Why MSPs need an AI security team, not another dashboard

Attackers have moved down-market. Small and mid-sized businesses rarely have anyone watching after hours, often pay to get back to work, and are now where a large share of attacks land. Coveware’s Q4 2025 data puts the median ransomware victim at just 200 employees, and companies with 11 to 100 employees were the single largest group of victims, at 38%.

For the MSPs who protect these companies, the problem isn’t a shortage of security products. It’s the absence of someone to run them. This post explains why another dashboard doesn’t fix that, what an AI security team does instead, and how to tell a real one from a chatbot bolted onto an alert list.

The real gap is people, not tools

There is no shortage of security tools. There is a shortage of analysts to work them. Cisco’s 2025 Cybersecurity Readiness Index, a survey of 8,000 security and business leaders, found:

• 86% of organisations say the shortage of skilled security people is a challenge.
• 70% already run more than 10 point tools, and 26% run more than 30.
• 77% say that complexity slows down how fast they can respond.
• Only 4% reach the “Mature” readiness level.

Most small businesses have no security budget of their own and nobody in-house who knows security. They rely on their MSP. And most MSPs don’t have spare analysts either. The tech who answers the helpdesk at 9 am is often the same person who gets the alert at 11 pm.

Adding another point tool just adds another console nobody is watching at 3 am. Each new tool brings its own alerts, its own login and its own idea of what “critical” means. The tools multiply. The people don’t.

Where an analyst’s time actually goes

Ask any security analyst what eats their day and the answer is rarely “fixing things”. It’s working out what happened before anyone can act:

1. Is this real? Most alerts are routine: a director on a business trip, an admin script that runs every Monday, a known update tool.
2. What else is connected? An odd sign-in means little on its own. The same sign-in followed by a new inbox rule and a strange program on the laptop is an attack.
3. How far did it get? Which accounts, which machines, which mailboxes, which data.
4. What is the right fix, and is it safe? Ending a session is quick and can be undone. Cutting a whole office off the network is not something you do on a hunch.
5. Did the fix work, and what do we tell the client?

Steps 1 to 3 are where the hours go, and they are exactly the steps a dashboard leaves to a person. That is the work an AI security team should take on.

What an AI security team does differently

An AI security team is a set of AI agents that do the investigation work, not just raise the alert. In Centeye, that layer is called Praxis, and it has two agents:

• Kavach, the security operations agent, works every case day and night. It reads what each Centeye product finds, joins related signals into one case, runs the playbook for that kind of attack and takes each step at the setting you chose.
• Argus, the advisor, explains every case in plain English and writes the weekly report for each client.

The difference from a dashboard is simple. A dashboard shows you forty alerts. An AI security team hands you three cases, each with the evidence attached, the steps already taken and what is left for a person to decide.

You stay in charge of every important decision

The fair worry about AI agents in security is that they will do something drastic on their own. A good AI security team is built the other way round. In Centeye, every kind of fix has a dial with three settings:

• Tell me: Centeye finds it, explains it and recommends the fix. You do the rest.
• Ask me first: Centeye prepares the fix and runs it the moment you approve, from a phone if needed.
• Handle it: Centeye fixes it on its own, checks it worked, and tells you after.

You set the dial per client and per action. Protective steps that can be undone, like ending a stolen session or quarantining a file, can run on their own if you want them to. Disruptive steps, like rolling files back from a snapshot or anything that touches a whole client, always wait for a person, whatever the setting. Every fix is checked afterwards at Microsoft or Google, and every step is recorded.

Why one case across everything matters

Real attacks don’t stay in one lane. A phishing email leads to a stolen login. The attacker sets an inbox rule to hide replies, grants an app access to the mailbox, and drops a remote-access tool on a laptop. An AI agent with too much access starts reading files it shouldn’t.

A single-surface tool sees one step of that and treats it as a low-priority blip. An analyst flipping between five consoles might join the dots, if they have time. An AI security team that reads identity, endpoint, email, AI use and AI agents together sees the attack as one story from the start. That is how you catch what a point tool would miss, and it’s also how you avoid waking someone up five times for the same incident. For a deeper look at how attacks move between surfaces, read how attacks chain across surfaces.

A worked example: one weekend, twelve clients

Here’s an illustrative weekend for a small MSP with twelve clients and one on-call tech.

Without an AI security team. Over the weekend the tools raise 40 alerts across the twelve clients. The on-call tech gets pinged through the night, checks the most alarming-looking ones, and leaves the rest for Monday. On Monday morning there are 40 alerts in the queue and no clear picture of which ones matter. One of them, an odd Saturday sign-in at a dental practice, was the start of an account takeover. Nobody joined it to the forwarding rule set ten minutes later.

With an AI security team. Kavach reads all 40 signals as they arrive. It closes 33 as routine, with the reason written down for each. It joins the other seven into three cases. At the dental practice it sees the odd sign-in, the new forwarding rule and the session that doesn’t match the user’s laptop as one account takeover. Because the MSP set “Handle it” for ending sessions and removing inbox rules, Kavach does both, then checks at Microsoft that they worked. Revoking an app the attacker connected waits for one tap from the on-call phone. The tech starts Monday with three cases, not forty alerts, and Argus has already drafted the note to the practice manager.

The time saved is real. Centeye’s own working estimates put ending every session and resetting sign-in at about 45 minutes of hands-on analyst time, and isolating a laptop and gathering the evidence at about 90 minutes. Multiply that across a client base and the hours add up fast.

Checklist: what to look for in an AI security team

Use this when you compare options, ours included:

• It investigates, not just alerts: each case arrives with the evidence, the timeline and a recommended fix.
• It joins signals across identity, endpoint, email and AI into one case per attack.
• You choose how far it acts on its own, per client and per action.
• Disruptive steps always wait for a person, whatever the setting.
• Every fix is checked afterwards and can be undone where possible.
• Every step is recorded: who, what, why and when.
• Corrections make it more careful for that client, never less.
• Cases open as tickets in the PSA you already run, and approvals reach your phone.
• It writes plain-English reports your clients will actually read.
• It is multi-tenant: one view across every client, with policy per client.

Built for the way MSPs work

Centeye is multi-tenant from the start: one view across every client, policy per client, and cases that open as tickets in the standard PSA and RMM tools you already use. Alerts and approvals reach Slack, Microsoft Teams and email. When Kavach gets a call wrong, the tech tells it why, and that lesson can only make it more careful for that client from then on.

Frequently asked questions

Is an AI security team the same as an AI SOC?

They describe the same idea. A security operations centre (SOC) is the team that watches alerts, decides what is real and responds. An AI SOC uses AI agents to do much of that work. Read what an AI SOC is for the longer answer.

Will it replace my techs?

No. It takes on the digging that eats their time, so they spend it on decisions and on clients. The measure of success is how quickly a person gets the facts they need to decide.

Can it take action without asking?

Only where you allow it. You set the dial for each kind of fix. Disruptive steps always need a person.

What does it need to run?

Praxis comes with every Centeye product. Connect Microsoft 365 or Google Workspace and roll out the endpoint agent through your RMM, and it starts working on what those products find.

How is this different from a managed detection and response service?

A managed service rents you a team of people watching your alerts. An AI security team gives your own techs the investigation work already done, inside your tools, on your rules. Some MSPs use both. Read EDR vs MDR for how the two fit.

Clients asking about Copilot agents? Read Copilot agents for MSPs: adopt, secure and sell agent security. Want to see it run across multiple tenants? Start a free trial. New to the terms above? Our security glossary has plain-English definitions. To see where the guard rails come from, read what guardian agents are.

Figures cited from the Coveware Q4 2025 quarterly ransomware report and the Cisco 2025 Cybersecurity Readiness Index. Time estimates are Centeye’s own working figures for hands-on analyst minutes per action. The worked example is illustrative. Centeye does not publish customer metrics it has not earned.