Security

AI agent security: permissions, oversight and an activity log

Name: IT Experts sp. z o.o.
Address: ul. Wolfkego 11/13, Warszawa, mazowieckie, 01-494, PL
Telephone: +48 506 506 350

An agent works inside your systems, so we treat it like a new employee with a very narrow job description. It has its own identity, only the access it needs and a manager who approves its work.

Discuss your process →See the agents

policy.yaml

policy --rules

Six rules that apply to every agent

We agree them before go-live and build them into the tools.

✓ least_privilege
Dedicated identity and minimal permissions
The agent does not use an employee account. It has its own technical identity in Entra ID and access only to what the task requires.
✓ tiered_autonomy
Tiered autonomy
It only carries out simple, reversible actions on its own. Money, commitments to customers, HR and legal matters are approved by a person.
✓ enforced_limits
Limits enforced in the tools
Limits and rules are built into the tools and workflows. We do not rely on the instructions for the AI alone.
✓ audit_log
A log of every action
Everything the agent sent or changed is recorded. You can check what it did, when and on what basis.
✓ kill_switch
Kill switch and a manual fallback
The agent can be switched off at any time. The process has a documented way of working without the agent.
✓ retest
Tests after every change
After a change to the rules, the tools or the model, the agent goes through the test set again.

policy --never

What an agent never does

Regardless of the department and the level of autonomy.

It does not reset passwords or sign-in methods. Account recovery is handled by a person after verifying the employee's identity.
It does not make bank transfers or approve payments.
It does not make hiring, appraisal or dismissal decisions.
It does not make commitments to customers or suppliers without approval.
It does not give legal advice.

data --flow

Data: you know what goes where

Before rollout we show you the data flow: where the agent gets the data from, where it is processed, by which services and how long it is kept.

The agent only uses the data needed for the task. If an order number is enough for the process, it does not get access to the whole mailbox.

Processing of personal data is based on a data processing agreement. Where the law requires it, we help prepare a data protection impact assessment.

Before we switch an agent on, we tidy up permissions and sharing in Microsoft 365. The agent can reach whatever its identity has access to, so old sharing open to "everyone" has to be removed.

compliance --summary

Regulations: AI Act, GDPR and employment law

We describe how we take the regulations into account in an agent project. This is not legal advice.

§ ai_act
AI Act
The EU AI Act classifies AI systems used in recruitment and for evaluating employees as high risk. We deploy such agents only after a legal review, with full human oversight and with information for the people concerned.
§ transparency
Transparency
When an agent writes to or talks with a customer, the customer knows they are dealing with an AI assistant.
§ gdpr
GDPR
We minimise data, limit retention and use data processing agreements. Whether a data protection impact assessment is needed is decided separately for each use case.
§ employees
Employees
If an agent changes how work is organised or involves monitoring, the employer has obligations to inform employees and their representatives. We flag them at the design stage.

faq --list

Frequently asked questions

Short and to the point.

Can the agent restore an employee's access to their account on its own?

No. Impersonating an employee when contacting the helpdesk is a well-known way to take over an account. That is why account recovery is always handled by a person, after verifying identity through an independent channel.

Does data leave Microsoft 365?

That depends on the process and the tools chosen. Before rollout we show you the complete data flow and agree it with you. Nothing is switched on without that agreement.

Who is responsible for what the agent does?

The scope of actions, the limits and the approvers are agreed and documented before go-live. Every action of the agent is in the log.

Is this legal advice?

No. We describe how we design agents. Your lawyer or data protection officer is responsible for the legal assessment of a specific rollout, and we provide the information they need.