For years, AI helped security teams detect threats faster while humans remained responsible for taking action. Agentic AI changes that model.
Today’s tools can isolate endpoints, disable accounts, block traffic, and launch remediation workflows automatically. The speed benefits are clear, but they also introduce new risks. As AI takes on more responsibility, MSPs need stronger governance to maintain control.
“Agentic AI shifts the SOC from alert processing to multi-step decision making and remediation,” says Eric Hulse, Head of Research at Command Zero. “They can block users, alter configurations, and launch defensive capabilities. They do so at machine speed, which increases response velocity.”
Faster response can stop threats before they spread. However, mistakes can also have immediate consequences.
“Like someone locked out of their account at the worst moment, an important system switched off in the middle of the working day, or a genuine supplier blocked by mistake,” says AJ Thompson, Chief Commercial Officer at Northdoor plc and a member of IBM’s Worldwide Security Advisory Council.
False positives mean more than wasted time
Emily Hartstone, founder of Hartstone Institute, says many SOC teams still rely on assumptions developed when alerts only generated tickets. “Every SOC on earth tuned its detection thresholds around one assumption: a false positive costs an analyst twenty minutes,” she says.
That calculation changes when AI can take direct action. “It does not hold when the output is an isolated endpoint, a disabled account, or a quarantined mailbox. Same false positive rate, entirely different blast radius.”
Hartstone argues that many organizations blur the line between AI that investigates threats and AI that can act on them. “It’s a correctly functioning agent given too much scope by someone who assumed ‘AI-assisted’ meant ‘AI-supervised.'”
MSPs must address the accountability gap
For MSPs, governance becomes more complicated when AI operates inside customer environments. “When your agent acts inside a client’s environment, whose authority is it carrying?” Hartstone asks. “A shared admin credential and a blanket MSA answer none of the key questions around authority, permissions, and verification.” She believes this area could create future liability challenges for MSPs.
Thompson raises a related concern. “If an AI tool decides to switch something off, can you clearly explain afterwards why it made that call, and who’s responsible for the outcome?”
Without clear answers, organizations risk losing trust in the technology and the providers deploying it.
Build guardrails before scaling automation
The experts agree that governance should come before broader adoption. Hulse highlights least-privilege access as a foundational control. “MSPs must deploy least-privilege identities for non-human agents so that they do not accumulate too much privilege.”
Hartstone recommends applying the same access discipline to AI agents that organizations apply to employees. She also advises requiring human approval for destructive or irreversible actions.
Thompson believes people should remain involved in decisions that could affect critical systems or business operations. MSPs should also test failure scenarios and ensure automated actions can be reversed when necessary.
Prioritize visibility and control
As AI takes on more responsibility, auditability becomes increasingly important.
“Agent actions need the same logging rigor as privileged user activity,” Hartstone says, “because when something breaks, the first question is going to be who or what made that call.”
Hulse adds that audit trails, verification controls, and override mechanisms should be standard features of any deployment.
Before selecting a platform, Thompson recommends a simple test: Can the vendor clearly explain what the AI decided and why?
The promise of agentic AI is real, but speed alone is not enough. MSPs that achieve the best results will be the ones that pair automation with clear governance, accountability, and human oversight.
Photo: Evdokimov Maxim / Shutterstock
This post originally appeared on Smarter MSP.