AI built for project work,
not chatbot novelty

Both. Onplana ships with dual-provider support: Claude (Anthropic) and Azure OpenAI. Admins toggle one or both per org from the AI configuration screen. Tier abstraction (fast / balanced / powerful) maps to a concrete model on each provider; per-endpoint overrides let you pin, for example, risk detection to Claude and plan generation to Azure OpenAI. If both providers are enabled, an admin-set primary acts as the default and the other is a fallback.

No. Onplana sends prompts to Anthropic and Azure OpenAI inference endpoints under their standard enterprise terms, which do not train on customer inputs by default. We don't fine-tune our own models on your data either. Your project data stays in your org's database, AI accesses it at request time through retrieval over the tenant's own rows.

Yes. The Azure OpenAI provider is configured to point at your Azure subscription's endpoint and deployments. Three deployment slots (fast / balanced / powerful) map to the tier abstraction so the same code paths work regardless of which underlying model names you've provisioned. Result: inference stays inside your Azure tenant, your Azure credits cover the spend, and AI features keep working even if your security review forbids outbound calls to Anthropic.

Most AI features are gated at PRO+ (chat, plan generation, status reports, NL parsing, suggestions) and BUSINESS+ (risk detection, portfolio insights, agent layer). The Schedule Health Check tool runs on all plans (including a free public version with no signup). Cross-project reports are also universal across every plan. Marketing tone aside: the gating is real, the feature flags in the codebase actually enforce it.

Retrieval-augmented at request time. The chat panel passes a typed tool catalog to the model, the model calls tools (list_tasks, get_project, list_risks, etc.) scoped to the calling org's X-Organization-Id, and the tools return only that org's rows. No vector embeddings sit in a shared index, no cross-tenant leakage path. Every tool call is audit-logged on AiOperation with input, output, and the requesting user.

No, by design. AI tool calls write to a preview state first (the AiOperation row records the proposed change), the user confirms in the UI, the operation commits. For low-risk reads (list_tasks, get_project, summarise) AI returns the answer directly. For mutations that change data (delete_task, move_task_to_sprint, advance_proposal_stage) the UI requires a human confirmation. Onplana also exposes a per-org "AI guard rails" setting that tightens this further if your org wants AI to never propose certain operations.

Each paid plan includes a bundled per-seat AI token allowance computed at sign-up. You consume from that allowance at no extra cost. If your org hits the bundle ceiling, an admin can either upgrade plan tier, request a top-up from your sales contact, or set a monthly cost cap (WARN at 80%, BLOCK at 103%) to control spend. No surprise overage bills, the cap rejects requests with a clear "AI quota exceeded" error rather than silently charging.

Yes, via the MCP server at mcp.onplana.com. Onplana implements Model Context Protocol with OAuth 2.1 (Dynamic Client Registration per RFC 7591) and Personal Access Token auth. Claude Code, Continue, Goose, and any other MCP-compliant client can connect and use the full Onplana tool catalog (200+ tools, same surface as the in-app AI). See /mcp-project-management for the full integration guide.

Yes. Org owners can disable AI at the org level (no chat panel, no suggestions, no risk detection runs, no scheduled report generation). The setting is enforced server-side so even a user with a direct API call can't bypass it. SCIM-deactivated or matrix-restricted users also lose AI access through the same gating layer.

AI is bad at: judgment calls between competing priorities, reading interpersonal team dynamics, deciding whether a sponsor is genuinely going to push back, and anything that requires sustained context across more than a few weeks of work. It's good at: surfacing the signals you didn't see, drafting a status report you would have written yourself but slower, parsing unstructured input into structure, narrating math the engine already computed. Use it as a force multiplier on the boring 60% of PM work, not a replacement for the 40% that actually needs a human PM.