AI agents now have access to your business systems: why access control is the weak link

More and more organisations are connecting an AI agent to a system that actually matters. No longer just a chatbot answering questions, but a workflow in n8n or Zapier that independently processes leads, creates invoices, or sends emails on the company's behalf. It often happens in an afternoon, built by an enthusiastic employee without IT ever finding out.

That pace is exactly where things go wrong. IBM's Cost of a Data Breach Report 2026 found that in 92 percent of breaches where AI played a role, the AI system involved had no proper access control. No scoped permissions, no separate identity, no log of what actually happened. Incidents involving shadow AI, agents and tools set up outside IT, nearly doubled: from 20 to 43 percent of all cases studied.

What is missing when an agent simply gets access

In most cases, an agent does not use its own identity, but the credentials of the employee who built it. An API key that got full access 'just for now' during the first test often stays exactly as it was created. The result: the logs show the employee's name, not the agent's. Afterwards, there is no way to tell who carried out which action, the person or the script running on their behalf.

That difference with a regular SaaS tool is significant. A tool you operate yourself only does something when you click a button. An agent carries out the same action dozens or hundreds of times a day, without anyone reviewing every step. A small mistake in a trigger or an overly broad filter scales instantly as a result.

A realistic scenario

Say a marketing team builds an automation that moves new leads from a form into the CRM and sends a welcome email. For convenience, the integration gets write access to the entire CRM, instead of just the lead field it was built for. Months later, someone adjusts a filter, and within minutes the agent overwrites thousands of existing customer records.

This is exactly the pattern that keeps coming up in governance research. An agent gets more rights than its task requires, and nobody notices until something goes wrong. No one was specifically watching what that one script was actually allowed to do.

Why automation projects still fail

Analyst firm Gartner predicts that more than 40 percent of agentic AI projects will be cancelled before the end of 2027. Not because the technology does not work, but because of rising costs, unclear value, and too little grip on risk. Access control is not a side issue in that. A project with clear boundaries from the start, who can do what, which actions require approval beforehand, survives the pilot phase more often than one that has to explain afterwards what exactly went wrong.

Five steps to get this in order now

Access control for AI agents does not need to be a large project. Five steps already provide a lot of grip in practice.

Build an overview of which n8n flows, Zapier zaps, scripts and AI agents run within your organisation, and which systems they touch. Most organisations have no complete picture of this, simply because agents get set up by separate teams.

Give every agent its own account. Never let an automation run on an employee's personal login. A separate service account with exactly the rights it needs immediately makes clear who, or what, carried out an action.

Build in an approval step for actions that cannot be undone. Large send actions, deletions and payments deserve a human look before they go through, even if that delays the automation by a few minutes.

Log what an agent does, separately from human actions. Without a log, there is no way to reconstruct afterwards what happened, and that is exactly what IBM's figures point to.

Periodically review which agents are still needed. An integration built for a one-off project often stays active long after everyone has stopped using it. Revoke access as soon as that is the case.

Automation with AI agents delivers real time savings, provided the access behind it is set up as carefully as the automation itself. Want help finding out which agents and integrations are active within your organisation and how to set them up securely? Zarioh helps with inventorying, scoping and maintaining AI agents and SaaS integrations. Feel free to get in touch for a no-obligation conversation.