AI analysis overview
CLARITY includes a built-in AI analysis layer that validates and enriches all cost optimization findings. AI reviews your recommendations, insights, and anomalies to add context, verify accuracy, and provide actionable explanations.
How AI analysis works
AI analysis operates in two distinct modes, each designed for a different workflow.
Batch validation
After a manual cloud data sync — or when a cloud account is first connected — CLARITY sends the current findings to the AI for review. The AI evaluates recommendations, insights and anomalies in context and returns:
- Confidence badges —
Agree,ModifyorDisagree - Contextual notes — why it agrees or disagrees, referencing usage patterns and cloud best practices
Scheduled syncs do not run batch validation
The automatic background sync does not call the AI. Badges therefore refresh when you trigger a sync by hand, not on the automatic cadence. If badges look stale, run a manual sync.
The model returns a prioritised subset rather than a verdict on every finding, so a finding it did not rank carries no badge at all — an absent badge means "not reviewed", not "disagreed". Anomaly cards do not currently render badges.
Batch validation runs in the background. Results appear as badges on your recommendations and insights.
On-Demand explain
Click the Explain button on any recommendation, insight, or anomaly to get a detailed AI-powered breakdown. On-demand explanations include:
- Root cause analysis — Why this finding was generated
- Impact assessment — What happens if you act (or don't act)
- Step-by-step actions — Specific steps to resolve the issue
- Risk considerations — Potential side effects or caveats
Explanations are cached for 24 hours, so repeated clicks return instantly.
Reading AI validation badges
Each finding displays a badge after AI review:
| Badge | Meaning |
|---|---|
| Agree | AI confirms the finding is accurate and actionable |
| Modify | AI agrees there is something to act on but would change the recommendation — read the note |
| Disagree | AI believes the finding may be inaccurate or not actionable |
Use these badges to prioritize your optimization work. Focus on findings where both the rule engine and AI agree.
Data privacy
CLARITY redacts data before AI processing:
- AWS account IDs, AWS ARNs, IP addresses, AWS network identifiers (VPC, subnet, security group and similar), AWS access-key ids and email addresses are masked
- Credentials and secrets are never included in AI prompts
Redaction is AWS-specific, and prompts may leave your deployment
Two limits you should know before enabling AI:
Redaction covers AWS identifiers only. Azure subscription GUIDs, Azure resource IDs (/subscriptions/...), GCP project IDs and GCP resource paths are not masked, and resource names are never masked on any provider — they are what makes a finding readable.
Where analysis runs depends on AI_PROVIDER. With claude or gemini, prompts are sent to a third-party API (api.anthropic.com or generativelanguage.googleapis.com) and leave your infrastructure. Only ollama and vllm keep analysis inside your own deployment. If data residency matters to you, use a self-hosted provider — or leave AI_PROVIDER at its default of none, in which case no data is sent anywhere.
TIP
When AI analysis is not available, CLARITY uses deterministic rule-based analysis for all insights, recommendations, and anomalies. The platform is fully functional in rule-based mode.
Next steps
- Learn about cost insights on the dashboard
- Explore optimization recommendations