On-premises AI for IT and security teams

Your IT and security tools, in one conversation.

IT Sentinel runs on your own hardware, reads your SIEM, endpoint protection, firewall, Active Directory, Microsoft 365, ticketing and more, and answers in plain language, with the evidence.

What PC is John using?

What PC is John using?

John Demo is signed in on PC-042 (10.20.30.42). Last sign-in: today, 08:52.

Sources: Active Directory, SIEM agent · read 09:14 today

Which computers have AnyDesk installed?

4 computers have AnyDesk installed: PC-017, PC-042, PC-108 and LAPTOP-23.

Source: help-desk software inventory · read 09:20 today

Did last night's backups run?

17 of 18 jobs finished. One failed: DEMO-FILE-02, because the repository was full.

Source: backup server · read 07:05 today

Any critical alerts in the last 24 hours?

One. A sign-in from a new country, followed by a new mail-forwarding rule, on Demo User's account. The other 212 alerts were routine.

Sources: SIEM, Entra sign-in log, Exchange Online · read 09:31 today

What keeps failing to sign in?

WEB01's computer account, about 14 times a minute against SRV-SFTP: a scheduled job uses an account that doesn't exist there. Find the job and give it a valid account.

Source: SIEM sign-in events · read 09:40 today

Why can't the label printer reach the print server on port 9100?

The office firewall blocks it. The printer listens on TCP 9100, but no policy lets the print server open that connection. Add one, for TCP 9100 only.

Sources: the firewall's own policy lookup, a port check of the printer · read 10:02 today

Which shared mailboxes can be signed into directly?

31 of 40 shared mailboxes have an account that can sign in with a password. Listed with who has access to each, and the steps to block sign-in.

Sources: Exchange Online, Entra ID · read 10:15 today

How many calls did support miss today?

7 missed calls, all between 12:00 and 13:10, when one agent was logged in. 5 callers are customers in the CRM; 3 left a voicemail and haven't been called back.

Sources: phone system call records, CRM · read 13:30 today

What does error 0x80180014 mean?

"Your organization does not support this version of Windows." An Intune enrollment restriction is blocking it: allow Windows (MDM) for that user.

Source: Microsoft's enrollment troubleshooting documentation

Is anything exposed to the internet that shouldn't be?

2 items. RDP to SRV-APP02 is open to any internet address, and the firewall's admin page answers on the internet side. The other 14 published services are limited as they should be.

Source: firewall policies and published services · read 10:40 today

Are my domain controllers healthy?

Yes, all 3. Replication is up to date on every link, the clocks agree, and every role holder answers. DC03, added last week, was picked up by itself.

Source: the domain controllers, read live · read 10:50 today

What's in quarantine from the last 30 days?

312 messages held: 268 spam or bulk, 31 phishing, 9 high-confidence phishing and 4 malware. The 13 dangerous ones are listed on their own; 2 messages were released, both on request.

Source: Exchange Online quarantine · read 11:10 today

Can any user get a certificate as someone else?

Yes, through one template. WebServer-Copy lets any domain user enrol and put any account's name in the request. Turn off "supply in the request" or limit who can enrol. The other 11 templates are fine.

Sources: the internal certificate authority, Active Directory · read 11:20 today

Who reset the domain admin's password?

Dana Demo, today at 08:41, from PC-IT-03. admin-ops counts as privileged because it is in Domain Admins. The Teams alert went out the same minute.

Source: domain controller security events · read 09:05 today

Try:

Illustration with demo data, not a live system.

Tell us about your environment → See how it looks

One assistant instead of a dozen consoles

Answering a simple question means opening the directory, the endpoint protection console, the security logs, the asset inventory and the firewall, one after another. IT Sentinel reads them all for you and answers in one place. It runs on your own hardware, in your network, with its own AI model, so no question and no company data ever goes to an outside AI service.

Ask

"Is anything exposed to the internet that shouldn't be?" One question, and it queries every relevant system.

Get told

Alerts arrive in Teams already investigated. A daily briefing and a findings list, each with a one-click ticket and what your other systems know about it.

Hunts

Every hour it searches your logs for problems no single alert shows: sign-ins that keep failing, PCs that can't reach a domain controller, services that keep crashing, where lockouts come from, and computers nobody monitors. Each comes with the likely cause and the fix.

Changes, explained

When someone edits a Group Policy, a sign-in policy, a privileged account's password or who receives a mailbox's mail, Teams gets one card saying who did it and what changed, in plain words. Code writes it from the system itself, not the AI.

Review

One button, or one sentence: "review our firewall", "audit our backups". Findings by severity, with the fix for each. It also finds the machines nobody is watching.

Cloud check-up

Microsoft 365 and Google Workspace, checked every day against a security baseline: the evidence and the fix for every gap, next to what is actually happening in the tenant.

AI use, on evidence

Which AI tools and services people use, by person and PC, from the endpoint, the firewall and Microsoft 365, including Copilot agents employees build themselves. A ready-made dashboard in your SIEM lets your CISO look directly, and the Sentinel answers from the same data. It names a person only on evidence, and never reads prompts or content.

Investigations on record

One click turns an investigation into a help-desk knowledge article. The facts come from the checks it ran; the AI only drafts the summary, and a person signs off. Next time, it brings the record back. An alert can also be linked to an existing ticket or recorded as a change, so later, "why was this changed?" has an answer.

Vulnerabilities by real risk

It ranks your vulnerabilities by how likely they are to be exploited in the next 30 days, not just by their severity on paper, so the few that matter come first.

What the internet sees

Every day it checks what the internet sees of you: your public addresses, public certificates and look-alike domains. It reads public data only.

Your decisions stick

Mark an alert as a false positive, with a reason and an expiry date, or accept a finding as a risk until a review date. It stays quiet until then, and comes back if things get worse. Some alerts can never be muted, and a mark that suddenly matches far more than usual is flagged.

Prove the fix

After you change something, ask "did my fix work?" It re-runs the check live: fixed, still there, or new.

Behind every answer: read-only tools fetch live data, a local model on your own GPU writes the answer, and code checks every concrete claim against the data before you see it. Where a mistake would matter, such as reading a cloud change or a sign-in policy, the decision is made by fixed checks in code, not by the AI.

What you'd ask before letting an AI near your network

Where does my data go?

Nowhere. The model, the data, the logs and every answer stay on your hardware. Nothing is sent to an AI service, or to us. Even the vendor documentation it reads for how-to answers is kept on your hardware.

What access does it need?

A dedicated account for each system, read-only except for help-desk tickets. You create the accounts, and you can see and revoke each one.

Can it change anything?

Nothing on its own. It finds problems and writes the exact fix; a person makes it. The one thing it can change is a help-desk ticket, which it opens or updates when you ask.

What if it makes something up?

Code checks each concrete claim against what the tools returned. A claim that isn't in the data is corrected or marked "not confirmed". For quick yes/no decisions the AI also says how sure it is; when it isn't sure, the Sentinel doesn't rely on it. Every command in an answer is checked against the vendor's command reference first, and anything that doesn't match is marked.

Who can use it?

Your own directory accounts, limited to a group you choose. Roles can give managers and help-desk staff different tools and data.

What if it stops?

A watchdog checks the SIEM, the model and its GPU, and posts to Teams when one of them stops. The SIEM's analysis engine restarts itself within minutes.

How does it get better?

Your team rates its answers. The ratings, and the full record of what it checked, stay on your server: wrong answers are reviewed there, and good ones become tests it must pass after every change. Nothing leaves your network, and nothing trains the AI on your data.

What hardware?

No server needed. A desktop with one NVIDIA graphics card is enough, for example an older Core i5 with 32 GB of RAM and a 12 GB card.

If it has an API, IT Sentinel can talk to it

IT Sentinel isn't tied to particular products. Any system with an API can be connected, read-only, so it works with the tools you already run. These are connected and in daily use today:

SIEM / XDR
Wazuh, with Sysmon and Windows audit logs
Endpoint protection
SentinelOne, Palo Alto Cortex XDR
Firewall
FortiGate, one site or several
Directory
Active Directory and Group Policy, one domain or several
Cloud and identity
Entra ID, Microsoft 365, Exchange Online, Azure, Google Workspace
Email security
Perception Point
Help desk and monitoring
GLPI, Zabbix
Backup and virtualisation
Veeam, Rubrik, Keepit, VMware Cloud Director
Phone system
FreePBX / Asterisk
Threat intelligence
abuse.ch, Spamhaus, AlienVault OTX, VirusTotal
AI usage
Copilot activity, firewall AI traffic, AI tools on PCs
Alerts, SOC and compliance
Microsoft Teams, an outsourced SOC, ISO 27001

Running something else? If it has an API, we can connect it. Tell us what you run.

Things it found that nobody was looking for

A log that went quiet

A group of servers had stopped sending a whole category of Windows logs for weeks. No alert fires on silence. Missing data is now flagged the day it stops.

An audit setting that switched itself off

Two kinds of audit policy were overriding each other, turning sign-in auditing off and on every few minutes. The review named both; one was removed.

An update that kept coming back

A server's update kept coming back after every restart. The answer was already in the update client's own failure events, which no alert ever read.

One phone, three addresses

Guest Wi-Fi sessions to a known command-and-control server looked like three machines. Lining up the times pointed to one phone, which was blocked.

A backup that failed in silence

A server's backup failed every night for weeks with a message that explained nothing. It found the reason in the backup system's own records, and it was fixed within the hour.

A million crashes in a week

A program on one PC crashed and restarted over a million times in a week. The flood filled the monitoring agent's queue and left gaps in that PC's log. It named the program, the hours missing and the fix: a delay before the service restarts.

A deployment, not a subscription to someone else's cloud

  1. 1 · TalkYou tell us which tools you run and what you'd ask first.
  2. 2 · SizeWe advise on hardware. No server needed: a desktop with one NVIDIA graphics card, like a gaming PC, is enough.
  3. 3 · ConnectOne account per system, read-only by default. Alerts are routed to Teams.
  4. 4 · TuneRules are tuned to your noise, and answers grounded in your policies.
  5. 5 · Hand overDocumentation, a walkthrough, and help as things change.

First answers usually come within days; tuning continues over the first weeks.

Questions we get asked

Is it air-gapped?

The model and your data never leave your network. IT Sentinel does make outbound read calls to the cloud services you connect (Microsoft 365, a cloud EDR) and to threat-intelligence feeds. Leave those out and it needs no internet at all.

Why doesn't it fix things itself?

By design. It's built to find problems and tell you exactly how to fix them. Making the change takes responsibility, and AI isn't ready to carry that without a person checking. So the fix stays with you, and afterwards it re-checks that the fix worked.

Which AI model does it use?

An open-weight model that runs locally; the exact model depends on your hardware.

How fast is it?

Most answers take a few seconds. A long plan or a full investigation can take up to a minute or so.

Does it work in languages other than English?

It answers in the language of the question. English and Hebrew are in daily use today.

We already have an outsourced SOC. Does this replace it?

No, it works next to it. It can forward the logs your SOC needs in real time, and it answers the questions a SOC doesn't.

Is it a product or a service?

Both: the software plus the work of connecting and tuning it.

Tell us about your environment

Tick what you run and say what you'd ask it first. We reply within two working days. No newsletter, and the form stores nothing.

What do you run? (optional)

Thanks, your message is on its way.

We'll reply within two working days.