Skip to main content

​
Configure the Modernization AI

KOSTRA can narrate modernization findings in plain language and produce an executive summary, using a large language model you configure and control. The AI only explains the deterministic findings the engine has already computed — it never invents metrics or savings. This guide shows how to enable it, and how it protects your data.

​
What the AI does — and does not do

The modernization engine computes every number: readiness, effort, impact, and estimated saving are all deterministic and evidence-backed. The AI layer sits on top and adds two things: a readable narrative for each finding and a ranked, plain-language recommendation for the chosen priority. It is strictly narration. If you leave it disabled, KOSTRA still gives you complete, evidence-based findings — the AI is an enhancement, not a dependency.

​
Before you begin

You need an API key for a supported model provider. KOSTRA currently supports:
  • Anthropic
  • OpenAI
  • DeepSeek
Support for additional models is planned. Configuring the AI is available to users with the appropriate FinOps role in your organization.

​
Enable and configure

  1. In the left navigation, open Settings.
  2. Open the Modernization AI tab.
  3. Select a provider from the list.
  4. Choose a model offered by that provider.
  5. Enter your API key. The key is write-only — you can set and rotate it, but it is never shown back.
  6. Use Test connection to confirm KOSTRA can reach the provider with your key and model.
  7. Enable the feature and save.
Once enabled, subsequent modernization scans include an AI narrative and recommendation alongside the deterministic findings.

​
How your key and data are protected

KOSTRA stores your provider key encrypted, using the same protection applied to cloud credentials, and it is decrypted only within the modernization worker to make the calls you configured — the key never crosses an internal service boundary in the clear. Because the feature is opt-in and provider-configurable, you decide whether any workload metadata is described to an external model at all. If your policy is that nothing leaves your environment, leave the AI disabled and rely on the evidence-based content.

​
Keeping data in your environment

If your policy requires that no workload metadata be described to an external model, leave the AI disabled and rely on the evidence-based content — the modernization findings are complete without it. Customers who self-host KOSTRA keep the rest of the analysis inside their own environment regardless, and support for additional model providers is planned.

​
If narration fails

Narration is designed to be non-fatal. If the AI step cannot complete — a missing key, a provider outage, or a transient error — the scan still finishes and the deterministic findings are produced as normal. You will simply see evidence-based content instead of an AI narrative for that run. This means enabling the AI never puts your assessment at risk.