ReplyXDocs
Guides

AI Assist

How to connect your own OpenRouter account, teach the assistant your policies, let it answer or draft replies, take a conversation over from it, and route and size tickets with AI.

AI Assist adds an assistant to your helpdesk. It can answer customers in the chat, or write drafts that your agents review. It works from your policies and your store's orders. When a conversation needs a person, it hands the conversation to your team. Your organization brings its own OpenRouter account and pays OpenRouter directly for what the assistant uses.

The admin pages are under Admin → AI. AI settings, Complexity and the group AI descriptions need the AI settings permission. Policies needs AI knowledge, and Usage needs AI usage. The Admin role has all three. See Roles and groups to give them to other roles.

Connect OpenRouter

  1. In OpenRouter, create a key under Keys. Give the key a credit limit: it caps what the assistant can spend.
  2. In ReplyX, open Admin → AI → AI settings and choose Add key. Paste the key and choose Save key.
  3. Choose Test connection. The result says whether the key works and how much credit it has left.

The key is stored encrypted. After you save it, ReplyX only shows its last four characters (•••• AB12). To change it, choose Replace key. To stop all AI, choose Remove key.

If OpenRouter rejects the key or the account runs out of credit, AI pauses for your whole organization. Admins get a notification, and AI settings shows an AI paused banner. While AI is paused, every conversation goes to your team as if AI were off. To resume, replace the key or add credit, then choose Test connection.

Choose how the assistant replies

Under Assistant → Replies in AI settings:

ModeWhat happens
OffThe assistant does not answer.
Drafts for agentsThe assistant writes a draft reply on each ticket. Customers never see it until an agent sends it.
Auto-replyThe assistant answers customers directly in the chat, and hands over to your team when it is unsure.

Under Models you pick the model for each level (Fast, Standard, Advanced). Each model shows its price per million tokens. Most conversations use Fast. See Complexity for how the level is chosen.

Add your policies

The assistant answers only from your policies and your orders. It never makes up rules, and it never promises refunds, discounts or exceptions.

  1. Open Admin → AI → Policies and choose Add policy.
  2. Give it a Title, then either Type text or Upload file (a text, Markdown or PDF file up to 10 MB).
  3. Choose Add policy. The policy shows Processing, then Ready once the assistant can use it. If it shows Failed, the file had no readable text or could not be checked.

To change a policy, open it and choose Edit policy. The assistant keeps using the old text until the new one is Ready, so customers never get half an answer. Deleting a policy removes it from the assistant's answers immediately.

Short policies with clear headings work best: one heading per topic (returns, refunds, shipping, warranty).

Connect your Spree store

With a store connection, the assistant can look up an order's status and tracking when a customer gives an order number.

  1. In the Spree admin, create a read-only API key that can read orders.
  2. In AI settings → Store (Spree), choose Connect your store. Enter the Store address (the public https:// address of your store) and the Store key (read-only).
  3. Choose Test connection.

The assistant only shares an order with the customer whose email address is on that order. Any other order number gets the same answer as a number that does not exist. Private-network and non-https addresses are refused.

Hand-off, take over and hand back

The assistant hands off to your team when:

  • the customer asks for a person
  • the customer asks for a refund, compensation or an exception
  • the customer is angry
  • it is not confident in its answer
  • the conversation stays unresolved for two turns in a row

On a hand-off it stops answering on that ticket, tags the ticket AI hand-off and puts it back in your team's queue. In Auto-reply it also tells the customer: "I've passed this to our team; a person will reply here soon."

When an agent sends a reply on a ticket, the assistant stops answering on it. On a ticket where the assistant is answering, you can also:

  • choose Take over to answer the customer yourself. You become the ticket's owner and the assistant stops.
  • choose Hand back to AI to let the assistant answer again from the customer's next message. This is only available in Auto-reply, and not while AI is paused.

The "Now replying" badge

At the top of the ticket, and in the customer's chat header, a badge says who is answering: Assistant · AI, the agent's name, or Support team. It changes live when someone takes over or hands back. Messages the assistant wrote show Assistant · AI as their author, for your team and for the customer.

Drafts

In Drafts for agents mode, an AI draft bar appears above the reply box when a draft is ready. Choose Use to put the draft in the reply box, edit it, and send it like any reply. Choose Discard to throw it away. A ticket has at most one draft. A newer customer message replaces it.

Routing new tickets with AI

Your routing rules always run first. When no rule matches a new customer ticket, AI routing can choose a group for it:

  1. On Admin → Groups, edit each group that should receive AI-routed tickets. Fill in its AI description, for example "Invoices, charges, payments and refunds". Groups without a description, and inactive groups, are never chosen.
  2. In AI settings → Routing, turn on Route new tickets with AI and set Confidence needed (70% by default). Choose Save routing.

If the assistant is less sure than the confidence you set, or no group fits, the ticket stays in Needs Triage. If someone sorts the ticket by hand first, their choice stands. The ticket's history records each AI routing decision and how confident it was.

Complexity: which model answers

Admin → AI → Complexity decides which conversations need a more capable model. Each factor gives a conversation a value from 0 to 1. Factors with more Weight count more. Their weighted average is the conversation's complexity score:

Factor typeWorth 1 when…
Yes or no questionthe AI answers yes
Scalethe highest level fits (lower levels are worth less)
Multiple choicethe option you valued at 1 fits
Customer has a tagthe customer has the tag
Times reopenedthe ticket was reopened 3 times or more
SLA breachedthe ticket has missed an SLA target

The two sliders under Levels set where Standard and Advanced start. The preview below them shows the score range of each level. The default settings start Standard at 45% and Advanced at 80%, so most conversations stay on the cheaper Fast model. When you first turn AI on, a starter set of factors is added: customer frustration, several issues in one message, and a technical problem, plus a VIP factor if you have a tag named "VIP".

Usage

Admin → AI → Usage lists every AI call: when it ran, what it was for, the model and level, the tokens, the cost and the outcome. You can filter by date and purpose, and the totals follow your filters. A cost marked ≈ is an estimate. No message content is stored with these records.

On this page