Checking has long been done — registers, business information, creditworthiness, one's own history. What is new is not that checks are made, but that a program obtains this information in seconds and lifts out the few cases in which it contradicts itself. The gain does not lie in a program deciding, but in a human only seeing the cases where it comes down to them.
The question is not whether an AI can decide this, but how big the stack is that afterwards still lies on a desk.
All the information says the same thing, nothing contradicts one's own history. Here the machine decides — and better than a human on a Friday afternoon: it forgets no query and does not get more impatient at the seventh check than at the first.
No valid licence, a sanctions hit, an account that arrived with the order. A program can do that too — with one condition: a machine no needs a route on which somebody can contradict it. Whoever does not provide for that loses business to a typo in somebody else's register.
The rest — and this is where the pre-check earns its money. Not by deciding after all, but by preparing the decision: what contradicts what, since when and against which source. Whoever is handed that needs minutes instead of an hour.
Their ratio is the actual key figure of this automation. Whoever does not measure it does not know whether the pre-check achieved anything or merely resorted the stack — what can be calculated from it.
Business information, credit index, commercial register, VAT number, sanctions lists — all of that exists as an interface. The part that can be fully automated, and at the same time the one into which too much is most often read.
A credit index is a model value. It has an issuer, a method and a survey date; without those three it is a figure without provenance. A program may fetch it and compare it. What it may not do is treat it like a measurement.
Whoever bases a refusal on it must be able to say, in a dispute, whose index that was and from when — otherwise a third party's information ends up standing there as one's own decision.
An index on its own says little. One that has fallen two grades since the last order says a lot — and no human notices that in passing, while a program never notices it too late. That takes no AI, only the stored previous value.
The cheapest step on this page and the one most often missing: most people query and throw the answer away. Whoever keeps it has a time series after a year — the only yardstick no credit agency supplies.
And the gap no interface closes. All this information describes a company. None of it answers whether the person writing to you today is that company — the agency knows the firm, not the sender. That is why, with a perfect data situation, one check remains that no data record takes over: confirming the contact over a connection you looked up yourself. Whoever supplies the number decides who picks up. And one piece of information is worth reading up on before querying it automatically: what lies behind the Community licence and the permit determines whether a register hit answers the question asked at all.
A blockchain is usually promised here in the wrong place. It does not prove identity; it proves that an entry has not been altered since it came into being — something else, but not nothing.
Eight months after the release the insurer asks what was checked. The answer is a log: which piece of information, when, with what result, and who gave the go-ahead. One that anybody could change afterwards proves nothing at this point.
A running chain of checksums changes that: whoever touches an entry afterwards breaks everything after it. That does not make the information true — what becomes provable is that it already stood that way at the time of release.
As long as only one house has to trust this log, a signed checksum chain does the same at a fraction of the effort. A blockchain earns its cost from the moment several parties have to trust the same entry without trusting each other: shipper, carrier, insurer, platform.
The criterion is therefore not the technology but the number of parties. This question is rarely asked — which is why a chain often ends up where a signature would have done.
For individual documents the same distinction applies — in detail under digital fingerprint: a hash confirms that a document has not changed, not that it comes from whoever is named on it.
Between “everything goes to the human” and “the machine waves it through” lies a setting that few people deliberately make — with two exits that both belong on the same desk.
The last two figures are the whole point of contention. Do not estimate them and do not take them from the vendor — you measure them by letting the pre-check run alongside for a while and decide without its decision being valid: Shadow operation.
The gain is not the hour saved — that goes into something else next week. The gain is the minutes left over for the difficult case. That is why the second line shows a time per case and not a saving.
The third figure is the uncomfortable one: it grows with every percentage point you entrust the pre-check with. It contains both directions — waved through what should not have been, and stopped what should have been allowed. The first costs a load, the second a customer.
A model value, not a measurement. The calculation is open: onto the desk comes “releases × (100 % − clearly settled)”, the time per case is “time budget ÷ what remains”, and wrongly decided is “releases × clearly settled × error rate”, times 21 working days. Nothing is transmitted, nothing stored; it runs in your browser.
Not because a human would be cleverer, but because here the information is missing from which a model could conclude anything at all.
The credit agency reports a deterioration, one's own history shows twelve trips without complaint. A model can weigh both — which source is right here is in neither of them. It then weighs by the average of all earlier cases, and the case in front of you is not one of them.
On the first transaction the model has nothing to compare. That is where the risk is greatest and the pre-check weakest — no weakness of the implementation but of the matter: a procedure that learns from the past has none for what is new.
An unusual route, a company renamed at short notice, an account in another country. All of that occurs in fraud and in honest transactions too. Only one person knows the difference: the one who can explain it — and they explain it to a human, not to a form.
From this follows the rule that applies everywhere on this portal: autonomous is the checking, never the releasing. It is not just an attitude but also the safe side: whoever decides about a sole trader is deciding about a natural person. If that runs exclusively automatically and affects the person significantly, Art. 22 GDPR may apply — then a human must be able to intervene and the decision must be explicable. A pre-check that prepares instead of decides fulfils that by itself; one that decides through has to retrofit it.
What follows from that for liability and the duty to log is set out under Responsibility and limits. And whoever hangs the pre-check on an agent that also purchases writes both into the same mandate — otherwise the fastest award is the one with the thinnest check.
We go through which information can be fetched over interfaces, where the threshold sensibly lies and what has to be logged so that in a damage case it is traceable what was checked and when.