AI and your data, on your terms.
TurboFinOps uses AI to explain findings, suggest plans and summarize posture. It runs on a bring-your-own-provider model so you choose the vendor, control the data flow and keep cryptographic control of the keys.
How AI fits in
TurboFinOps is not an AI product wrapped around a chatbot. It is a multi-cloud control plane that uses AI sparingly, where it materially improves the human workflow: explaining a finding in business terms, drafting a remediation plan, or summarizing posture for a non-technical reader.
AI is optional. Every customer-visible workflow remains usable with AI disabled, and none of the deterministic features (scanning, scoring, action execution, audit evidence) depend on a model call.
Bring your own AI (BYOAI)
AI capabilities are activated by configuring a provider credential in your organization settings. The credential is encrypted, scoped to your organization and only used to authenticate upstream requests on your behalf. Removing the credential immediately disables AI features for the organization.
| Provider | Default model | Provider terms |
|---|---|---|
| OpenAI | gpt-5.4 (default), other GPT-family models | Open ↗ |
| Anthropic | claude-sonnet-4-6 (default), Claude family | Open ↗ |
| Azure OpenAI Service | Customer-deployed models in customer Azure subscription | Open ↗ |
| Google Gemini | gemini-3-pro (default), Gemini family | Open ↗ |
| AWS Bedrock | Metered for AI cost attribution (Converse / InvokeModel) on Enterprise | Open ↗ |
| xAI Grok | Available where enabled by the organization | Open ↗ |
Customers are responsible for accepting the upstream provider's terms and for confirming that the provider satisfies the customer's own regulatory requirements.
What may be sent to your AI provider
- Aggregated finding text and rule identifiers (for example, "FIN-CMP-001: 14 unattached EBS volumes detected").
- Resource metadata derived from cloud inventory: type, region, tag keys, identifiers in a form the customer chose to share.
- Cost and utilization summaries used to explain a recommendation.
- User-typed prompts in the AI copilot interface.
What is never sent
- Customer cloud credentials, access keys, or service-account JSON.
- Workload contents, application logs, file contents stored in object storage, or database row data.
- Customer AI provider API keys (these remain encrypted at rest and are used only to authenticate the upstream request).
- Personal data of end users beyond what is contained in the resource metadata the customer chose to scan.
Safeguards
Bring-your-own provider
AI capabilities require an organization to configure its own provider credentials. We do not bundle a shared model. Customers control which provider sees their context and can revoke at any time.
Encrypted keys at rest
Provider API keys are encrypted at rest using AES-256 envelope encryption with customer-configurable KMS (AWS KMS, Azure Key Vault or GCP Cloud KMS via HSM integration) and are never returned in API responses or logs.
No training on customer data
Where the upstream provider offers a no-training contractual mode (for example OpenAI business terms, Anthropic commercial terms, Azure OpenAI), TurboFinOps uses that mode and does not opt-in to model improvement programs.
Per-org isolation
Prompts are constructed from data already scoped to a single organizationId. The AI service never aggregates prompt context across tenants.
Deterministic fallbacks
If the configured AI provider is unavailable or fails validation, the platform falls back to rule-based recommendations rather than blocking the workflow.
Grounded answers, no invented numbers
Finn AI natural-language answers are generated from a deterministic spend analysis computed from billing-backed data. The model is instructed to use only those figures and never to invent numbers; with BYOAI off, the deterministic explanation is returned verbatim.
Audit trail
AI provider configuration changes (provider, model, region) are recorded in the audit log with actor and timestamp.
Transparency commitments
- AI output is a suggestion, never the sole basis for a state-changing action.
- Every cost-saving or remediation suggestion is grounded in real inventory and findings, not an AI inference about your environment.
- Finn root-cause analysis walks a deterministic chain (anomaly → top mover → resource → change events) built from your data; the AI only narrates the likely cause from that chain.
- The conflict guard and approval workflow apply equally to AI-suggested and rule-suggested actions.
- The Finn remediation agent only drafts action requests — every one routes through the conflict guard and waits for a human to approve it in the Actions queue. The agent never executes a change itself.
- Customers can disable AI entirely at the organization level without losing any non-AI feature.
A note on AI accuracy. Language models can produce plausible but incorrect statements. Treat AI explanations as starting points, not authoritative answers. Always validate destructive actions through the approval workflow and the audit trail, both of which operate independently of the AI layer.
Disabling AI
To disable AI for your organization, remove the provider credential under Settings → AI provider. All AI surfaces (explain, plan, summary, copilot) immediately revert to deterministic fallbacks. No further provider calls are made. Existing AI-generated suggestions remain in the audit log for evidence purposes.
Related
Find recoverable spend before the next invoice lands.
Connect one AWS, Azure or GCP scope, approve the safest savings actions, and give finance a receipt when the savings verify.