# AI features and data handling

Axiad Mesh includes AI-powered features that help you investigate your data and
set up your tenant. This page describes where AI appears in the product, what
data those features can access, how sensitive values are protected, how AI
activity is audited, and which questions to raise in your organization's
privacy review.

## Where AI appears in Mesh

### Mesh AI assistant

The Mesh AI assistant is available from anywhere in the dashboard through the
floating action button. You can ask free-form questions about your assets,
identities, and risks in plain language, or start from guided prompts such as
"Show me expiring certificates." Answers are grounded in your own data — the
assistant works from the same data you see in the dashboard.

### Conversational tenant onboarding

When an administrator signs in to a tenant that hasn't completed setup, Mesh
opens a full-page, guided onboarding experience instead of the dashboard. The
onboarding assistant walks you through capturing your organization profile,
setting up your identity provider, choosing recommended connectors, and
validating each configured integration.

How the flow behaves:

- Progress is saved at the tenant level and shared across all administrators,
  so any administrator can resume exactly where the flow left off.
- You can skip onboarding. If you do, you return to the flow on your next
  sign-in.
- Users without administrator access go straight to the dashboard.

### Guided Universal REST connector setup

The Universal REST API connector uses a guided, conversational setup. It walks
you through the application name, the base URL, the OpenAPI or Swagger
specification, and the authentication scheme, then validates the specification
and helps you select which data to collect.

### Agent-to-agent (A2A) integrations

Mesh supports the open [A2A (Agent2Agent) protocol](https://a2a-protocol.org)
in both directions, so the Mesh AI assistant can participate in your broader
agent ecosystem:

- **Outbound — Mesh calls your agents.** Administrators can register external
  A2A agents with Mesh. Once registered, the Mesh AI assistant can discover
  and invoke them during an investigation — for example, asking a remediation
  agent in another system to act on a finding. Each registered agent uses the
  authentication your administrator configures (API key, bearer token, or
  OAuth 2.0), and credentials are stored encrypted.
- **Inbound — your agents call Mesh.** Your organization's own AI agents can
  connect to the Mesh agent over A2A and ask it questions or delegate
  Mesh-related tasks, with the same authentication and tenant isolation as
  any other access to your data. Mesh publishes a standard A2A **agent card**
  describing its skills and how to authenticate — see
  [Connect your agent to Mesh](./connect-your-agent-to-mesh.md) for the full
  integration guide.

Nothing is connected by default: both directions require explicit
administrator configuration.

## Data access and isolation

The assistant only sees data your signed-in account is permitted to see, and
your organization's data stays fully isolated. Answers reflect your own
permissions in Mesh — the assistant doesn't grant access to anything you
couldn't already view in the dashboard.

## Sensitive-data protection

Every AI request passes through Mesh's PII protection gateway before it
reaches any external large language model. The gateway works in both
directions:

- **On the way out**, personal data in your request and in the tenant data the
  assistant consults — values such as names, email addresses, and account
  identifiers — is replaced with opaque tokens. The tokens carry no meaning
  outside your conversation, so personal data is not sent to the model
  provider.
- **On the way back**, the gateway restores the original values in the model's
  response before you see it — you read real names and addresses; the model
  only ever saw tokens.

Token mappings are scoped to a single conversation, held only for the life of
that conversation, and expire automatically. The categories of data that are
tokenized are centrally defined and maintained by Axiad.

## Auditability and feedback

- AI interactions are recorded in the [audit log](../admin-guide/audit-logging.md), so you
  can review how the assistant has been used in your tenant.
- You can rate each assistant response with thumbs up or thumbs down.

## Current limitations

The following capabilities aren't included today:

- **No direct changes in Mesh** — the assistant can't modify data inside
  Mesh on your behalf, for example renewing a certificate. Actions in
  external systems are possible only through the A2A agent integrations your
  administrators explicitly configure.
- **No outbound-data review** — there is no review or approval step for
  outbound data before it is sent.
- **No multi-region or data-residency controls.**
- **No usage-analytics view** for the assistant.
- **No advanced feedback workflows**, such as flagging or escalating a
  response.

## Questions for your privacy review

Some details that your privacy or security team may ask about aren't covered in
this documentation, including:

- Which large language model provider processes requests.
- How long prompts and responses are retained.
- Whether your data is used to train models.
- Whether the assistant can be turned off for your tenant.

For your organization's data-processing and subprocessor terms, contact your
Axiad representative.