The Runtime Control Engine for Agentic AI
Don’t just authorize AI.
Authorize the right action.
Axtor evaluates consequential AI actions across scope, intent, evidence, policy, authority, and risk before they execute. It returns a clear outcome: allow, block, clarify, or require approval.
Currently in development.
Runtime architecture
Control the action before it reaches the enterprise system.
Axtor sits between an AI agent proposing an action and the tool, API, or system that would execute it.
Axtor Control EngineScope · Intent · Evidence · Policy · Authority · RiskALLOWBLOCKCLARIFYAPPROVAL
Illustrative scenarioFictional data. No actions are executed.Runtime control scenario
See Axtor evaluate an action before it executes.
Follow a proposed AI action through the request, evidence, and control decision.
Scenario preview. Press Play to follow the evidence.
1Customer conversation
An end customer request
I was charged even though the service was down. Could you sort this out?
AI
Let me check the charges on your account.
Scripted customer conversation
2Proposed action and evidence
What the AI plans to do and supporting data
Proposed action
Refund $750
Issue a refund to the customer’s original payment method.
Relevant evidence
Subscription charge$750
Monthly plan, billed to account
Outage-related charge$140
Service disruption linked to this charge
Prior credit$50
Recorded as a goodwill credit
3Axtor control assessment
Evaluates the action across relevant control dimensions
ScopeIn scope
Refund workflow
IntentAmbiguous
Which charge does the customer want resolved?
EvidenceInsufficient
Does not establish this action
PolicyPermitted
Fictional refund policy
AuthorityWithin limit
Refunds up to $1,000 permitted by the fictional policy.
RiskFinancial
Changes a customer balance
Runtime outcomeCLARIFY
The evidence links the outage to a separate $140 charge. It does not establish that the $750 subscription is the intended transaction. The $50 credit also needs context.
Illustrative scenarios using fictional data.
This fictional customer service scenario shows how runtime control evaluates more than permission. An AI proposes a $750 subscription refund, while the outage evidence points to a separate $140 charge and an existing $50 credit. Axtor identifies that the intended transaction is not established, so the action requires clarification before it can proceed. The evidence does not establish that a $90 refund is correct.
An AI action can be allowed and still be wrong.
CAN Capability
Can the AI execute this action?
MAY Permission
Is the AI allowed to execute it?
MEANT Intent
Does this action match what was actually requested?
PROCEED? Control
Do evidence, policy, authority, and risk support execution?
Axtor Control Engine
- Scope
- Intent
- Evidence
- Policy
- Authority
- Risk
Across workflows
One control engine. Different consequential actions.
Customer Service
- Human request
- “Refund the charge caused by yesterday's outage.”
- Agent proposes
- Refund $750 subscription payment.
Axtor outcomeCLARIFY
The evidence points to a different outage-related charge. The intended transaction is not established.
Enterprise Access
- Human request
- “Give Maria analytics access for the project.”
- Agent proposes
- Grant production administrator access.
Axtor outcomeBLOCK
The proposed privilege exceeds the scope of the request.
Agentic Workflow
- Human request
- “Make the king disappear from this scene.”
- Agent proposes
- Delete the King permanently from story canon.
Axtor outcomeBLOCK
A scene-level request does not establish authority for a permanent state change.
How it works
Follow the evidence before an action executes.
Customer request
“I was charged even though the service was down.”
Transaction
$750 subscription charge
Support case
Service outage · separate $140 charge
Prior credit
$50 credit issued
IntentClarification needed
Which charge does the customer want resolved?
Start with the request
Identify what was requested and what AI proposes.
Follow relevant evidence
Connect related conversations, transactions, and previous resolutions.
Resolve uncertainty
Clarify missing intent or evidence before execution or approval.
- Try the evidence walkthrough
Our starting point
Starting with customer service.
Axtor is starting with customer service and contact center workflows, where AI agents increasingly move beyond answering questions to taking actions that affect accounts, transactions, refunds, and customer outcomes.
This is the first proving ground for the broader Axtor Control Engine. It evaluates whether a proposed action is sufficiently supported by scope, intent, evidence, policy, authority, and risk before execution.
Why We’re Building AxtorHelp shape Axtor around your workflows.
We’re looking for enterprise teams building or using AI in customer service to help us test and refine Axtor.