Skip to content

Agentic NetOps

The AI engineer for your network.

Any vendor. Any environment. Always under your control.

It investigates the alert, writes the root-cause analysis, updates the ticket and executes the fix you approve.

Acts, not just alerts

10 related alerts correlated into one incident. No engineer in the path.

Runs where others can't

Fully air-gapped, with local model inference. Nothing leaves your perimeter.

You stay in command

Every command previewed and approved before it runs, audited to your SIEM.

Prefer to watch first? Watch real product demonstrations. For your own infrastructure, discuss a paid pilot.

How NetPrompt AI handles an alert

Real execution
  1. Alert Received

    SolarWinds · CHE-FW1 · CriticalAlert ID: 120 · Critical device state
  2. Intake Alert

    Alert normalized and enrichedDevice: CHE-FW1 · Site: CHE
  3. Gather Data (Parallel)

    Device Health · Interface Status · SyslogsImpact Analysis · Historical Correlation
  4. Management Plane Check

    ICMP Reachable · SNMP ReachableCLI Reachable: No
  5. AI Investigation & RCA

    Transient management-plane failurewith correlated network instabilityConfidence stated · Diagnostic gaps noted
  6. Smart Ticket Handling

    Existing incident foundINC0011215 → Relationship: RELATED
  7. Ticket Updated

    RCA, impact, timeline and remediationadded to existing ticket
  8. Acknowledge Alerts

    10 related alerts acknowledged

Execution summary

10

Related alerts

1

Incident

No engineer paged

Full audit trail. SIEM ready.

Executes across

CiscoAristaArubaJuniperFortinetD-LinkDellHPEPalo Alto
LinuxWindows ServerVMware ESXiProxmox VE
120+ platform registry

Any ITSM with an OpenAPI spec connects in a single sitting.

Supported platforms

Works with the infrastructure you already have

NetPrompt AI layers onto your estate rather than replacing it — logging into your devices and reading from the tools your team already runs.

Network devices

Cisco
Arista
Aruba
Juniper
Fortinet
Palo Alto
D-Link

ITSM, monitoring and log analytics

ServiceNow
SolarWinds
Splunk

Also supported today: Linux, Windows Server, VMware ESXi and Proxmox VE, plus AWS resource operations. Anything else — build a governed connector from an OpenAPI specification in a single sitting.

On the roadmap: Microsoft Azure and Google Cloud resource operations · Jira Service Management · Palo Alto Panorama & Cortex · Prometheus and Grafana.

All product names, logos and brands are the property of their respective owners and are shown to indicate compatibility. Their use does not imply endorsement, affiliation or partnership.

The situation

Your team was told they could not have agentic AI. They were told correctly — about everyone else’s.

02:00

The 2 a.m. problem

An alert fires. Someone is woken. Diagnosis starts from zero — device state here, monitoring there, ticket history somewhere else. By the time the picture is assembled, the incident has been running for an hour.

MIXED

The multi-vendor problem

Your incumbent's AI assistant is genuinely good — on your incumbent's equipment. On the other three vendors in your estate it does nothing, so the work falls back to the same small pool of senior engineers who already carry it.

DENIED

The sovereignty problem

Every platform that could help wants your device configurations, addresses and credentials in its cloud. Security has already said no. The capability was never the blocker — the deployment model was.

Acts, not just alerts

Ten related alerts in. One diagnosed, acknowledged incident out.

Observability tells you something is wrong. ITSM tracks that you know. NetPrompt does something about it — and the whole loop is auditable afterwards.

  1. 01

    Alert

    A webhook fires

    The monitoring platform raises a critical alert. No polling delay, no queue, no engineer trigger — the workflow starts the instant the event arrives.

    SolarWinds · ServiceNow · Splunk

  2. 02

    Investigate

    It logs into the devices

    Not a query against a telemetry store — an actual session against the affected equipment, gathering device state, syslog, interface counters and neighbour status in parallel.

    Six OEMs · 120+ platform registry

  3. 03

    Analyse

    It writes the root-cause analysis

    Evidence from the devices is correlated against monitoring data, recent changes, ticket history and logs, then reasoned against the original alert to produce a cited analysis and remediation steps.

    Cited · reviewable · reproducible

  4. 04

    Ticket

    It files the ticket, diagnosis attached

    If no matching incident exists, the ITSM record is created with the RCA already in it. If one does, the analysis is attached to that incident and the alert related to it rather than opening a duplicate — so the engineer who picks it up starts from a conclusion rather than from a blank incident and a device name.

    ServiceNow · opens or relates

  5. 05

    Acknowledge

    It acknowledges, and nobody was paged

    The loop closes. What remains for a human is the decision, not the evidence-gathering that used to consume the first hour of every incident.

    Full audit trail to your SIEM

One execution

10 alerts → 1 incident

No engineer paged.

Up to 70% MTTR reduction; typically 40–60% for L1/L2 incident classes.

See it run against your own alert pattern

Vendor-modelled against internal incident data, not independently validated. We will measure it against your own baseline during a pilot.

See NetPrompt in action

Watch a critical network incident investigated

The recorded firewall event shows parallel evidence gathering, an explicit diagnostic gap, structured RCA and an existing incident updated—without inventing a fix that did not occur.

View the dedicated video page

You stay in command

Every action previewed. Every approval yours.

The question every serious evaluator asks is what happens when the AI gets it wrong. It is the right question, and it has a structural answer rather than a reassuring one.

Approval thresholds, per user and per category

Junior roles are constrained, architects are not, and read-only operations are governed by different rules from configuration writes. Set once, enforced before the model is even scored.

Every command previewed, and editable

Nothing state-changing runs unreviewed. The generated command is shown before execution and can be edited in place — so the engineer stays the author of what touches production.

Tamper-evident audit, forwarded to your SIEM

Every action, approval and rejection is logged and shipped to the customer’s own SIEM. The audit trail lives in your estate, not in a vendor console you have to request access to.

Designed on the assumption the model will be wrong

The control architecture does not depend on the AI being correct. That is the point of it. Where you want measurement, we will instrument command-accuracy tracking during the pilot as an agreed success criterion.

Diagram: NetPrompt AI proposes a configuration change on device CHE-FW1, showing the exact command and the interface state it would alter. A NOC L2 engineer, whose threshold for configuration writes requires approval, can reject, edit or approve. Rejecting or editing means nothing reaches the device. Only approval executes, after which the change is validated and the action, approver and outcome are forwarded to the customer's SIEM.

Runs where others can’t

Runs on your laptop. Runs in your data centre. Runs with the cable pulled.

Sensitive values — addresses, credentials, hostnames, SIDs — are tokenised before any model call, cloud or local, and restored only in the final response. The pattern manager that does it is administrator-configurable, so you decide what counts as sensitive.

On-premises
Your data centre, your hypervisor, your network.
Your cloud
Inside your own AWS or GCP tenancy — not ours.
Fully air-gapped
Local model inference, signed offline update bundles.

Cloud scope, precisely

Model inference via Google Cloud, AWS and Azure; available on the AWS and Google Cloud marketplaces; conversational cloud resource operations for AWS.

Model provider is a separate choice from location — local Native AI, a cloud provider, or a switchable mix, on-premises or in your cloud tenancy. Full breakdown on the deployment page.

Diagram: NetPrompt AI running entirely inside the customer's perimeter — device sessions, tokenisation and local inference stay inside the boundary; only optional external lookups may cross it, and everything still functions fully air-gapped.

Stated plainly: in a fully air-gapped deployment, external lookups — CVE feeds, end-of-life data and vendor documentation — degrade gracefully, because nothing reaches out. Everything that runs against your own estate continues to work.

Any vendor

Their agent is excellent — on their equipment.

Platform-native copilots work properly on their own vendor's hardware. That covers part of your estate. NetPrompt executes across the rest of it under the same interface and the same governance model.

Network devices

  • Cisco
  • Arista
  • Aruba
  • Juniper
  • Fortinet
  • Palo Alto
  • D-Link

Driven by a registry of 120+ platforms — new vendors onboard without a code change.

Servers & hypervisors

  • Linux / Unix
  • Windows Server
  • VMware ESXi
  • Proxmox VE

NetOps and SysOps under one governance and audit model, with a locked-down read-only mode for production hosts.

Your existing tooling

  • ServiceNow
  • SolarWinds
  • Splunk
  • AWS

Anything else — build a governed connector from an OpenAPI specification in a single sitting.

Stated openly: device-level execution is genuinely multi-vendor. Deep controller integration for SD-WAN and wireless is Cisco today (vManage and the 9800 WLC), with vManage also reachable through the OpenAPI connector library. If your controllers are non-Cisco, we will scope those out of a pilot rather than promise parity.

Where it earns its licence

The shift work your engineers dread, run reliably and around the clock.

Not a feature list — four routines that currently occupy rostered hours, and what happens to them.

01NOC Manager

Alert-to-ticket RCA

Event-driven triage on every P3 and P4 that fires overnight. The queue your team currently works through at 08:00 is already diagnosed and ticketed when they arrive.

Read more
02L1 / L2 Engineer

Autonomous morning health checks

The scheduled sweep across sites and device classes that consumes the first hour of the early shift, run on a schedule and reported as an exception list rather than a full readout.

Read more
03NOC Manager

Automated shift handovers

A written handover assembled from what actually happened — incidents, changes, outstanding tickets and their state — rather than from what the outgoing engineer remembered to write down.

Read more
04Head of Network Operations

SLA breach prevention

Continuous threshold and trend monitoring against the commitments you are measured on, with the investigation already done by the time the number becomes a conversation.

Read more

The common thread is that none of these require a human to start them, and all of them currently do. That is the difference between an assistant you open and an engineer who is already working.

Early adopters and partners

Working with infrastructure teams, data centres and service providers

Managed service providers, hosting and data-centre operators, and enterprise infrastructure teams running NetPrompt AI in evaluation and early production.

Redington
IBM
Tata Teleservices
ACL Digital
Binary Global
Hughes Systique
Cyient US
Datadog
Broadcom
Manage Engine
BitsnByte
E2E Networks
Yotta
Signellent
RiseTech Software
NovoTek UK
TDM Networks
NetforChoice

Where we are, honestly

Our ecosystem today includes customers, partners, and organizations evaluating NetPrompt AI through pilots and early-production engagements. We do not publish performance numbers unless they have been measured and validated. Before a pilot begins, we agree on success criteria in writing, establish a baseline, and measure outcomes against it. Where the results are meaningful, we ask the organization whether we can publish the findings as a case study.

Founding-customer terms are available for qualifying organizations willing to participate in a quantified case study and reference calls.

Deployment and AI ecosystem

Buy it the way your procurement already works

Transact through a cloud marketplace against committed spend, or run entirely inside your own perimeter.

Purchasing through a marketplace draws down committed cloud spend, which routinely removes months from enterprise procurement and bypasses new-vendor onboarding entirely. Model inference runs on Google Cloud, AWS or Azure — or fully locally, on premises or air-gapped, with no data leaving your network.

Technology partners

Development and infrastructure relationships — not additional purchase channels. Every sale outside the two marketplaces above is direct.

NVIDIA Inception

NVIDIA Inception Program member · not a purchase channel

Next step

See it run against one of your own incidents.

Begin with a short technical walkthrough. If the fit is clear, proceed to a 30-day paid pilot with success criteria agreed before day one.

Already planning a paid evaluation? Here is how the pilot works:

The fee
Roughly one month of one L1 engineer’s loaded cost — credited in full against your first-year contract.
The baseline
We measure your current incident handling before we deploy. Without that, no honest claim about the outcome is possible afterwards.
The criteria
Three or four, agreed in writing, all measurable. Signed off by whoever owns the budget — not only by the champion.
The exit
If the criteria are not met, you have a measured answer and owe nothing further. That is the point of agreeing them first.

Book a pilot

A 30-day paid pilot, on your infrastructure

Success criteria agreed in writing before day one. The fee is set at roughly one month of one L1 engineer’s loaded cost, and is credited against your first-year contract.

We use these details to scope the pilot and nothing else. No newsletter, no sequence.