AI Assistant
Definition. The AI Assistant is a console-based operator copilot. It helps a signed-in operator find and summarize supported Noxtica information without turning the assistant into an autonomous administrator.
The problem it solves is investigation overhead. Policies, domains, recent assessments, risk distribution, and audit records can answer an operational question, but finding the relevant slice takes time. The assistant provides a conversational starting point; the underlying product records remain the source an operator reviews.
Key points
- An operator copilotIt helps a signed-in operator find and summarize supported Noxtica information without turning the assistant into an autonomous administrator.
- Read-only toolsThe current assistant tool boundary is read-only. Write tools are not part of the current assistant contract.
- Your permissions, your tenantReads remain subject to the signed-in operator's role and the active tenant.
- AuditableAssistant tool activity is recorded so operators can review how tenant information was accessed.
- Navigation, not evidenceTreat it as navigation and synthesis, not evidence that a person is fraudulent or that a policy is correct.
Operator console
Signed-in operator
The assistant runs in the context of the signed-in operator.
Noxtica
Permission and tenant checks
Tool calls are checked against the operator's current role-based access. Reads are limited to the active tenant.
Noxtica
Read-only tools
Retrieve and summarize supported information. Tool activity is recorded.
Back to you
A summary to review
A narrower investigation path. Verify material conclusions against the linked console records.
When it applies
Use the assistant for questions such as:
- Which policies and rules are relevant to this domain?
- What does recent risk activity look like?
- Which recent fingerprints or audit records should I inspect?
- What context should I collect before a policy review?
Do not use an assistant response as the sole basis for blocking traffic, approving a transaction, attributing intent, or closing an investigation. Verify material conclusions against the linked console records and your normal operating procedure.
How customers use it
- Sign in to the operator console with an account that has the Admin role. Lower roles do not see the assistant button.
- Open the assistant if it is enabled for your tenant.
- Ask a bounded question about the tenant and time period you are investigating.
- Review the returned summary and follow up in the relevant policy, domain, fingerprint, risk, or audit view.
- Make any operational decision through the normal product control or your own application workflow.
The useful output is a concise summary and a narrower investigation path. It does not create a new source of truth, expand the operator’s permissions, or perform a traffic action.
How Noxtica does it
The public contract is defined by the controls around the assistant, not by a particular model:
- Operator-session scope. The assistant runs in the context of the signed-in operator.
- Permission checks. Supported tool calls are checked against the operator’s current role-based access.
- Tenant boundary. Reads are limited to the active tenant and supported product surfaces.
- Usage allowance. Assistant availability and usage limits follow the tenant’s provisioned plan.
- Auditability. Assistant tool activity is recorded so operators can review how tenant information was accessed.
Supported deployments can offer more than one model provider. The available provider set depends on the tenant’s plan and configuration; a provider name is not a guarantee about response quality. Review assistant output before it informs a customer-impacting decision.
Phase 1 — read-only tools
The current assistant tool boundary is read-only. Subject to operator permissions and tenant availability, it can retrieve and summarize:
- policies and rules;
- domains;
- recent fingerprints;
- risk distribution; and
- audit information.
It cannot change configuration, create or delete records, flip feature settings, or challenge, block, or otherwise act on traffic.
Phase 2 — write behind a flag
This heading is retained to make the boundary explicit: write tools are not part of the current assistant contract. Prior roadmap references do not make mutation available or commit a delivery date. Plan integrations against read-only behavior unless a future product document expressly replaces this boundary for your tenant.
Not in scope
It cannot change configuration, create or delete records, flip feature settings, or challenge, block, or otherwise act on traffic.
Policies and rules
Scope
Read-only, subject to operator permissions and tenant availability.
Useful next decision
Decide which configuration screen an operator should review.
Domains
Scope
Read-only, subject to operator permissions and tenant availability.
Useful next decision
Decide which configuration screen an operator should review.
Recent fingerprints
Scope
Read-only, subject to operator permissions and tenant availability.
Useful next decision
Open the source assessment and evaluate its score, confidence, and flags.
Risk distribution
Scope
Read-only, subject to operator permissions and tenant availability.
Useful next decision
Choose a time range, risk tier, or segment for deeper investigation.
Audit information
Scope
Read-only, subject to operator permissions and tenant availability.
Useful next decision
Identify the actor, time, or operation that needs normal audit review.
Outputs and decisions
| Assistant output | Useful next decision |
|---|---|
| Policy or domain summary | Decide which configuration screen an operator should review |
| Recent-risk summary | Choose a time range, risk tier, or segment for deeper investigation |
| Fingerprint summary | Open the source assessment and evaluate its score, confidence, and flags |
| Audit summary | Identify the actor, time, or operation that needs normal audit review |
Assistant prose can omit context or require clarification. Treat it as navigation and synthesis, not evidence that a person is fraudulent or that a policy is correct.
Eligibility and limitations
- The assistant is available only where included and provisioned for the tenant.
- The assistant needs the Admin role. Reads remain subject to the signed-in operator’s role and the active tenant.
- Usage allowances and available read surfaces may differ by plan.
- The assistant does not browse customer systems, execute customer code, or mutate Noxtica through its current tools.
- Questions that require unsupported data, another tenant’s data, or a write action remain outside scope.
Where it fits
The assistant is the “we operate agentically” part of the Agentic Security model. It is distinct from MCP, which lets an eligible external agent read approved Noxtica context, and Know Your Agent, which evaluates supported identities of agents reaching your site.
Who it fits
The assistant fits operators who work in the console and have it enabled for their tenant. It helps them choose where to look next. The questions it suits are listed under When it applies. What each summary points to is in Outputs and decisions.
How teams connect
You write no code to use the assistant. It runs in the console. The steps are under How customers use it. Model routing and provider keys for the assistant are set in Backoffice; see AI assistant, MCP tokens and coding-agent keys. To let your own AI client read tenant data, see MCP Integration.
Limits
The assistant is a helper for operators, not an authority.
- Controlled availability. Access is agreed with Noxtica.
- Assistant text can be wrong. It is not a verdict.
- Risk scores and levels it summarizes are probabilistic and can be wrong. Your team decides what to do with them. See Calibration.
What the assistant cannot reach or change, and where it is available, is listed under Eligibility and limitations.
Common questions
Can the assistant change my policies?
No. Its tools are read-only. How Noxtica does it lists what it can read and what it cannot do.
Can it see more than I can?
No. The permission checks and the tenant boundary under How Noxtica does it apply to every read.
How is it different from MCP access?
Each has its own audience. The assistant serves a signed-in operator in the console. MCP serves an approved external client that holds a token. They are provisioned separately. See Where it fits.
Read deeper
- Agentic Security — the three agent workflows and their decision owners.
- MCP Integration — the read-only external-agent counterpart.
- Features: AI-assisted operations — the source product views and read boundaries behind an investigation.
- Calibration — how to interpret risk outputs before acting.
- Contact — confirm tenant eligibility and provisioning.