The three ways a transaction can still fail after the agent says yes.

It sold out in the gap between the check and the commit

Say you run a small-batch candle business and an AI agent has just told its person "done, it's on its way." From the outside that looks finished. From the business's side, three separate things still have to hold true for that sentence to end up correct, and each one fails in a different way if it doesn't.

Stock can move between the moment an agent checks availability and the moment it places the order, especially on something small-batch with genuinely limited quantity, a seasonal scent that runs to forty jars and no more. If nothing re-checks at the point of commit, the agent has already told someone their order is confirmed before finding out it isn't. The fix isn't a faster check, it's a check that happens at the moment of commitment itself, not a moment earlier when the answer was still true.

A condition wasn't really met

Some purchases carry a condition that isn't obvious from the product page: a minimum order size for a trade discount, a delivery address inside a specific postcode range, an age requirement for a scented candle with an alcohol-based fragrance oil in some regions. An agent that never checked the condition can confirm a purchase that then has to be quietly cancelled or corrected on the business's side, which is a worse outcome for everyone than catching it before confirmation. The customer hears "confirmed" twice, once wrongly, and the second correction lands worse than a single honest delay would have.

The payment doesn't go through

A declined card, an expired authorisation, a payment method the agent had permission to check but not to charge: any of these can sit undiscovered until after the agent has already said yes. The business is left holding an order it can't fulfil and a customer who's been told something that wasn't true. This is the failure most worth catching early, because it's the one most likely to look, from the agent's side, exactly like success right up until it isn't, since everything up to that point genuinely was fine.

None of these three are exotic edge cases. They're the ordinary ways a transaction that looked complete at the moment of confirmation turns out not to be, and each one is a timing gap between what an agent checked and what was still true a moment later. A hotel losing a room between check and confirm, a business confirming an order that turns out to miss its own delivery-area rule, a payment that fails after the "yes" was already given, all three are the same shape wearing a different vertical's clothes.

What makes all three worse than the equivalent failure with a person at the till is that nobody's standing there to absorb the awkwardness in real time. A cashier whose card machine declines a payment has a quiet, immediate conversation and the customer tries another card. An agent that finds out after telling someone "done" has to deliver a correction to a person who already believes the matter's settled, which lands as a broken promise rather than a normal hiccup, even though the underlying event, a declined card, is exactly the same thing that happens at a till every day without anyone thinking twice about it.

Closing all three means the same thing: a live connection to the systems that hold stock, conditions and payment status, checked at the point of commitment. That's the specific layer Selfe exists to build, so "yes" from an agent means the same thing it would mean from a person standing at the till.

Can't the agent just retry if something fails?

Sometimes, but a retry after telling someone "it's confirmed" is a worse experience than catching the gap before that sentence was said at all.

Is this more of a risk for smaller or limited-stock businesses?

Yes, disproportionately. A business with deep, stable stock rarely hits the first failure; one selling something genuinely limited hits it often, which is exactly why the check matters more there, not less.