I wrote a while ago that what our clients were waiting for was AI on the edge. Models running on the institution’s own infrastructure, on the institution’s own data, with nothing leaving the building. That is what unblocked the conversations that InfoSec and Legal had been sitting on for two years, and it is still true.
It was also half the answer. Here is the other half.
An alert has two sides
Take an incoming payment. The beneficiary is your customer. You have the account, the history, the KYC file, years of behaviour. You can reason about that side of the transaction because you own the evidence for it.
The party who sent the money is a name on a message from another bank. You hold nothing on them at all. No account, no history, no file. And in a great many cases, that is the side the decision actually turns on.
Edge AI settles the first side completely. It cannot settle the second, and not because the models are not good enough. It cannot settle the second because the evidence for the second side was never inside the bank. It is on the public record, outside, where a system running on your own infrastructure has nothing to read.
One architecture cannot do both jobs
This is where the compromise usually gets made, quietly, in one direction or the other.
Keep everything inside the perimeter and you give up the frontier: the strongest models are not the ones you can host, and the public record is not something you can bring in-house. Send everything out to a hosted service and you have solved the second half by exporting the bank’s own data, which is precisely the thing nobody was willing to do in the first place.
Neither trade is acceptable to a compliance function that has to defend its decisions. So we stopped asking one architecture to do both jobs.
Two tools, split by where the evidence lives
The EyesClear Platform runs on your infrastructure, inside your perimeter. It consolidates every alert source into one queue, assembles the evidence, drafts the narrative, and returns the signed outcome to the system that raised the alert. Every inference on your data happens on your hardware. Nothing is exported. That has not changed and will not.
EyesClear Investigations does the other half. It screens a company or a person against the public records using frontier models, and returns a report in which every finding carries the quote and the source it came from. What leaves the bank is the name being screened and the search terms derived from it — never a customer file, never a transaction, never an alert. We state that plainly on the security page rather than rounding it down to “nothing leaves”, because the precision is what procurement is reading for.
The split is an engineering decision about where evidence lives. It is not something the compliance team has to operate. Both sides arrive in the same case file, traceable to source, and one analyst signs it.
What changed in the room
The change in the sales conversation was sharper than the change in the product.
- Meetings that ran an hour and a half now run fifteen minutes.
- Onboarding has moved the same way, for the same reason: there is less to argue about before anything can start.
- GPU and cost objections that used to stall a deal for weeks are now a ten-minute conversation.
None of that is because we got better at explaining. It is because the architecture answers the objection before it is raised. When the question is “where does our data go”, an answer that requires no trust and no contract clause ends the discussion faster than any deck.
The workload is not going anywhere
I would rather be writing this in a calmer decade.
The OCC is easing some of the anti-money-laundering burden. Almost everywhere else, the sanctions perimeter is widening — and it widens a little further every time the world divides a little further. One of those trends does not cancel the other. A compliance function does not get to reallocate the capacity freed on one side to the side that is growing, because the growing side needs different evidence, gathered faster, on names the institution has never seen before.
Technology is moving faster than any of us. Financial institutions are picking it up faster than they ever have, and it is still not fast enough. That gap — between how quickly the tools improve and how quickly a regulated institution can adopt them — is the whole job now. It is why we care so much about the fifteen minutes.
Who signs, and what that is worth
None of this moves the decision. The AI is the maker; the analyst is the checker. Every output traces back to source data and needs a person’s approval before it counts, on both sides of the case. On the public-record side that discipline is stricter, not looser: a quote is checked against the page it came from before the finding is kept, anything that cannot be located in its source is discarded, and where nothing is found the report says nothing was identified in the sources searched — never that the subject is clean. How a report is produced is published in full, because a method you cannot read is not a method you can defend.
That is also the honest answer to the comparison question. Every vendor has AI narratives now. The things worth comparing are where the AI runs, who decides, and whether data leaves the bank — and, increasingly, whether a platform can reach the half of the case the institution does not own.
Where the rest of this is written down
The Platform is the longer conversation, and it is a fifteen-minute one now. The services that sit on it and the questions procurement actually asks are both published, as is what nine years in production looks like. None of it is a deck. If a claim on any of those pages cannot be checked, it should not be there.
Thank you to the clients who stood behind the version of this that was harder to explain, and to a team that kept building for the moment the objection would finally have an answer.
Best, Erkin