Skip to content
RaiCentral

AI Agent Governance & Assurance

Give AI agents authority.Keep control of their actions.

Register agents, define their permissions, and apply policy checks and human approvals to integrated actions. Bring decisions, evidence, and AI spend into one place.

Book a demoExplore the platform

Alpha · Scoped pilots available

Refund requestInteractive example

Requested action

Issue refund of $1,240.00

Billing system · order ORD-48213

Policy decision

REQUIRE_APPROVAL

Refund authority: Exceeds the $500 threshold. Held. Nothing runs until a reviewer decides.

Human approval

Pending · Support lead

Approval applies to this action and amount only

Agent
support-refund-agent v3.2.0
Owner
Customer Ops
Receipt
rcpt_7f3a9d04
Illustrative interface · Sample data

AI agents are taking action. Who controls the authority?

  1. 01

    Unclear ownership and permissions

    Agents are deployed across teams with credentials that outlive their purpose and no named owner accountable for what they do.

  2. 02

    Policies disconnected from runtime

    Governance lives in documents. The moment an agent calls a tool, nothing checks the action against those rules.

  3. 03

    Fragmented evidence and spend

    Logs, approvals, and model costs sit in separate systems, so no one can say what happened, why, or what it cost.

How RaiCentral works

A decision point between agents and the systems they act on.

  1. 1

    Register the agent and define its authority

    Record the owner, version, runtime identity, and the actions and limits it may use.

  2. 2

    Evaluate integrated actions

    Before a governed action runs, RaiCentral checks it against policy and approval requirements and returns a decision.

  3. 3

    Review decisions and evidence

    Every decision produces a receipt; reported execution results and spend are linked to it.

Agent

support-refund-agent · v3.2.0

RaiCentral decision

policy · limits · approval rules

Decision point

Approval

when required

Business system

on ALLOW only

Decision receipt

decision · result · cost

Illustrative flow · Sample data

Platform

Know your agents. Control their actions. Trace the outcome.

RaiCentral governs the actions your applications route through it. Your agents keep their own business logic and prompts.

01 / Know and authorize

A system of record for every agent you operate.

Know which agents exist, who owns them, and exactly what each is allowed to do.

  • Agent inventory
  • Immutable versions
  • Dependencies
  • Accountable owners
  • Runtime identities
  • Scoped credentials
  • Delegation
agents / inventory
AgentOwnerCredentialsStatus

support-refund-agent

v3.2.0 · immutable

Customer Opsscoped · 2 toolsActive

invoice-reconciler

v1.4.1 · immutable

Finance Systemsscoped · 3 toolsActive

repo-maintainer

v0.9.0 · immutable

Platform Engdelegated · MCPReview

procurement-intake

v2.0.3 · immutable

Procurementscoped · 1 toolActive
Illustrative interface · Sample data

02 / Control and protect

Decide at the moment an agent acts.

Integrated actions are evaluated against policy before they run. Approvals are bound to the specific action requested.

  • Action authorization
  • Policy evaluation
  • Action-bound approvals
  • Structured-data access controls
  • Masking
  • Budgets
  • Intervention
policy / support-tier-2

  • billing.issue_refund

    ≤ $500 allow · ≤ $5,000 approval · else deny

  • crm.read_customer

    fields: name, tier · mask: email, phone

  • budget.monthly

    $2,000 · intervene at 90%

Approval request · pending

Refund $1,240.00 · ORD-48213

Approval applies to this action and these parameters only.

customer.email j•••••@•••••.com(masked)
Illustrative interface · Sample data

03 / Observe and prove

Evidence your reviewers can follow.

Trace each decision to its outcome and cost, and report across the whole agent portfolio.

  • Trajectories
  • Decision receipts
  • Recorded execution results
  • Evaluation gates
  • Incidents
  • Portfolio reporting
  • Spend attribution
evidence / receipts & spend

  1. 14:02:11 Input observed

    session ses_88a1

  2. 14:02:13 Action requested

    billing.issue_refund

  3. 14:02:13 Decision · REQUIRE_APPROVAL

    rcpt_7f3a9d04

  4. 14:09:40 Approved · Support lead

    bound to receipt

  5. 14:09:41 Refund executed · billing system

    ORD-48213

  6. 14:09:42 Execution reported · SUCCEEDED

    result recorded

AI spend · this month · by source

  • Estimated$420

    From token counts × list price

  • Metered$1,180

    Reported by integrated gateways

  • Billed$960

    Reconciled with provider invoices

  • Unavailable—

    2 agents without a cost source · shown, not guessed

Illustrative interface · Sample data

Use cases

Where authority matters most.

Enterprise agents on business systems

1
The action
An operations agent updates a vendor record in the ERP.
2
The governance control
Write access limited to approved fields; changes to bank details require a named approver.
3
The evidence produced
Receipt with requested change, decision, approver, and reported result.

Coding agents with tools and MCP servers

1
The action
A coding agent calls an MCP tool to run a migration.
2
The governance control
Tool calls checked against repository and environment scope; production targets denied.
3
The evidence produced
Trajectory of tool calls with each decision and its reason code.

Customer-support agents

1
The action
A support agent requests a refund or account change.
2
The governance control
Thresholds decide allow, approval, or deny; customer PII masked before model use.
3
The evidence produced
Decision receipt linked to the case, plus model spend for the conversation.

Developers

Ask before acting. Report what happened.

Your application requests a decision before a consequential action and executes only on ALLOW. Approval and denial outcomes never execute.

Coverage depends on the integration and capabilities you choose. RaiCentral governs actions your application routes through it; it does not intercept traffic automatically.

$ pip install rai-control-plane-sdk

decision = session.authorize_action(
    tool="issue_refund",
    parameters={"amount": 1240, "currency": "USD"},
)

if decision.allowed:
    # 1. Execute your application's business action.
    refund = billing.issue_refund(order_id, amount=1240)

    # 2. Report the actual result once it has run.
    if refund.succeeded:
        decision.report_execution(client, status="SUCCEEDED")
Full quickstart in the repository

Security & trust

Precise controls, stated plainly.

Product status

RaiCentral is in alpha.

Pilots run under agreed terms. Data protection depends on the integration and configuration you choose. We do not yet hold third-party certifications.

  • Scoped identity

    Each agent runs with its own identity and credentials limited to the actions it is authorized for.

  • Human oversight

    Consequential actions can be held for a named reviewer. Approvals apply to the specific action requested.

  • Evidence access

    Decision receipts and execution records are available to the people responsible for review, under your access rules.

  • Fail-closed configurations

    Integrations can be configured so that if a decision cannot be obtained, the action does not proceed.

FAQ

Common questions

Book a demo

Put your first AI workflow under clear control.

Tell us about one workflow where an agent takes consequential action. We'll show how it would be authorized, approved, and evidenced.