AI Agent Access Management

By Juliya Turchyna, Product Owner at IBA Group

Consider an AI agent working in a finance department. It opens incoming invoices, extracts information from attached documents, checks supplier details against the ERP system, creates accounting entries, and prepares payments for approval. When something does not match, it opens a ticket for an accountant.

The workflow looks efficient. A task that previously required an employee to move between several applications can now be handled largely by software. Yet the agent has also acquired something that enterprise technology teams have spent decades carefully managing: access to business-critical systems and the authority to perform actions within them.

In this example, the agent interacts with email, document storage, ERP, payment systems, and the service desk. Its actions can affect financial records, operational processes, and potentially other people.

An employee performing the same work would have an identity, assigned responsibilities, access restrictions, and a manager accountable for overseeing the process. Companies need equivalent controls for AI agents, while accounting for the differences in how these systems behave.

As businesses move from AI assistants that generate information to agents that execute tasks, the ability to delegate work is developing faster than some organisations’ ability to govern that delegation.

When an AI mistake becomes a business event

An incorrect answer from an AI assistant can influence a business decision. Employees may trust it too readily, overlook an error, or act on information that appears convincing.

Autonomous execution introduces a more immediate consequence. An error can become a modified customer record, an incorrectly processed payment, or a message sent to the wrong recipient before anyone reviews what happened.

An invoice-processing agent, for example, might encounter malicious instructions hidden in a supplier document. If it treats that content as an instruction rather than untrusted data, it could attempt to change payment details or disclose information through one of its authorised integrations.

Other failures may be less dramatic but equally consequential. An agent can complete thousands of operations in a short period, repeating an unnoticed mistake across records. Individual actions may appear permissible while their combined effect violates the intended business process. A model or prompt update can also change the agent’s behaviour without any corresponding change in its access rights.

These risks become particularly difficult to investigate when an organisation cannot distinguish between actions performed by an employee, actions performed independently by an agent, and actions approved by a person following an agent’s recommendation.

The security question therefore extends beyond whether the model produces accurate outputs. Companies must understand what the agent is capable of doing when its reasoning fails, its instructions are manipulated, or the circumstances differ from those anticipated during deployment.

An employee’s credentials are not an agent’s identity

Corporate identity and access management has traditionally been organised around employees and applications. Employees receive permissions according to their roles. Applications and service accounts receive permissions to perform defined technical functions.

AI agents introduce a different operating pattern. They can select actions dynamically, work across applications, and perform tasks on behalf of individual employees.

Existing identity systems provide important foundations, but authentication alone cannot determine whether an agent’s particular action is appropriate within the context of a business task.

This becomes especially relevant in legacy environments where applications recognise only a user account or service account. An agent may be connected using an employee’s existing credentials because the system has no native way to represent delegated authority.

That shortcut creates several problems. The agent inherits permissions beyond those required for its task. Its activity becomes difficult to distinguish from the employee’s. If something goes wrong, disabling the agent may also mean disabling the person’s access.

A dedicated agent identity creates the basis for more precise control. The organisation can establish on whose behalf the agent acts, which operations it may perform, and how long that authority remains valid.

For systems that cannot enforce these restrictions directly, a policy gateway can check requests before they reach the target application. Short-lived delegated credentials and action-level authorisation allow the organisation to restrict the agent to the work it was actually assigned.

An agent processing supplier invoices may need to read purchase orders and prepare payment records. It does not necessarily need permission to change supplier bank accounts or release unrestricted payments.

The relevant unit of authorisation becomes the business task and the actions required to complete it.

Accountability cannot be delegated to software

Access management becomes an organisational challenge when agents cross departmental boundaries.

A finance team understands how invoices should be processed and which exceptions require approval. IT security understands authentication, access controls, and monitoring. Technical teams understand integrations and operational dependencies. Legal, compliance, and risk functions understand the obligations that apply to the process.

The business owner should define the agent’s purpose, acceptable outcomes, and permissible actions. Security should translate those requirements into enforceable permissions. Technical owners should ensure that the integrations, monitoring, and emergency controls work as intended. Compliance should establish applicable constraints and evidence requirements.

The business process still needs an accountable owner who understands what the agent has been authorised to accomplish.

This is particularly important when companies introduce agents through individual departments. A team may see a relatively simple automation opportunity, while the resulting workflow requires access to systems and information owned elsewhere in the organisation.

The approval process should examine the complete sequence of actions, including what happens when the agent encounters unexpected information, needs additional permissions, or attempts an operation outside its original scope.

Human oversight must be operationally meaningful

Companies frequently describe human oversight as a safeguard for autonomous AI. The practical value of that safeguard depends on what the person can actually see and control.

A reviewer who receives an approval request without sufficient context may simply confirm the agent’s recommendation. A supervisor who discovers an action only after execution has little opportunity to prevent it. An organisation that cannot stop an agent independently may have oversight procedures without effective intervention capabilities.

The appropriate degree of autonomy should reflect the consequences of the action.

Reading information, classifying documents, preparing drafts, and creating internal tickets can often operate with relatively broad autonomy, subject to logging and periodic review. Reversible changes or low-value transactions may be appropriate within defined limits, with exceptions escalated for approval.

Changes to bank details, significant payments, deletion of important records, access-right modifications, and decisions with serious consequences for individuals require much stronger controls. Some use cases are also subject to specific legal obligations concerning human oversight or automated decision-making, including relevant provisions of the EU AI Act and GDPR.

A meaningful approval mechanism should show the reviewer what the agent intends to do, why the action is proposed, and which information supports it. The reviewer must have enough time and authority to intervene before execution.

If an agent’s behaviour becomes unexpected, the organisation should be able to suspend execution, revoke delegated access, and investigate the actions already performed. These capabilities need to be designed and tested before deployment.

A practical example: establishing the foundation for autonomy

The importance of these controls became clear in a project involving a managed IT services provider operating within the scope of NIS2.

The company wanted to introduce an AI assistant to help third-level support engineers resolve incidents involving legacy platforms. The assistant needed to search code repositories, operational logs, and historical tickets across several client environments.

The business case was straightforward: engineers were spending substantial time locating technical information and reconstructing previous incidents.

The access requirements were considerably more complicated.

Operational data contained shared service accounts, secrets, and personal information. The organisation also had to protect information belonging to different clients. Without appropriate isolation, access controls, and auditability, the security, compliance, and client stakeholders could not approve deployment.

The solution was deliberately designed with limited autonomy.

The assistant received its own corporate identity and accessed information on behalf of an engineer, within that engineer’s permissions. Its access remained read-only, and an architectural restriction prevented it from executing commands or modifying production systems.

Client data was logically isolated, sensitive information was masked before reaching the model, and incoming logs and tickets were treated as untrusted content. Each request generated a structured record identifying the relevant users, information sources, policy decisions, and model configuration.

The deployment went live after the required approvals. According to the project’s internal measurements, L3 incident resolution time decreased by 25%, time spent searching for technical knowledge fell by 80%, and onboarding time for new engineers improved by 30%.

These figures describe the results of this particular deployment, rather than a general benchmark for enterprise AI.

The distinction is important: this was an advisory assistant, not an autonomous operational agent. It could analyse information and support engineers’ decisions, but it had no authority to execute changes.

If the organisation later wanted the same system to reset credentials, modify infrastructure configurations, or execute remediation commands, the existing controls would provide a starting point, but they would not be sufficient on their own.

Each new action would require explicit authorisation tied to the incident and the engineer initiating it. Permissions would need to be restricted to particular systems and operations, with limits on how long they remained valid. Changes affecting production availability, access rights, or client data would require appropriate human approval.

The architecture would also need to enforce those decisions before execution, record the resulting changes, and support emergency shutdown, access revocation, and rollback where technically possible.

This is how an organisation can approach greater autonomy without treating it as a single leap from assistance to independent execution. The controls established for a read-only assistant become the foundation for deciding which additional responsibilities the system may safely assume.

Auditability requires more than an activity log

Once agents begin operating across enterprise systems, the ability to reconstruct their decisions becomes a management requirement.

Traditional application logs may show that an API was called or a record was updated. They may not establish who initiated the task, which agent performed the action, whose authority it used, or whether a person approved the operation.

An auditable record should connect the agent’s identity with the original request, the relevant business task, the systems accessed, the operations performed, and the outcome. Where approval was required, the record should identify who authorised the action and what information was available to that person.

Model versions, prompt configurations, and policy decisions also matter because an agent’s behaviour can change while its technical permissions remain the same.

Monitoring serves a related but different purpose. Unusual activity volumes, attempts to access unexpected data, or repeated policy violations can indicate that an agent is behaving outside its intended scope.

Together, traceability and monitoring allow an organisation to understand past actions and identify emerging problems while intervention is still possible.

Regulated organisations may face additional requirements concerning logging, retention, incident reporting, and human oversight. Those obligations depend on the particular use case, sector, and applicable legislation.

The next governance challenge is delegated authority

AI agents are likely to become increasingly capable of coordinating work across business applications. Their usefulness will depend partly on how much operational authority organisations are prepared to grant them.

Permissions that are too broad increase exposure to errors and misuse. Restrictions that are poorly designed can make the agent ineffective or force employees to work around the controls.

A more sustainable approach starts with a clearly defined business process. The organisation determines what the agent needs to accomplish, which actions are necessary, what level of autonomy is acceptable, and where human decisions remain essential.

This also gives organisations a way to expand automation responsibly. An agent may begin by searching and preparing information, then progress to reversible updates or transactions within defined limits as the company develops evidence that its controls work.

The future of enterprise AI will depend as much on organisations’ ability to govern increasing autonomy as on the capabilities of the agents themselves.

LEAVE A REPLY

Please enter your comment!
Please enter your name here