There is a moment in every AI project where somebody says: "why not let it just do it?"
The AI has drafted the purchase order. The order is correct. Everyone can see it is correct. Making a human tap approve feels like ceremony - friction added for its own sake, the kind of thing that makes software annoying.
We have thought about that moment a lot, and we have landed firmly on the other side of it. Rai drafts. You approve. Always.
The asymmetry nobody mentions in demos#
Reading and writing are not two versions of the same risk.
If Rai reads your sales and gets a figure wrong, you look at it, something feels off, and you check. The cost is a few minutes and some mild irritation. The mistake is contained to a screen.
If Rai writes and gets it wrong, the mistake goes into your books. It flows into your stock counts, your supplier balances, your accounts. You do not discover it when it happens - you discover it weeks later when a reconciliation does not balance, and by then you have to unpick which of a hundred entries was the fabricated one.
What Rai actually does at the boundary#
Rai has two features that would change your data. Both stop at the same wall.
Drafting a purchase order. Rai notices a fast-moving item is running low, works out a sensible quantity from your own sales history, picks a supplier based on what you have actually paid each of them, and prepares the order. Then it shows it to you: item, quantity, supplier, price, and the reasoning. Nothing is created until you say so.
Reading a supplier invoice from a photo. You snap a picture of an invoice or a delivery note. Rai reads it and drafts the purchase entry - lines, quantities, prices. Again, it shows you the draft next to what it read, and waits.
In both cases the useful work has already happened. The typing is done, the arithmetic is done, the lookup is done. What is left for you is the one part a machine should not do on its own: deciding that this should become real.
Automate the labour. Do not automate the decision.
The four things behind every write#
Approval on its own is not enough - it is just a button. Underneath it there are four things, and they matter equally.
Permission scoping. Rai acts as the person using it, through the same permission model your system already enforces. A staff member who cannot create purchase orders by hand cannot get Rai to create one for them either. The AI is not a back door around your access rules.
A full audit trail. Every question asked, every query run, every answer given, and every write approved is recorded - what happened, and who approved it. This is the part your accountant will care about, and it is also the part that lets us diagnose anything that ever looks strange.
Rollback. Every write path can be undone. Not "we can probably fix it if you call us" - actually undone.
Read-only by default. Everything that is not one of these two features is read-only, structurally. Rai cannot modify things it has no write path for, regardless of what it is asked.
The benefit people do not expect#
Owners come to the approval step expecting to tolerate it. What they usually report back is that it turned out to be useful for a different reason.
When Rai drafts an order and shows its reasoning - this item, this quantity, because it has been selling at this rate and you have this much stock left - the approval screen becomes a small, regular lesson in your own business. You start noticing the patterns behind the drafts. Some owners have told us they began changing their reorder habits after a few weeks of reading Rai's reasoning, independently of anything Rai actually did.
An action taken silently teaches you nothing. An action proposed, explained and approved teaches you something every time.
How to judge any AI tool on this#
If you are looking at AI for your business - ours or anyone's - three questions will tell you most of what you need to know.
What can this system change on its own? If the answer is anything at all, find out exactly what and under what conditions.
Can I see what it did? If there is no audit trail, there is no supervision - only hope.
Can I undo it? If not, then a single bad automated decision is permanent, and you are one bug away from a very bad week.
We built Rai so the answers are: nothing without your approval, yes in full, and yes. Not because we think our AI is unusually likely to be wrong, but because a system you can check is a system you can keep using after the first time it surprises you.
