AI ON AEX
 

Your engineers are spending their week on diagnostics they've run a hundred times.

AEX is building operational agents for fiber networks, grounded in live network data and real resolution history, with the customer holding the autonomy dial. 

~85% 
of operators lose a fifth or more of the engineering week to manual diagnostics

~25%  
lose more than 60% of that week to the same manual work

50%
picked network troubleshooting as where AI should apply first

 

The real constraint is not capability. It is permission.

A majority of operators polled said they'd allow an agent to diagnose but not to act, and most of the rest wanted human approval before any change. That's the right instinct: a confidently wrong recommendation is a nuisance in a chat window and an outage on a network. So we designed for it rather than around it.

Identity-scoped Execution

Every call runs using the identity and token of the user it acts for. Your existing auth, RBAC, and identity management stay in force. No privileged service account, no second permission surface to audit.

No Free-form Commands

The agent cannot compose arbitrary commands. Actions are constrained to defined action steps, and every change is authorized by a human.

An Autonomy Ladder You Control

You advance autonomy one fault class at a time, on the evidence of that fault class's measured first-pass resolution rate. Not a switch you flip and hope. 

Full Audit Trail

Every diagnostic step, every action, and every authorisation is recorded. 

 

Why this works when a chatbot over an API does not

Every OSS/BSS vendor will claim AI this year, and most will ship a chat window over their existing APIs. General-purpose models fail in live networks in specific, predictable ways: they lose critical detail inside long runbooks, they produce different answers to the same question, they hallucinate actions, and they have never seen the long tail of operator, OEM, and firmware combinations where real incidents live. 

Bounded fault classes

The agent retrieves and reasons over only the fault classes relevant to the event, keeping context small enough for retrieval to stay reliable. Each fault class defines explicit diagnostic steps and branch conditions.

Grounded in real networks

Continuous monitoring of OLTs, switches, BNGs, rectifiers,  batteries, and backhaul. Proactive fault detection, field and third-party coordination, and root-cause analysis and reporting.

Your data stays yours

Run through Microsoft Foundry on Azure, not consumer AI services. Prompts and completions aren't shared with other customers or used to train foundation models. Stored in your designated geography, AES-256 encrypted.

Built on an open standard

Capabilities expressed as use cases served over MCP, the default integration standard for agent tooling.

Where we are now

LIVE NOW AI-assisted network troubleshooting with measured performance scoreboards per fault class
IN PROGRESS An expanding fault-class library and agents across provisioning, coverage and premise, billing and product, customer and service, and field service.
NEXT Self-healing networking: troubleshooting integrated with monitoring so faults are resolved before they become tickets

See the troubleshooting agent work a live fault.