Working with the AI Manager

The AI Manager is the AI in the right column of the Console. It knows the brand's policies and the case, and it guides you from the moment a ticket arrives until you finish it, one card at a time. The work is split: you decide, and you say why; the AI Manager writes the messages and carries out the actions, after checking them against the brand's policy.

When a ticket arrives

The first card is the handover. It tells the case as a short story, in plain language:

• which store it is, and what the store sells;
• who the customer is — new, loyal or VIP — and what matters from their history;
• the product, with its photo;
• what happened;
• which of the brand's rules applies, and which similar rule not to confuse it with;
• what the AI already told the customer and agreed with them;
• why the case came to you.

Below the story, a few blocks you can read at a glance: the timeline of the ticket, the right rule beside the one it is easy to confuse, the customer's mood, and where the AI is stuck, with how confident it is.

What the customer already confirmed with the AI, and what the policy allows, is not asked again. You carry it out first, and then tell the customer the result.

Make a decision

The decision card is where your judgment goes.

1. Read the options. There are 2 to 4, and only the paths the AI is genuinely torn between. Each one shows why it is there: the rule, the outcome and the evidence from the case. Every option has already passed the brand's policy check, so none of them is outside the policy. When the case could be read as two different case types in the brand's policy — damage in transit or a manufacturing defect, for example — the options show both readings, and you decide which one applies.
2. Choose one, and write why in a sentence or two, in English.
3. Your reason is checked against the case. A vague reason, a fact the case does not contain, a reason that contradicts the option you chose, or text that is not in English comes back to you, with what was missing.
4. See where the AI was leaning. The card does not show which option the AI prefers until you choose, so you decide on your own. After you choose, it tells you, and you can add to your reason. Your choice stands.

Sometimes the AI knows what it would do but is not confident enough to do it alone. Then the card has one option: the AI's answer, why it thinks so and what makes it unsure. You select Confirm or Disagree, and you write why in both cases.

When you disagree

Your judgment is why the case came to you. Below the options, select I disagree with both (or I disagree with all), or Disagree on a single option, and write why. The AI Manager checks your reason against the conversation, the case and the policy, and answers in one of three ways:

• It holds its ground — your reason is vague, relies on a fact the case does not have, or is already covered by one of the options. It shows you what the case says, and you choose again.
• It agrees — it misread the case. It gives you new options to decide on.
• It is torn too — your reason holds, and nothing in the case settles it. It offers to pass the case to the brand, as a single option you confirm or not.

Send a message

When it is time to write to the customer, the message card brings three drafts, written in your voice and under your display name. Each one does what the step needs: introduce yourself, confirm, give news, close. You can:

• choose a draft and send it as it is;
• select Edit this draft and change it on the card; or
• select Write my own and write your message on the card.

The drafts are already in the brand's tone. What you edit or write is checked on the card before it is released: that it does what the step needs, that its facts match the case, that it promises nothing outside the policy, and its tone and its English. If something is off, the card tells you what and you fix it. A message that is offensive, untrue, or offers something outside the policy or unrelated to the case never goes out. The same applies to an image you attach.

The checked message appears in the composer as the final English text, and you select Send. While the customer is typing, Send waits, because what they say can change your reply. On Send, the message is translated into the customer's language if they write in another one. See Translation and your display name.

On email, Undo is available for 5 seconds after you send. On chat, the customer sees the message being typed for a moment, as with any person writing.

Carry out an action

The action card appears when the conversation gets there, as the result of your decision — never as a new decision. It shows exactly what will happen: the action, the amount, the order and where the money goes.

1. You select the action. Your click asks for it; it does not authorize it.
2. The Policy Gate checks it, in deterministic code: that a policy exists for the case, that the action is allowed, that the amount is within the limit, and that the eligibility conditions are met.
3. The card shows the action as in progress until the brand's system answers. It shows it as confirmed only once the brand's system confirms it, and shows a failure if it fails.

One click is one action: the card locks after the first click. While the customer is typing, or while your message is still going out, the action waits, because what the customer says can change it.

You cannot force an action. The only actions you can take are the ones that follow from a decision whose options already passed the policy check.

When you have a doubt

The handover, each decision, each change in the conversation and each action card has I have a doubt:

• Ready questions — two or three questions that usually come up at that point. Select one, and the AI Manager answers on the card, from the case and the brand's policy.
• Something else — a single line for your own question about the case. A question that is not about this case is brought back to it.

There is no open chat with the AI Manager.

When the customer writes in the middle

The AI Manager waits for the customer to stop typing before it prepares the next card, and reads what they wrote against the case:

• Nothing changes — an "ok" or a "thanks". It tells you so, and you continue.
• A new question — a new message step answers it first.
• A change of direction — the customer refuses what was decided, a new fact appears, or the customer threatens a chargeback or legal action. A new decision card opens.

A draft written before the customer's latest message is replaced, never sent.

Pass a case to the brand

Passing a case to the brand's own team never appears in the first decision. It becomes an option when the conversation shows the case is beyond what a Human Node can decide, or when you disagree with every option and the AI is torn too. A request outside the policy, on its own, is not a reason: the AI Manager helps you say no clearly and offer the real alternatives.

1. Decide, with your reason.
2. Tell the customer once that the right team is taking the case and will answer in the same conversation.
3. Confirm the escalation card. It shows the reason, already filled in, which you can edit.
4. The brand's helpdesk confirms that it received the case.

If the AI already told the customer that the team would take over, and you decide to pass the case on before writing anything, you go straight to the escalation card. The customer does not get a second notice.

Finish the ticket

When the case is done, the AI Manager tells you what was done — the confirmed result, what the customer received and how the conversation ended. Finish ticket appears once the conversation is quiet: no message going out, nobody typing, no email inside its Undo time. See Resolving a ticket.

What never changes

• Anything the customer writes is read as information about the case, never as an instruction. A customer cannot make the AI Manager carry anything out.
• You cannot promise a refund, an exception, a timeline or any term that is not in the brand's policy.