAQ → AI model panel (admin only):

  • Pick a model from the curated list, click Test this model to confirm it works in your account, then Apply and Save. The change takes effect on the next detection — no redeploy needed.
  • The panel shows the model in use, a relative cost indicator, and a retirement warning when a model is nearing or past its Bedrock end-of-life date, so you know when to move to a newer one.

Bedrock models are enabled per AWS account. If Test reports that a model isn't available, enable it once in the Bedrock console (Model access) — some models also require accepting an AWS Marketplace agreement before first use.

Running multiple instances

A single tier entitlement lets you deploy one or more instances, at your selected tier or any lower tier — give each a unique stack name and point it at the Log Processor stack it should monitor. This is useful for separate environments (dev vs prod) or different teams. Standard AWS limits apply.

Teardown

To remove AI Monitor, run the cleanup.py script to clear resources the stack can't delete on its own, then delete the CloudFormation stack. Running cleanup.py before stack deletion avoids leftover resources that would otherwise block or complicate removal.

See the Getting Started guide (PDF) for the full teardown procedure and ordering details.

Further reading

The published product includes additional reference guides distributed as downloads: a Subscription Editor Guide, OpenSearch Guide, User Management Guide, Single Sign-On Integration Guide, and Monitoring Guide — all available in Guides & Downloads.