The assistant drafts. You confirm.

There is one rule and the whole feature is built around it. The assistant may read your shop. It may not change it. Everything it would change comes back as a draft, and a person presses the button.

The page says the same thing at the top: Grounded in this shop's data. It can look things up on its own; anything it would change comes back as a draft for you to confirm.

What it can look up

Nine things, and nothing else. It can find a customer by name, company, phone, email, plate or VIN, then read that customer's next due service, open orders, unpaid balance and last visit from the same summary as their record. That summary only reads; it never writes anything. It can read a vehicle's history, including visits imported from your old system. It can find a repair order by number or by the customer's name. It can tell you how today looks — cars in the shop, ready for pickup, waiting on approval. It can read Customer Aging for unpaid accounts and today's End of Day report. It reads those reports; it does not chase, call, or message anyone on them. It can search your own spec library. And it can search this help centre, so product questions get answered from real articles rather than from a model's imagination.

Under each answer it prints which of those it used, starting looked up: — so you can see whether it actually went and checked, or was talking.

It never sees more money than you do. Money visibility is four switches now rather than one, and the figures the assistant is willing to quote are governed by the one called The shop's own figures. Without it the assistant is told not to state them and they are stripped out of what it is shown, rather than merely left unmentioned. Totals a customer would see on their own invoice are still fine. Who can see what a part cost you has the rest.

The drafts on the assistant page

A text to a customer, an internal note, a job on an order, or an estimate. Asking for any of those creates a draft — nothing more. The draft appears on the same page under Waiting on you, showing what it would do in one line.

1. 1

   Step 1: Read it

   The summary says exactly what will happen: who the text goes to, which order the job lands on.

2. 2

   Step 2: Edit the wording if you want

   The body is an ordinary editable box. What you send is what is in it when you press the button.

3. 3

   Step 3: Press Confirm and do it — or Discard

   Confirming runs the same code as if you had typed it yourself, which means the same consent checks, the same permission checks, and a line in the audit trail recording that a draft was confirmed and whether a human edited it first.

If a confirm is refused — most often because the number has no texting consent — the draft stays where it is with the reason attached. It is not thrown away and it is not half-sent.

There is a fifth kind of draft that does not start on this page. A technician can say what they found out loud on their own ticket, and the assistant sorts it into a cause and a correction — as a draft, on the job, with the same confirm step and the same rule.

There is also a pricing-matrix draft. The assistant can analyse sold, non-test part lines and propose changes to margin bands, with the current values, proposed values and evidence kept together. It does not appear under Waiting on you. It is reviewed on the pricing-matrix settings page, where a person with the settings and money keys confirms or discards it. The analysis never changes a matrix by itself.

Running with no model connected

Torqix works with no AI service configured at all. In that state the assistant answers direct lookups from your own data and keyword-searches the help articles, using plain rules and no network call, and it says so on the page: Running without a model connected: it answers direct lookups and searches the help articles, nothing more. Same tools, same drafts, same confirm step. Connect a model and the conversation gets better; the rule does not change.