Skip to main content

Automation & Analytics

Introducing ARIA: AI for Regulatory Intelligence & Authoring, our new AI agent, built on everything you already know from NuMantra.

Home / Technology and trust

Built for adaptable AI workflows and governed access.

NuMantra works with your quality, IT and regulatory stakeholders to review data flows, roles, system connections and the intended use of AI in each workflow. ARIA keeps expert review within the process and records the approvals and agent actions that govern work. We provide current security and product documentation during evaluation so your team can assess fit against its requirements.

Key takeaways
  • Database-level tenant isolation
  • Multi-vendor models chosen by evaluation
  • Approvals, limits and audit for agent actions

Which controls and technologies does ARIA use?

Tenant isolation

Row-level security in the database, with retrieval that re-authorizes every result.

Flexible AI models

Hosted models from several vendors plus an open-source embedding model on NuMantra infrastructure.

Evaluated model choice

Models are chosen after scored evaluation, then routed by accuracy, cost and privacy.

Agent-ready services

Customer agents call classification, retrieval and validation through governed tools, including MCP.

Approvals and limits

Costly actions wait for approval, and spending limits apply.

Audit trail

Activity records are tagged to the agent and the person it acts for.

How does an agent request flow through ARIA?

An agent authenticates with a scoped, short-lived credential and calls an ARIA tool. Costly runs wait at an approval gate, results return with evidence, and every step is recorded.

Agent→Authenticated request→ARIA service→Approval gate→Result with evidence→Audit record

Where can you find security documentation?

Security documentation, data-handling terms and compliance information are available on request. Contact NuMantra to review them against your quality and IT requirements.

How does ARIA protect tenant data?

  • Row-level security in the database
  • Retrieval re-authorizes every result
  • Access scoped to the user’s role
  • Idempotent job interfaces
  • Activity records for every action
  • Any cross-organisation signal uses a separate governed pipeline

How does ARIA choose and govern AI models?

TaskApproach
ClassificationA cost-efficient model, validated against labelled examples
Fact extractionA stronger model, selected after a scored evaluation on a labelled sample
EmbeddingsAn open-source embedding model on NuMantra infrastructure
DraftingA model matched to the template and the quality evidence for that section

What controls apply to agent access?

01

Scoped credentials

Short-lived credentials limited to what the agent needs.

02

Quotas and spending caps

Per-tenant limits keep usage predictable.

03

Approval gates

Costly actions wait for human approval.

04

Tagged audit entries

Each entry names the agent and the person it acts for.

How do security reviews work?

Share requirements

Send your quality, IT and data-handling questions.

Review documentation

Walk through security documentation and data-handling terms.

Discuss models and data

Agree which models and providers apply to your content.

Agree deployment and roles

Confirm tenant setup, roles and approval rules.

Frequently asked questions

Does ARIA support MCP and agent access?

Yes. Customers’ own agents can call selected ARIA services through governed tools, within the user’s role, with approval for costly steps, spending limits and an audit trail.

Which AI models does ARIA use?

ARIA calls hosted models from several vendors and an open-source embedding model through a common layer, and routes tasks by accuracy, cost and privacy.

How is customer data separated in ARIA?

Tenant isolation uses row-level security in the database, and every retrieval result is re-authorized against the user’s permissions.

What approval controls apply to sensitive or costly actions?

Costly actions wait for human approval, spending limits cap usage, and audit entries record which agent acted for which person.

Where can I find security and compliance documentation?

Security documentation, certifications and data-handling terms are available on request. Contact NuMantra to review them with your quality and IT teams.

Can ARIA use self-hosted or adapted models?

Self-hosted or adapted models can be evaluated for sensitive or high-volume tasks, weighed against infrastructure cost and quality. NuMantra does not build foundation models.

Ready to simplify regulatory knowledge work?

Bring one workflow and a few representative documents. We map the workflow and outline a scoped onboarding plan.

Request a Workflow Review