AI Agents for MSPs: Decide What the Agent May Do First

The primary structural trend is the shift from reactive auditing of AI agent actions to proactive, policy-driven control over what AI agents are allowed to do within IT environments. This is illustrated by companies like Charles Schwab and JPMorgan Chase, which have established specific limits on agent actions rather than broad bans, and by technology providers such as Anthropic and Microsoft introducing tools to enforce granular permissions before AI agents are deployed.

Evidence underscores this shift: Acronis found AI agent tools on 9.4% of monitored Windows machines, with higher adoption among MSP-managed clients. Financial institutions are restricting critical actions—Charles Schwab only permits limited AI agent access through controlled servers and explicitly blocks agents from moving money, while JPMorgan Chase prohibits external AI agents altogether. OpenAI’s disclosures of agents circumventing safeguards led to increased automated oversight. Microsoft now contains agent permissions dynamically, citing the inherent risk in delegating broad privileges to AI.

Supporting developments include reports that less than 3% of legitimate agent activity can be reliably identified, highlighting the difficulty of external oversight. Tools like ThreatLocker’s Zero Trust controls and 1Password’s credential policies show vendors are responding with more rigorous permission and approval frameworks, but these solutions still require explicit configuration and maintenance by service providers.

For MSPs and IT service providers, this trend increases both governance obligations and operational workload. Action policies must be created, updated, and monitored for every client and internally for provider systems. The work of defining and maintaining these policies affects margins unless priced explicitly. Approval fatigue is another risk, as excessive prompts can reduce the effectiveness of human oversight. Providers should anticipate that clients will expect AI usage policies to be a formal part of agreements, requiring new approaches to contractual and operational management.

00:00 Banks and Sites Push Back on AI Agents

03:50 Why Watching Agents Afterward Fails

06:43 The Permission Layer Needs a Policy

09:41 Why Do We Care?

Supported by:

Pax8
HaloPSA

NinjaOne On-Demand Webinar: https://go.businessof.tech/p/ninjaone-pod

About Business of Tech: The daily MSP news and analysis podcast for managed service providers and IT services leaders. Host Dave Sobel covers AI for MSPs, MSP pricing and profitability, MSP M&A and private equity, vendor partner programs, MSP cybersecurity and liability, and MSP operations — and answers the question every story raises: Why do we care? Produced by MSP Radio.