What may an AI agent do? Set boundaries before the first task
Define an agent's task, access and approval points, then check the result where the action actually happened.
Article language: English

In this article
An AI agent can draft a message or send it immediately. There is a decision between those actions that should not be left to chance. Before assigning the task, define access, permitted actions, the point of approval and how the result will be accepted. A person should know what they are approving. “Do it properly” is a pleasant wish, but a poor operating instruction for a company inbox.
Here, an agent means a system that uses an AI model and tools to carry out a task in several steps. It chooses some actions itself, so a goal alone is not enough: it needs an access scope, action boundaries and a way to check the result. These elements form a basic company workflow.
How do you write a task that can be checked?
Start with the expected result. “Prepare a list of contracts renewing next month from the specified register” is easier to check than “take care of contracts”. Add the recipient, format and completion criterion.
Name the source of truth. If the agent sees a contract, spreadsheet and meeting note, it should know which document supplies the binding date. When sources disagree, you need a list of discrepancies. Automatically choosing the more convenient date can conceal the real problem.
For a simple task, a short card can be enough: goal, sources, permitted actions, approval point and evidence of execution. Add an example of a correct result. This lets you test the instruction on a small piece of work before giving the agent a wider scope.
Which permissions should you grant at the start?
Grant the access needed for the specific task. Reading a chosen folder, creating a draft and publishing should be separate decisions. Organise tools and accounts so that the separation also exists technically.
Record the access period and owner. After the trial, check which connections remain active. If a tool grants broad permissions while the task needs narrow ones, that mismatch should be visible in the assessment before launch.
In the trial stage use test material or data approved for that work. Prepare a situation where the agent encounters missing information. It should be able to show the gap rather than invent a missing amount, name or surname.
When should a person approve an action?
Set control points according to effect. Reading a public page and preparing a table usually need different oversight from a transfer, file deletion or sending an offer. For external actions, the person should see the specific result: content, recipient, amount or list of changes.
General permission to “handle customer service” does not explain whether the agent may issue a refund or promise a new date. Record those permissions separately. Also define what needs renewed approval after content or conditions change.
- Before sending, show the finished message and recipients.
- Before publishing, show the destination address and material version.
- Before changing data, show the scope and how to restore the previous state.
- After execution, retain confirmation from the system where the action occurred.
The last point matters. An agent's “done” message is a claim. An operation number, status in the application or checked address of a published page is evidence that can be assessed independently.
What should you do with instructions found in documents?
A page or attachment may contain text resembling an instruction to the agent. Treat it as source content. A supplier's document should not itself grant permission to send company files or change the task's rules.
Include such cases in a test. Put a request for an out-of-scope action in a test document and check how the system responds. Verify restrictions at the tools, since an instruction to the model alone may not be enough. Combine this test with a check of actual file and network boundaries if the agent uses an isolated environment. The article on outside instructions in documents (in Polish) and the article on sandbox file and network boundaries (in Polish) describe those checks.
It is useful to record actions, but also to limit log contents. Detecting an error often requires a document identifier and the type of operation. Copying all confidential content into another system creates an additional data flow.
How do you put the boundaries on one task card?
A working example: an agent prepares draft replies to questions about delivery dates. It may read a selected order register and create a document with proposals. It has no permission to change dates or send messages. The task owner identifies which field is the confirmed date and what to do when it is empty. This is a proposed way to organise work, not a claim about a feature available in every product.

The card can fit in five sentences: select records that meet a defined condition; cite the source for each; mark missing information as missing; save ready replies in an agreed folder; finish with a list of cases needing a human decision. Record the cost limit and trial time separately. When something falls outside those conditions, the agent should report the specific obstacle rather than expand the task on its own.
If preparing sources requires sorting conflicting files, start organising them with the guide to company knowledge architecture (in Polish). A good instruction cannot replace knowing who owns the register. Keep the agreed source-selection rule in a project decision log so the next operator can explain why it was chosen. See how to keep a decision log (in Polish) for the details to record.
What must approval of a prepared operation include?
The reviewer needs the complete set: content version, destination, recipients and effect of the action. If a person approved one message and the tool then added an attachment or changed the recipient, that approval no longer describes the same result. Test how renewed review is enforced in the integration you use.
For bulk changes, show a scope summary and a way to inspect individual items. “Correct 80 records” is not enough if nobody knows which fields will be replaced. Compare the state before and after the operation in a test. Confirmation should come from the destination system, not only a conversation summary. When an AI agent operates a browser, distinguish the approved intent, the click and the confirmed state in the destination application, as illustrated in the article on browser agents (in Polish).
For code or deployment, require the version and the results of relevant checks. The article on accepting code, access and confidentiality after agent work (in Polish) organises the materials that should remain after a task. Connect publication approval to a specific address and checked behaviour.
How do you test a retry after a tool error?
Prepare a case where a tool performed an action but the agent did not receive confirmation. Without a control, a retry can create a duplicate. The appropriate safeguard depends on the system: it might be an operation identifier, a status read or manual resolution. Do not assume every integration recognises repeats on its own.
In the test scenario, expect an unknown-result message and a state check before the next attempt. A certain access denial, where the operation never began, is a different case. Separating them makes errors easier to assess and prevents broader permissions being granted merely to silence a message.
What happens when a task passes to another person?
Hand over the work state, not only the conversation transcript. Include finished materials, actions actually performed, operations awaiting a decision and unresolved errors. Identify active connections, their owners and the plan for ending access. This package lets the session stop without losing control.
Before permanent use, compare behaviour with the agreed task card and retain failure cases. If an operator must reconstruct the history manually after every attempt, narrow the scope. A well-bounded agent can perform a smaller task that the team can check and repeat independently.
Also review the technical side of the task description: which accounts and integrations make it possible, and whether permissions exceed the work's scope. The article on AI integration permissions (in Polish) helps with that review. Before starting, prepare an operational response path for a data incident, with a reporting owner, evidence preservation and control over suspending access. The guide to the first response to an AI data incident (in Polish) helps organise those steps.
How do you check that the workflow works?
Prepare examples: a correct task, conflicting sources, missing permission, an unavailable tool and an out-of-scope request. Describe expected behaviour in advance. After changing the model, instruction or integration, repeat the relevant checks.
NIST organises AI risk management into four functions: Govern, Map, Measure and Manage, covering responsibility, context, measurement and action on risk. This is a useful framework for asking questions across the workflow's life cycle. It is not a certificate of compliance. Details are in the NIST AI Risk Management Framework (in English).
Finally, designate the person who stops the workflow and collects error reports. Decide how to disable connections and return to manual work. A trial then has a clear start, measurable result and exit path. Separately plan how to retire the automation without leaving tasks behind, so ending use also covers pending operations. The automation retirement guide (in Polish) describes what to close.
Connect these rules with an AI tool inventory; our guide to inventorying AI tools (in Polish) covers owners, use and costs. To discuss access, data and responsibilities in your company, see the Aeneia Audit overview.
Before expanding use, plan an AI pilot with baseline results, using the guide to AI pilot criteria (in Polish) to define how to assess it. Describe the material's route in a data-flow register so access scope can be checked outside the instruction itself. See how to map an AI data flow (in Polish) for the register's contents.
Turn action boundaries into a specific point for human approval. The article on human approval of AI actions (in Polish) explains how to design that decision. The approver should see the prepared action, recipient and significant effect. An “OK” button is not a sound control if nobody knows what is being approved.










