AI Assistant

The Assemblified AI assistant is a conversational layer over your Assemblified data. Ask questions like “how many shirts can I build with current stock?” or “which bills are blocked on materials?” — and the assistant pulls the answer directly from your live data. Ask it to change something, and it shows you exactly what it is about to do before it does it.

Where to find it

Two ways in, one assistant:

The AI Assistant page — the AI Assistant entry in the app nav opens the full-page workspace, with room for long answers, tables and file work.
Ask AI, from the page you are on — a page’s title bar carries an Ask AI action. On a wide screen it opens the assistant in a column beside the page, so you keep the list or the record in view while you ask about it. On a narrow screen there is no room for two columns, so it opens the full page instead.

What it can do

The assistant reaches your data through exactly the same surface a connected external agent uses. That means its abilities are the ones listed in What the assistant can read and on the operating pages — bills of materials and assembly BOMs, materials and stock, work orders and build runs, reports, catalogues, settings and the rest. Nothing is special-cased for the in-app assistant, and nothing is hidden from it that an agent with the same access could see.

In short, it can:

Answer questions about your data — what you can build, what is short, what something costs, what a setting does, what a run did and why it failed.
Find things — the same cross-entity search the app’s search field runs, so “the Oak table one” resolves before it goes and reads anything.
Make changes, with your approval — create and edit bills, adjust stock, plan and run work orders, change settings.
Work with a spreadsheet you attach — drop a CSV, XLSX, JSON or text file into the conversation and it can inspect the sheets, tell you what is in them, propose how they map onto your data, and prepare an import for you to approve.
Link you to the right page — when it names a record or a page, it writes a link. It never navigates you somewhere on its own; you decide whether to click.

Two things it deliberately will not do: run an assembly flow, and edit your Shopify catalogue. Those stay where you can see them.

It knows what page you are on

The assistant sees the page you had open when you asked, and the rows that page was showing: the bills of materials in the list, the materials you had ticked, the work order you were looking at. So “what is short on these” works on a list with three rows selected, and “how much does this cost” works on a detail page without you naming the item. It sees a bounded summary, never your whole catalogue, and selected rows always come first.

On a settings page it goes one step further: it knows which settings that page owns, and it is told to check whether a setting can be changed — and what would change with it — before it offers to change it. That is why asking “turn this on” on a settings page usually gets you an explanation of the consequences rather than a silent switch.

How conversations work

Each conversation is a thread, kept under its own name, with a chat switcher for moving between them and New chat for starting a fresh one. A thread you no longer want is deleted outright; there is no archive.
Threads persist across page loads and between the panel and the full page — it is the same conversation in both.
Long threads are trimmed. A thread keeps its recent history rather than growing forever, so a conversation that has run for days may no longer remember its own beginning. For a new topic, start a new chat: it is faster and the answers are better.
Answers stream in as they are produced, and the steps the assistant takes are shown as it takes them, so you can see what it read before it answered.

Approvals

Reading your data never asks for permission. Anything that CHANGES something does: the assistant shows an approval card that names exactly what it is about to do before it does it, and nothing happens until you press Approve. Rejecting leaves everything as it was. The assistant never moves you around the app on its own: when it mentions a page or an item, it links it, and you decide whether to click.

Every change made this way is recorded against YOUR staff account, not against “the assistant”.

The approval gate is enforced where the change would happen, not by the assistant’s own judgement: if it tries to write something you were not asked about, the write is refused rather than quietly performed.

When a run fails or is cut short

Sometimes a step the assistant takes fails partway, or its answer is too long to show in full. The step then carries a short safety note in the conversation, so you are never left guessing:

Whether anything was saved. The note says the step stopped before saving any changes, that it may have saved changes (naming the actions involved and asking you to check what was saved), or that it is not known. Missing information is never treated as “nothing changed”.
Whether repeating it is safe. A step that may have saved changes is marked as not safe to repeat.
What was left out. When an answer was shortened, the note says so and how large the full answer was.

The assistant follows the same rule itself: it never runs a step again when that step may have changed something. Instead it checks what actually happened with small, targeted reads (fetching just the records or the page it needs) and tells you what it found. It only retries on its own when the failed step is known to have saved nothing.

Privacy and data handling

Your question, and the data the assistant reads to answer it, are sent to a hosted AI model so it can produce a reply. Treat a conversation the way you would treat any other cloud service you paste business data into.
It only ever sees your store. Every read and every write is scoped to the store you are signed in to.
It reads what it needs, not everything. Answers are built from specific reads it makes at the time, and the summary of the page you are on is capped in size before it leaves your browser.
Files you attach are stored so the assistant can work through them across several turns, and they belong to your store like any other data in the app.
Everything is logged. Assistant activity is recorded in the same access log as API keys and connected agents, attributed to your staff account — see Managing access & security.