Stock control

AI and automation in Stock

What the AI can do in Stock and how to choose how much of it runs unattended, task by task.

AI and automation in Stock

Stock uses AI in several places to save you manual work — suggesting how a product maps to ingredients, drafting recipes, reading invoice photos, proposing reorders, and surfacing recommendations. AI & Automation, under Settings, is where you control how much of that happens without a person checking it first, one task at a time.

The six tasks

Each of the following is controlled independently:

• Product mapping — suggesting how a product should be tracked
• Modifier mapping — the same, for modifier options like extras and add-ons
• Invoice reading — extracting line items from a photographed or uploaded invoice, and from the Invoice Inbox
• Order suggestions — drafting purchase orders from par levels and sales
• Par levels — proposing changes to the pars themselves, on the Par levels page under Ordering
• Recommendations — proactive suggestions like flagging dead stock or noticing a supplier's fill rate has dropped

On/Off tasks, and tasks with levels

Some tasks can only be on or off. Those show a simple On / Off switch — Off means the AI is never used for that task, and On means it proposes and you approve. There's nothing else to choose, so there's nothing else shown.

The tasks that genuinely have more to offer show a dropdown instead:

• Off — the AI is never used for this task
• Suggest — the AI proposes an answer, and a person always reviews and confirms it before anything changes (the default on a new install)
• Apply when confident — the AI applies its own suggestion automatically when it's confident enough, and only queues lower-confidence results for you to check. Choosing it reveals a min confidence box where you set the bar.
• Always apply — the AI applies its suggestion every time, no review

The list you're shown for a task is exactly what that task will act on: if a level isn't offered, it's because nothing in Stock would behave any differently had you picked it. That's also why some tasks are a switch rather than a dropdown — a choice between two things that do the same thing isn't a choice.

The tasks that are simply On or Off

Order suggestions is the one task whose output leaves the business. Approving a suggested order is what emails it to the supplier — the approved order goes out automatically within about ten minutes — so "approve this for me" and "send this to a supplier unattended" are the same instruction. A mistake there costs real money and can't be undone by editing a record afterwards, so the approval always stays with a person. Off stops the nightly drafts being created; On creates them for you to review. There is nothing in between, and nothing above.

Par levels proposes figures on a review page and nothing else — there is no overnight job behind it and no model in front of it. The proposed par is worked out arithmetically from your recent usage, so there is no AI confidence for a threshold to compare against, and applying pars for you would be changing what gets ordered without anyone reading it. Off stops the suggestions appearing at all, for a venue that sets pars by hand; On shows them for you to accept or turn down. See Suppliers and ordering.

Recommendations flags price rises, dead stock and suppliers whose fill rate has slipped. Every one of those names a judgement rather than an action — there is nothing for Stock to "apply" on your behalf, since closing the alert without doing anything would just throw the warning away. So they always wait for you in the inbox. Off stops the AI wording being written (the alerts are still raised, in plainer language); On has the AI write them up.

Changed recently. Recommendations used to offer Auto when confident and Full auto as well. Neither ever did anything different — the one kind of recommendation that could be applied automatically was par levels, and that moved to its own page — so the task is now simply On or Off. If you had picked one of those two, you'll see On, which is what you were already getting. Nothing about your recommendations changes.

Recipe drafting is no longer a setting

AI draft recipe — the button on a row of the Product Mapping screen that proposes the ingredients and quantities behind a combination item — used to have a switch here. It no longer does, and there is nothing you need to do.

Drafting only ever ran when you pressed that button yourself, one row at a time, and the result opened in the recipe builder for you to check before anything was saved. That row can be a product or a modifier option — a sauce or a side salad is made from several ingredients just as a dish is, so both can be drafted the same way. Because a person starts it and a person reviews it, there was never an unattended step for the setting to hold back.

If you had turned recipe drafting off, that is a real change for you, and the only case where anything changes at all. New businesses always started with drafting switched on, so the setting only ever did something for someone who deliberately turned it off — and what it did was let the button still appear on the screen and then refuse when pressed, which looked like a fault rather than a choice. The button now always works. If you would rather your team didn't draft recipes with AI, that is now a matter of who can create recipes at all — see Users and roles.

Choosing which AI model Stock uses

Further down the same page are two model settings. You pick from a list of the models your account can actually use — you no longer type a model name in by hand, so there is nothing to spell correctly and no version number to keep up with.

• General model — used for catalogue classification, draft recipes, order-suggestion wording and recommendations. Left alone, it follows whatever model you chose in POS, which keeps the two in step.
• Invoice reading model — used for invoices from every source: uploads, the Invoice Inbox and photos of delivered paperwork. Until you pin one here it covers photos only, and uploaded PDFs go to the general model instead — the setting says so underneath it. Pin a model here if you want one model reading every invoice however it arrived.

Open either one and you get a search box and two groups. Recommended comes first: a short list named by family — Gemini Flash, Claude Sonnet, GPT mini, Gemini Pro — each with a line on when to pick it and the exact model it works out to today. Below that is All models, everything your account can reach, grouped by who makes it. Type in the box to narrow either group.

Each row shows a rough running cost beside it, so you can weigh accuracy against price without leaving the page. Models that cost nothing to run read Free, and a dash means the price could not be read — a dash is not a claim that the model is free.

The recommended names are families rather than fixed versions, so when a maker releases a newer version the recommendation follows it on its own. Choosing one saves the specific model it resolved to at that moment, so nothing changes underneath you until you come back and pick again.

Why the invoice list is shorter

The invoice reader is handed a picture — a scanned page or a photo — so it can only use a model that can look at images. Most models cannot. Rather than let you choose one that would fail later, mid-invoice, the invoice list only offers models that can read an image, and the recommended group shrinks to match.

Two things you might see

"Currently set" in amber, with a warning. The model saved here is not in your account's list any more — usually because the maker retired it. It keeps being used until you pick another, and nothing is changed just by opening the page. Pick a replacement when convenient.

"No AI provider connected yet." Stock runs on your company's shared OpenRouter key rather than one of its own. Connect it in POS under Settings → Integrations → OpenRouter and the model lists fill in. The task switches above are still yours to set in the meantime — they just have nothing to run on until a key is connected.

Neither picker saves on its own. Choosing a model fills the box; Save model is what stores it.

A sensible way to turn things up

Most businesses leave everything proposing rather than applying at first, watch how good the suggestions are for a week or two, then move the tasks they trust — usually product mapping first, since the name-matching suggestions are already reliable before AI even gets involved — up to Apply when confident. There's no requirement to change anything: Stock is fully usable with every task left on Suggest indefinitely.