Suggested Replies

What It Reads First

Before every suggestion, the assistant reads:

• the whole conversation on the ticket, including your team's internal notes (which it never quotes to the client);
• the client's other tickets, by their email address and their company;
• their chats from your website (Chat);
• the automated emails sent to them (the subjects and which app sent them, from Recent Emails), and what your other systems told Helpdesk about them (see From Your Other Systems);
• their projects and proposals;
• their subscriptions, invoices, and whether anything is past due;
• your catalog's prices, when the client asks about money.

It reads the newest messages in full and sums up older ones in a line each, so even a long history fits. Under the suggestion, What It Looked At lists each thing it read, with a link to each record (action helpdesk.replies.get).

A suggestion shows only to people who may see the ticket and everything the suggestion read, including prices when it read any. Anyone else sees Make One for Me, which makes a suggestion from only what they may see.

The Suggestion

The card above the reply box shows the suggested reply, the kind of case (such as Pushback on price or Outage complaint), and anything to check before sending. Its buttons:

• Use This Reply puts it in the reply box, where you change anything you like and send it with the usual Send buttons.
• Regenerate (action helpdesk.replies.regenerate) makes another one. Add a short note first, such as "shorter", "firmer", or "mention the staging link", and the new one follows it.
• Dismiss (action helpdesk.replies.dismiss) puts it away until the client writes again.

A new suggestion is made a minute after the client writes (so two emails in a row are read together), and again when the ticket changes: the client writes again, someone adds an internal note, or your other systems say something new about the client. When your team replies, the suggestion is done.

How It Writes

Every suggestion is friendly and creatively firm:

• Calm and warm, never defensive or emotional, however the client writes.
• It acknowledges how the client feels without agreeing to facts or work it has not checked.
• It holds the line on scope, price, and timelines, and always offers a way forward: options, the next step, or a paid change with a quote.
• It never blames the client, argues, or uses sarcasm, and never promises a date, refund, credit, or outcome your records do not show.
• It never makes up a price. It quotes only prices in your records (the catalog, a proposal, an invoice, a subscription), and otherwise says you will send a quote.
• It writes in plain US English and never uses dashes as punctuation.

The Review-Risk Check

Before you see a suggestion, it is checked for wording that could read badly if the client posted it in a public review: anything rude, dismissive, blaming, sarcastic, shouted in capital letters, or a promise your records do not support. Any dollar amount that is not in your records is flagged too. Each flag says why and offers softer wording; Use Softer Wording puts it in place.

How It Learns Your Voice

The assistant learns from the replies of the people you choose (your owners, unless you choose others), and from your team's edits:

• The style profile describes how your team writes: tone, phrasing, greetings and sign-offs, and how you handle pushback. It is written from your strongest replies, with recent years counting more than older ones, and written again as your team keeps editing suggestions.
• Similar past cases: for each new message, the assistant finds your closest past cases and follows how they were handled. Only the approach is used, never the facts: prices, dates, numbers, addresses, and links in past replies are blanked before anything is kept, and today's facts come only from today's records.
• Your edits: when someone changes a suggestion before sending, the suggestion and what was sent are kept side by side, and later suggestions move toward what your team actually sends.

Replies by people the assistant does not learn from never shape the style profile. They are left out, or kept as low-weight examples when they already match your style.

Everything it learns stays in your company's own database, the same as your tickets.

Learn from Past Replies

Once, an owner or admin can have the assistant learn from every past reply in your Helpdesk. Learn From Past Replies in Helpdesk Settings first shows what it would read (your tickets and replies by year, and how many are by the people it learns from) and what it should cost (action helpdesk.replies.history.preview), then runs in the background with its own spending limit when you choose Start Learning (action helpdesk.replies.history.start). You can stop it at any time (action helpdesk.replies.history.stop); what it learned so far is kept and used. After it, every new reply is learned from as it is sent.

Settings

In Helpdesk Settings, Suggested Replies (owners and admins; actions helpdesk.replies.settings.get and helpdesk.replies.settings.save):

• Turn On or Turn Off. Turning it on needs your Claude API key in Settings, AI Agents.
• The assistant's name your team sees, such as "What Assistant Looked At".
• Monthly limit: suggestions stop for the month when the month's cost reaches it (until the 1st), and your team replies by hand as always. Owners and admins are told in the workspace when most of it is spent. The card shows this month's cost, how many suggestions were made, and the cost of each on average.
• Learn from: whose replies it learns from.
• Style Profile: read what it learned (action helpdesk.replies.style.get), or Reset it so it is learned again (action helpdesk.replies.style.reset; the old one is kept).

From Your Other Systems

Your own systems can tell Helpdesk what they know about a client, for suggestions to read: automated emails they sent (reports, notices, invoices, receipts), the client's services and their states, open incidents, recent work, and billing notes. Send them with an API key and the action helpdesk.context.put: the client's email, a short name for your system (the source), and its list of items, each with a kind (email, service, incident, work, or billing), a title, and optionally a state, a link, and a date. Each call is one client's whole list from that source, and takes the place of its last list. Only titles and states are kept, never a message's contents.

Cost

Simple tickets use a faster, lower-cost model, and complex or heated ones (pushback, refunds, scope, outages, long conversations) the most capable one. The rules and your style profile are cached, so most of each request costs a tenth as much. Each suggestion's cost is recorded, and a typical one costs a few cents.