AI Configuration
An administrator configures P4SaMD's AI capabilities per organization, among them the Whisper AI Assistant and AI-powered compliance evaluation. This page covers how to enable AI features and connect your own LLM provider. AI configuration was introduced in v3.2.
Only users with the Admin role in the active organization can see and change AI configuration. See Roles and Permissions.
Enabling AI Features
A per-organization AI features toggle in the organization settings gates AI capabilities. When the toggle is off, the AI configuration forms are disabled and Whisper does not appear. Turning it on enables the AI configuration and, once a chat provider is configured, makes Whisper available to your team.
LLM Profiles
P4SaMD uses your own LLM provider credentials, configured on the LLM Profiles settings page (under Settings). Credentials are assigned per workload, so different purposes can use different providers or models. The two workloads are:
- Chat: powers the Whisper conversational assistant.
- Evaluation: powers AI-based compliance analysis (the suggestions Whisper surfaces).
Each workload is configured independently, so you can, for example, use a fast, inexpensive model for chat and a more capable model for compliance evaluation.
Configuring a workload
For each workload (Chat and Evaluation), provide:
| Field | Description |
|---|---|
| Provider | The LLM provider. Which providers you can use depends on the workload; see Supported providers. |
| Model | The default model used for the workload. |
| API Base URL | The provider endpoint (when applicable). |
| API Key | The credential used to authenticate with the provider. |
Click Save to store the configuration. P4SaMD validates the credentials live before saving, so an invalid key is reported immediately: "Invalid API key. Please check your credentials and try again."
Each workload shows its current status (Configured or Not configured) and the date it was last updated. On the Chat workload, Use same configuration as Evaluation copies the Evaluation provider settings across.
Supported providers
The settings page lists four providers, but not every provider works for every workload yet:
| Provider | Evaluation | Chat (Whisper) |
|---|---|---|
| Anthropic | Supported (Claude models) | Supported |
| OpenAI | Supported (GPT and o-series models) | Not supported |
| Azure OpenAI | On request: contact us before configuring it | Not supported |
| Google Vertex AI | Not supported | Not supported |
If you use OpenAI for Evaluation, configure Anthropic for the Chat workload, and don't use Use same configuration as Evaluation.
API keys are stored encrypted at rest (AES-256-GCM) in the database. They are never returned to the browser; the configuration form only indicates whether a key is currently stored. As of v3.3, every AI-powered evaluation workload (Requirement Quality Score, Brownfield analysis, and Whisper compliance suggestions) uses your organization's own configured Evaluation credentials end to end for every request.
Available chat models
A provider can optionally define an allow-list of available models alongside its default model. When an allow-list is set, those models become selectable in the Whisper chat model picker. When the allow-list is empty, only the default model is available. You can also discover the live model list directly from the provider (OpenAI and Anthropic) to populate or refresh the available models.
In chat, users can override the model per message by choosing from the available models. If a chosen model is unavailable or unrecognized, Whisper silently falls back to the default model. Evaluation always uses its single configured model.
Automatic (background) evaluation
Once an Evaluation LLM Profile is configured, P4SaMD evaluates requirements automatically, with no manual action needed. A background AI compliance evaluation runs whenever you:
- Create a new requirement.
- Edit a requirement's content: title, description, acceptance criteria, type, level, priority, or labels. Changes to status, assignee, or other metadata-only edits do not trigger a new evaluation.
- Import requirements in bulk: each imported requirement is evaluated individually.
Whisper's compliance suggestions for a document are kept up to date the same way: a few seconds after any of the changes above, P4SaMD automatically refreshes Whisper's findings for that version, so you don't need to manually ask Whisper to re-check it.
Brownfield analysis is not triggered automatically. You start it explicitly from the Brownfield Analysis workflow; see Brownfield Analysis Overview.
Every automatic evaluation above is an LLM request billed to your configured Evaluation provider. Read Cost management below before enabling AI features on an active project.
Stopping automatic evaluation
Switch off the AI features toggle (see Enabling AI Features) to stop automatic evaluation. While the toggle is off, creating, editing, or importing requirements does not send any request to your LLM provider, and Whisper's findings are no longer refreshed. Your LLM Profiles are kept, so turning the toggle back on resumes automatic evaluation from the next requirement change.
Requirements changed while the toggle was off are not evaluated retroactively. They are evaluated the next time their content is edited.
Cost management
P4SaMD does not cap, rate-limit, or monitor how much you spend with your LLM provider. Every automatic evaluation described above, and every Whisper chat message, is a request billed directly to the provider account configured in your LLM Profile. On an active project with frequent requirement edits, automatic evaluation alone can generate a large number of LLM requests. Cost enforcement is on our roadmap but not yet available. Until then, protect yourself using your provider's own controls.
We strongly recommend configuring spend limits or budget alerts directly with your LLM provider (for example, OpenAI usage limits, Anthropic workspace spend limits, Azure OpenAI quotas, or Google Vertex AI budget alerts) before enabling AI features on a production organization.
Recommended models
For the Evaluation workload, we recommend a low-cost, fast model, such as Claude Haiku 4.5 or an equivalent lightweight tier from another provider. In our internal testing, larger and more expensive models did not produce significantly better results when applying compliance policies, so the cost/quality trade-off generally favors the smaller model for this workload.
Contact us for more detailed guidance on which models we recommend and how they perform for your use case.
How this powers Whisper
Whisper becomes available to your team once AI features are enabled and a Chat profile is configured. The Evaluation profile drives the compliance suggestions Whisper displays for each version. If Whisper reports "No active chat provider configured", assign a provider and model to the Chat workload on the LLM Profiles page.