Skip to main content

Platform Assistant — AI System Information

This page is the system-information notice for the Waxell Platform Assistant, the AI assistant built into the Waxell control plane. It explains what the assistant is, how it operates, what it can and cannot do, and how humans stay in control. It exists so that anyone using the assistant — or auditing it — can verify how AI interaction is disclosed and governed.

You are interacting with an AI system

Every conversation with the Platform Assistant is a conversation with an AI system. This is disclosed directly in the product: the assistant dock carries a persistent AI badge in its header, the chat input is captioned "Responses are AI-generated," and the assistant home page states it before your first message. This page is linked from all of those surfaces via About this AI system.

What the assistant is​

The Platform Assistant is an AI operations advisor embedded in the Waxell platform. Its purpose is to help you operate your AI-agent estate: answering questions about your agents, runs, costs, and errors; explaining platform features; proposing governance policies; and surfacing proactive findings (setup gaps, anomalies, cost hotspots).

How it operates. The assistant is powered by large language models (LLMs). When you send a message, it:

  1. Classifies your intent and plans which data it needs.
  2. Queries your tenant's data only — observability data (runs, traces, costs, errors), your policy configuration, and the product documentation — through scoped, read-only APIs.
  3. Generates a natural-language answer grounded in what it retrieved, with citations to documentation where relevant.

Potential outputs. Text answers, inline charts, navigation suggestions, proposed policy cards, and proactive feed notifications. Outputs are AI-generated content and are labeled as such in the interface.

How to interact with — and override — the assistant​

  • Everything is initiated by you. The assistant answers when asked. Proactive findings appear as feed notifications you can open or ignore.
  • No action without approval. The assistant never changes your configuration on its own. Anything that writes — creating a policy, changing a setting — is presented as an explicit card or approval prompt that you click to execute. Declining or ignoring it is always available and always safe.
  • Override and disengage at any time. You can stop a response, clear the conversation, close the dock, or simply not act on any suggestion. Nothing the assistant says takes effect without a human decision.
  • Reach a human instead. The dock footer's Talk to a human hands the conversation to our support team at any point.

Limitations and accuracy​

Like all LLM-based systems, the assistant has limitations you should factor into how you use it:

  • Answers can be wrong. Generated summaries and explanations can be inaccurate or incomplete, even when grounded in real data. Verify important figures against the underlying dashboards (which the assistant links to) before acting on them.
  • Data-grounded, not omniscient. The assistant only sees your tenant's Waxell data and the product documentation. It cannot see systems you haven't connected, and its picture is only as complete as your instrumentation.
  • Not professional advice. Outputs are operational guidance for using Waxell — not legal, financial, or compliance advice.
  • Performance varies with context. Answers about well-instrumented agents with rich telemetry are more reliable than answers about sparsely instrumented ones. The assistant behaves the same way for every user and tenant — its capability differences track the data available, not who is asking.

Human oversight​

The assistant is designed around human-in-the-loop control:

  • Read/write asymmetry. Data queries and explanations run autonomously; every state-changing action requires explicit per-action human approval in the UI.
  • Role-aware capability. The assistant enforces your platform permissions — it can never do more than the signed-in user is allowed to do, and capability flags ride a per-message signed token that downstream services verify.
  • Full auditability. Assistant conversations and actions are governed by the same observability and audit infrastructure as every other agent on the platform — runs, traces, costs, and governance decisions are recorded and reviewable by your admins.

Technical requirements and lifetime​

  • Requirements. A supported browser and a Waxell account — the assistant is part of the control plane web application; nothing to install.
  • Operation. The assistant runs as a governed agent on Waxell's own runtime, subject to the same policy engine, guardrails (including tenant-scope and internal-leak protections), and monitoring as customer agents.
  • Expected lifetime and maintenance. The assistant is a continuously maintained, versioned component of the platform. Model routing, prompts, and capabilities are updated regularly; material changes to what the assistant can do are reflected in this documentation and the product changelog. It carries no fixed end-of-life date; if it is ever retired or substantially changed, users will be notified in-product.

Data handling​

  • The assistant reads only the requesting tenant's data, enforced by per-message cryptographically signed tokens scoped to your tenant and your permissions. Cross-tenant access is structurally prevented, not just policy-forbidden.
  • Conversations are stored to provide history and continuity and are subject to the platform's privacy policy and your data-processing agreement.

How this information is provided​

The disclosure criteria are deliberate: AI interaction is disclosed at the point of interaction (badge and captions in the chat surfaces, before and during use), in plain language (no expertise assumed — the assistant serves both technical and non-technical platform users), and in layered depth — a one-line disclosure in the UI, this page for the full picture, and the Platform Assistant guide plus the Meet your assistant tutorial as educational material. Reasonably foreseeable misuse — treating generated numbers as authoritative without verification, or expecting professional advice — is addressed by the always-visible accuracy caption and the limitations above.

Contact and support​

This page is reviewed and updated alongside assistant releases. Last substantive revision: August 2026.