
UpTrajectory Review
ZDNet's piece offers one blunt warning to small-business operators adopting AI agents: think twice about what permissions you grant and what actions those agents can take on your behalf. The available text is a single line, but the concern behind it deserves full attention. AI agents now book appointments, draft emails, process invoices, reconcile accounts, and query customer databases. Each capability requires access to your systems, and each access point creates exposure. An agent holding read-write privileges on your inbox or bank feed can act at machine speed, without pausing to double-check whether the instruction made sense.
Small-business operators face a sharper risk profile than large enterprises do. Big companies employ security teams to scope service accounts, monitor audit logs, and enforce least-privilege policies. You likely have one person handling IT between payroll, hiring, and customer service. When you connect an agent to your accounting software, your CRM, or your email, you hand credentials to a system that interprets instructions probabilistically. It can misread a request, hallucinate a contact, or execute the right action on the wrong record. The blast radius depends entirely on the permissions you granted during setup.
The advice to think twice is sound, but it stops short of a workable framework. We recommend a default-deny posture: grant the minimum access the agent needs for its specific task, choose read-only permissions wherever possible, and split duties across agents rather than giving one agent universal reach. Require human approval for any action that moves money, deletes data, or sends external communications. These controls take minutes to configure and can prevent a catastrophic error that would consume weeks of cleanup and damage client trust.
Permission creep deserves scrutiny. Vendors push broad access because frictionless setup drives adoption, and an agent that can only read your calendar looks less impressive in a demo than one that can reschedule meetings, email attendees, and update project boards. Resist the demo-driven urge to enable everything. Audit what you have granted at least quarterly, and revoke credentials for agents you no longer use. A dormant agent with live access is an open door, and attackers actively probe for abandoned integrations with valid tokens.
Watch for regulatory guidance on AI agent accountability. When an agent makes a mistake, liability rests with the business owner, not the vendor. Document your permission decisions and approval workflows so you can demonstrate reasonable care. If you lack time to manage this properly, start with agents that run in sandbox mode or require human-in-the-loop confirmation for every action. The productivity gains are real, but they only hold value when you control the blast radius.
“Think twice about what permissions you are providing your AI agents and what actions they can take on your behalf.” — ZDNet
Takeaway: Grant AI agents minimum read-only access by default and require human approval before any action that moves money, deletes data, or messages clients.
Excerpt from the original — ZDNet
Think twice about what permissions you are providing your AI agents and what actions they can take on your behalf.