An AI technician that reads the telemetry and a human who approves the fix: how NinjaOne is dividing the work, and what to confirm before you widen access.

Too much time goes to pulling context from several tools before a fix can begin, according to Randy Hyder, MSP Manager at Range MSP, and that gap is what NinjaOne targets with Katana AI. The company announced the agentic AI technician on October 13, built into its Unified IT Operations Platform. Katana AI reads device telemetry, history, and alert activity, then proposes a probable root cause and next steps, including remediation actions and scripts, inside the console where technicians already work.

Diagnosis is the right job to hand to an agent

Katana AI reasons over a customer's own device history, alert patterns, and the resolution knowledge the team has built, which NinjaOne contrasts with a generic model. A recommendation shaped by how your fleet behaves carries more weight than a plausible guess.

Reading telemetry leaves a device untouched, which makes diagnosis the safer place to start.

The technician's approval is the control that counts

Range MSP's Hyder points to two settings: what the Katana agent can access, and approval of its actions. Matt Hastings, SVP of Product Management at NinjaOne, says customers told the company they would trust AI with real access once they could set what it sees and does, and verify it themselves. Hyder puts the result this way: adopting Katana AI "does not mean giving up control of our environment."

I think that division of labor is a sound strategy. The agent saves the technician hours of diagnostic work, and the person who knows the business context decides whether a script runs on a finance laptop or a point-of-sale terminal. Time returns to the technician without handing over the judgment call.

MCP sends outside agents through the same door

The second announcement is the NinjaOne MCP server. Model Context Protocol is a common standard that lets AI agents connect to other software. NinjaOne's server lets external agents read NinjaOne data and take action in the platform under the role-based permissions IT teams already manage.

For a CIO running several AI tools, one permission model to audit is easier to defend than one per tool. NinjaOne describes it as a single orchestration point and a single governance model for every agent that connects.

Cost predictability belongs in the pilot plan

NinjaOne says its architecture is designed for token efficiency so customers can manage and predict costs. Tokens are the units AI providers use to meter and bill the text a model processes, so a diagnostic assistant that runs on every alert can add up. The release gives no figures, which makes per-technician usage a number to request during the trial.

Four details to confirm in the demo

The release sets the direction. These questions fill in the operating detail:

• Do approval prompts apply to actions requested by external agents through MCP, or to Katana AI alone?
• Can you grant read access and action rights as two distinct permissions for each agent?
• What does the audit trail record when an agent proposes a fix and a technician approves it?
• How is token usage metered and reported to your team?

NinjaOne reports six major releases and 43 minor releases in the last year, so the answers may change within a quarter. Ask for the current product roadmap alongside the demo.