What You're Actually Approving When You Press That Button
Arvio writes the change first, then stops and shows it to you. Nothing reaches your store until you press one button.
So the interesting question isn't "does it work." It's: what exactly is on the screen in the moment before you press?
Here's the whole card, transcribed from a real run on our own demo store:
| What's on the card | The text |
|---|---|
| Title | Confirm Description Rewrite — Acacia Round Tray |
| Reversibility marker | (reversible) — in green |
| Which product | Acacia Round Tray |
| Scope line | Only this product will be updated — all others left untouched. |
| Current Description | Solid acacia, finished by hand. Hand wash and dry immediately. — struck through |
| New Description | An editable text box, with a resize handle in the corner |
| The button | Confirm Update — black, not yet pressed |
Each of those is answering a question you'd otherwise have to go and ask.
What that card is actually telling you
This is the state we spend the most time looking at: the change is fully written, and none of it has happened yet.
The title names the operation and the object. Not "Arvio wants to make changes" — Confirm Description Rewrite — Acacia Round Tray. You know what kind of edit and to which product before reading anything else.
(reversible) sits right under it, in green. This is a property of the operation, shown up front rather than discovered afterward. It's telling you the cost of being wrong here is low — which is a different decision than one where the cost of being wrong is high.
The scope line is the one I'd read twice. It says: Only this product will be updated — all others left untouched. This is worth being precise about, because it's easy to over-read. The prompt for this run was "Rewrite the product description for one product only: Acacia Round Tray. Leave every other product untouched." So that line is Arvio confirming back the scope of this task — it's a receipt for what you asked, not a standing promise about how the product behaves in general. If you ask for a change across a category, the scope line will say something else. Read it as "here's what I understood you to want," and check that it matches.
The current description is shown with a strikethrough. You don't have to open another tab to remember what's there now. The thing being replaced is on screen next to the thing replacing it.
The new description is in an editable box, not a block of display text. More on this below — it's the part that changes what "approve" even means.
And Confirm Update is a button that hasn't been pressed. That's the state the whole system is designed to park in.
Around the card, the same screen shows a few things that aren't the decision but frame it: Plan complete 39s ▼ (you can expand what it did), 2/6 connected, Not subscribed, a Was this answer helpful? / Useful / Needs work prompt, and a standing line at the bottom: Arvio is AI and can make mistakes — please verify before acting. That last one isn't decoration. The card exists because of it.
We ran it once. Here's what happened to the catalog.
Card text is one thing; what actually landed in the store is another. So we diffed the catalog before and after.
The description on that one product went from 62 characters to 364 — roughly six times longer — and the other 67 products came out of the run byte-identical.
- 68 products before, 68 products after. Nothing created, nothing deleted.
- 1 product changed. The one named on the card.
- 62 → 364 characters on that product's description.
The old text was Solid acacia, finished by hand. Hand wash and dry immediately. — accurate, and about as much help to a shopper as a shipping label. The new text describes the grain, suggests what you'd put on it, and keeps the care instruction at the end.
That's one product, one prompt, one store — ours. It shows the shape of the edit, and that the diff matched what the card said it would be. What happens on a large catalog with several people editing at once is a different question, and we haven't measured it.
You can rewrite it before you approve it
The New Description field is a text box. There's a resize handle on it. That single detail splits "approve" into two different actions:
- Approve what it wrote. You read it, it's good, you press.
- Approve your edit of what it wrote. You change a phrase, fix a product detail it couldn't know, cut a sentence — then press.
The second one is the case that matters most in practice, and it's the case a pure yes/no gate throws away. If the only options are accept and reject, then a draft that is mostly right is a rejection, and you're back to writing it yourself from a blank field. With an editable box, a draft that is mostly right is a small edit.
This matters for exactly the things a model can't get from your catalog data. Whether the acacia is FSC-certified. Whether you've been calling it a "serving tray" in your ads for two years. Whether that care instruction is legally required wording in your market. Arvio drafts the structure and the language; you supply the facts only you have.
Whatever's in the box when you press is what gets written. That's the whole contract.
What this looks like when it's not one product
The run above is deliberately the smallest possible case: one product, so the diff is easy to check. If you came here with four hundred products, the question you actually have is what this card looks like multiplied.
Two things we can point at, and one we can't. Our App Store listing describes the scope choice as One product, a batch you pick, or all of them — you choose the scope, and the approval step as Approve or reject each change one by one; undo any edit in one click. That's the documented behaviour.
There is one recorded run on a batch rather than a single product, and it is public: a price change applied across a product grid, with the confirmation card read line by line before the press.
What we haven't published is a measured run at that size — how the cards are paced, what a hundred of them in sequence feels like, where the review stops being careful. We've written about the attention problem further down, because it's the part we'd want to know about if we were you. But we're not going to describe a four-hundred-product session we haven't recorded and shown you.
The card comes before the write
Here's what we measured: the card appeared, we pressed the button, and the catalog changed. The diff is the result of that press — 68 products in, 68 out, one of them different.
So the card sits before the write, not after it. Up to that press the change is a proposal held in the conversation: the draft exists, the store doesn't know about it. Nothing on the card is an edit that already landed and is waiting to be un-landed.
That last sentence is read off the order of events, not off a recording of someone declining — our capture of this run goes prompt → card → press → done, and we don't have footage of a rejection to show you. If you want certainty for your own store, it's a cheap thing to check on one product: note its current description, walk up to a card, close it without pressing, and look at the product again.
The timeline of our run shows one more thing. (Not a speed claim — that capture was sped up, so the durations in it would be misleading.) It's this: once the card is on screen, it stays there. The gap between the card appearing and the button being pressed is however long the person took. Nothing times out and decides for you.
When you should actually reject
The card is only worth having if you sometimes say no. Concretely, these are the moments to stop:
The scope line doesn't match what you asked for. You asked for one product; the card describes a category. That's not a bad rewrite, it's a misread instruction, and editing the text box won't fix it. Reject and re-prompt with tighter wording.
It asserts a fact you can't verify. Material, origin, dimensions, certification, compatibility. Our new description says "solid acacia wood" — true, because the old description said so and it's our store. If a draft introduces a specification that wasn't already in your catalog, treat it as a claim needing a source, not as copy. This is the failure mode I'd watch for most closely: fluent, plausible, and wrong is harder to catch than obviously broken.
It quietly drops something load-bearing. In our run the care instruction survived — hand wash and dry immediately shows up in both versions. Compression tends to delete qualifiers first, and qualifiers are often the legally or practically important part. Before you press, check that every constraint in the old text is still somewhere in the new one.
It's off-voice in a way that's easier to re-prompt than to hand-edit. If you'd have to rewrite more than half the box, don't. Reject, adjust the instruction, and let it draft again. Editing is for fixing a draft, not for salvaging one.
You're not in a position to read it properly. The card waits. That's the feature.
None of these require you to know anything about the model. They require you to know your products — which you do, and which is exactly why the gate is where it is. We wrote separately about why Arvio hands you a finished change instead of a list of problems to go fix yourself; this is the other half of that argument. A report puts the work on you. A finished change with no gate puts the risk on you. The card is the position where you carry neither.
After you press
The result card restates both versions in full, so the record of what changed lives in the conversation and not only in your store.
The confirmation resolves into a result card:
Description updated — Acacia Round Tray
All other products left untouched.
✓ Description updated
Old description: Solid acacia, finished by hand. Hand wash and dry immediately.
New description: Crafted from solid acacia wood and finished by hand, this round tray
brings warm, natural grain to any table. …
💡 This action can be undone. Let me know if you'd like to revert it.
Result Verification / View in Shopify Admin →
Three things here do real work:
Both versions, in full. Not "updated ✓" — the old text and the new text, one above the other, after the fact. If you approve a batch of these and something looks wrong tomorrow, the old string is still sitting in your conversation history.
This action can be undone. The (reversible) marker from before the press, restated as a fact after it. And the mechanism is a sentence, not a hunt through settings: you tell it to revert.
View in Shopify Admin →. The link out to the source of truth. We use it constantly, and I'd suggest you do too for the first several changes — not because we expect a mismatch, but because "verify before acting" is printed on the screen for a reason, and the fastest way to build a calibrated sense of when you can skim a card is to check a few of them all the way through.
If you want to see what the drafts themselves look like across more than one kind of edit, that's the subject of our walkthrough of what a finished Arvio change actually reads like.
Where this gets uncomfortable
The honest tension in this design: a gate you press a hundred times stops being a gate.
Card ten gets read carefully. Card ninety, at the end of a Tuesday, gets a glance and a press. We know this, and we haven't solved it. The design choices above — showing the old text struck through, restating both versions afterward, putting the Admin link in the result — are partly hedges against the moment you approve on autopilot. They don't eliminate it.
So front-load your attention. Read the first ten cards all the way through, click into Admin on a few, and build a sense of where this thing is reliably good and where it needs you. Then you'll know which cards you can move through quickly and which ones deserve the full read — and that judgment is worth more than any promise we could put in a marketing page.
The other limit worth naming: what we've shown you here is one operation type on one product. Description rewrites are a good case for this pattern because they're reversible, self-contained, and easy to eyeball. Not every operation has those properties. Price changes touch what customers pay; deletions don't have an obvious "before" to restore from memory. The test we'd apply before approving anything quickly is whether you could describe the undo out loud — if you can't say how you'd put it back, that's a card to read all the way through, whatever the marker says. We'd rather give you that test than generalize from the one operation we've measured.
Arvio: AI Store Operator — install it on the Shopify App Store. It drafts the change, then stops and shows you the card.
FAQ
Does Arvio write to my store before I approve?
The card appears before the write, and the diff we measured is the result of pressing the button. We only captured the approval path, so I'll describe what the structure implies rather than claim we filmed the alternative: at card time the change is a proposal, not something already applied. If you want certainty, test it on one product and check the result in Admin.
Can I change the text before approving?
Yes. The New Description field is an editable text box with a resize handle. Whatever is in it when you press Confirm Update is what gets written — so "approve" and "approve my edited version" are the same button.
What does (reversible) mean?
It's a marker on the card, shown before you press, and it's restated on the result card as This action can be undone. You revert by telling Arvio to revert it.
Does "all others left untouched" mean it will never modify other products?
No, and this distinction matters. That line confirms the scope of the task you asked for — in our run the prompt explicitly said "leave every other product untouched," and the catalog diff backs that up: 68 products before, 68 after, exactly 1 changed. If you ask for a broader change, the scope line will describe a broader change. Read it every time; it's a receipt for one task, not a permanent rule.
How long does the card wait?
As long as you do. In our run the gap between the card appearing and the press was however long the person took to read it — nothing expires or decides on your behalf. (I'm deliberately not quoting durations from that capture: the footage is sped up and any number would read as a speed claim we haven't earned.)
How much does a description actually change?
In the run we diffed, one product's description went from 62 characters to 364 — roughly six times longer. That's one product on our own demo store, and I wouldn't extrapolate an average from a sample of one. The number is here because it's what we measured, not because it's a benchmark.