Skip to content
Interface language: English
AI course / Signal

AI automation: what a person must approve before an action

Design an approval boundary, restrict access and test rejection and retries in a fictional complaint workflow.

Article language: English

Aeneia concept image: a woman guides a thin tool between red, navy and golden threads on a small loom.
In this article

An automation has drafted a complaint response. Now it is meant to click “send” or refund money. The responsibility changes between those steps: a working draft becomes an action affecting a customer. Put human approval at that boundary.

Follow a fictional shop process. Case R-17 concerns a damaged mug handle. The system may read the description and prepare a draft. It may not independently grant a refund or replacement or change the order status. The exercise uses invented data only; we do not send messages or initiate payments.

Using fictional case R-17, prepare a package for the handler to review.
List the problem, missing information, proposed response, proposed action
and the facts supporting it. If the material is insufficient, say “requires clarification”.
Do not perform actions or tell the customer a decision has been made.
Case: “The mug handle was cracked on unpacking. A photo is attached.
Please tell me what to do next.”

A correct reference answer summarises the problem and proposes a next step. It does not confirm a refund. The prompt prepares material; access restrictions must also be configured in the automation system.

Where to place approval

Automation is an established flow of actions between systems. An AI agent may additionally select tools and their parameters. If it can send messages, approval should cover the specific message, recipient and attachments immediately before sending.

n8n describes human-in-the-loop for tools: the flow pauses and a person approves or rejects execution with the proposed parameters. This is an example of a product feature, not proof that every n8n process has it correctly configured.

In the scenario, the handler sees the case number, problem, response and recipient address from a trusted system. They do not approve a general instruction to “handle the case”. The button must answer what will happen when clicked.

If the text or recipient changes after approval, obtain a new approval. Consent to version one does not automatically carry over to version three, even if both begin with a polite greeting.

Design the approval screen around concrete data. For fictional R-17 use recipient klient-test@example.invalid, no attachments, and the text: “We received your report. The handler will check whether the materials are complete.” This is a demonstration address; the exercise does not send a message.

Prepare an approval card for fictional action WYSLIJ_ODPOWIEDZ_R17.
Recipient: klient-test@example.invalid. Attachments: none.
Text: “We received your report. The handler will check whether the materials are complete.”
Show the case number, recipient, complete text and exact effect of approval.
List changes that must invalidate approval. Also prepare a rejection message.
Do not send a message.

The reference card shows all parameters before the button. A change to text, recipient or attachments requires fresh approval. Rejection leaves the case with no response sent. Check that the card does not hide part of the message under “standard response”. The person should approve what will actually reach the recipient.

An email prompt (in Polish) helps prepare the text. Record content correctness separately from permission to send. Even a correct response may still await the right time or decision-maker.

Restrict permissions to the task

Give the process access to the necessary case, not the whole mailbox and every refund. Separate reading, drafting, sending and financial operations. At the first stage, reading and saving material for review suffice.

OWASP identifies excessive agency as a risk category, pointing to excessive functions, permissions or autonomy as sources of harmful actions. Restrictions must be enforced in tools and accounts; a note in a prompt does not remove system privileges.

An external document or message can contain an instruction trying to change a model's behaviour. OWASP calls this prompt injection. A request in a complaint to disclose the entire customer database is sender-provided data, not authority to perform the operation.

Set the information scope beforehand using safe data practices for AI (in Polish). A person at the end cannot undo a disclosure that already occurred at the beginning.

Test rejection and retries

A working “approve” button is half the test. Check rejection: no message should leave. Then change the recipient after approval and check whether the old approval can still send the material. If it can, stop the rollout.

Also run the same case twice. The system should recognise that the approved action has already occurred. This property is called idempotency: a repeated request does not create a second effect. Checking it requires a record of the action identifier and its actual outcome.

Define what happens if the handler does not respond and if the connection fails. Silence must not mean approval. “Sending started” is not proof of delivery. Record approval, the execution attempt and the confirmed result separately.

Trace a fictional log. Action A-17 was approved. Sending began, but the connection broke before the result was received. A minute later the process ran again. The log contains neither confirmation of sending nor a confirmed service rejection.

Analyse the fictional action log A-17 above. Separate confirmed events
from the unknown sending result. Do not treat a broken connection as proof
that nothing was sent. Suggest what to check in the sending system before
retrying and identify the risk of a second message. Prepare a note for the
handler without claiming that the email arrived.

A correct analysis leaves the result unknown until checked. That differs from both “sent” and “not sent”. In a real process, recovery depends on the particular service; the designer must establish it before use. A prompt can describe the problem but cannot read an unconnected log.

If you analyse event logs in a table, apply the spreadsheet data check (in Polish). Two rows with A-17 may represent a retry of one action, not two separate cases. Establish what a row means before counting completed operations and declaring the process effective.

Accept one complete run

Test on a test account with a fictional case. Inspect input, proposal, human decision, execution and the recorded outcome. Separately exercise rejection and retry. Only that evidence lets you assess whether the boundary works beyond a diagram.

Add a scenario in which the handler does not respond. Determine who will see the pending case and when it needs manual takeover. Do not leave the model to choose the authorised decision-maker: the process must define that role. Record the reason for stopping. During acceptance, ask someone outside the implementation team to find R-17 and explain whether the message went out. If they cannot determine it, improve state visibility before starting the automation.

If the automation comes from team decisions, first check the meeting note and decisions (in Polish). Choose a platform against the task and tool acceptance criteria (in Polish). In the programme (in Polish) of the AI Course (in Polish), developed by Aeneia, automation follows a previously described process. Mark the last point on your diagram where the action can be stopped before contacting the customer.

*The source states that documentation was checked on 30 October 2026 and that the OWASP entries concern the 2025 edition. The complaints process is an invented scenario, not a deployed system.*

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Explore the AI course (in Polish)

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear