The Q&A at AWS Builder Loft started around 7 PM. Three demos, a panel of senior practitioners from Visa, Nium, and a large EV company — two hours of thoughtful, careful conversation about where fintech AI actually is, as distinct from where it is being marketed to be.
The room was small. Maybe twenty people, mostly men — engineers, founders, a handful of execs. The panel was three men. Every hand that went up in the Q&A belonged to a man. This is not an observation about the state of diversity in fintech AI. It is just the room.
Then the microphone went to the audience.
The questions were good. Better than the demos, in some ways. Technical, specific, practically motivated — coming from people who had clearly been living with these problems for months, not reading about them in newsletters. And they all pointed, in various formats, at the same thing: how do you actually handle hallucinations in production? What are your internal approaches? What's the secret sauce?
The phrasing changed. The question didn't.
The Senior Director at Visa handled it in two moves. First, a redirect: a personal anecdote about using AI to generate a product requirements document, outside of work, nothing to do with Visa. Technically responsive. Completely beside the point. The audience didn't register it as a sidestep. So they pressed again — and kept pressing, in different forms, for the rest of the Q&A.
When the explicit refusal finally came, it sounded something like: "I'm not comfortable having that conversation about the hallucinations we see at Visa."
Nobody paused. Another hand went up. Another version of the same question.
There are no answers to those questions in this room. Not because nobody knows them. Because nobody can say them here.
What changed was the room.
Mixed company is a social register phrase. It signals that the composition of the room changes what can be said. This is not a cynical observation. It is a structural one.
A public fintech AI panel contains, simultaneously: people who compete with the panelists, people who might invest in them, people who might acquire them, people who report to regulators, people who might become journalists, and people who have no idea who anyone else in the room is. The same person who will tell you everything over a private dinner will tell you almost nothing on stage at a public event. The information did not change. The room did.
Four filters run in parallel in anyone who has been doing this long enough. Competitive: what Visa does internally to manage hallucinations in production is a competitive asset — sharing it is not in their interest. Regulatory: how a regulated institution manages AI risk is subject to examiner scrutiny, and discussing it publicly creates exposure. Reputational: acknowledging specific failure modes in a room with an on-the-record transcript is a risk management problem. Relational: every partner, investor, and potential acquirer in the room is listening to everything.
The Senior Director at Visa was running all four filters simultaneously. The founders in the audience didn't seem to have them active yet. What's more interesting is that some may have understood the sensitivity and asked anyway -- which points at something different. They were doing what people do when they don't yet know the code: asking directly, in good faith, expecting an honest answer.
What the wall is made of.
The founders pressing the Visa Senior Director on hallucinations weren't being aggressive. They were navigating as though they were in an open field — walk toward what you want, ask for what you need, expect to receive it or be told plainly why you can't.
They were in a maze.
Walls in a maze are not obstacles. They are structural. They don't mean you've done something wrong or that the person on the other side is being evasive. They mean this particular corridor does not go through. The adult in the room doesn't experience a wall as a dead end. They experience it as direction.
Notice the sequence. The first move was a redirect — a personal anecdote about using AI to draft a product requirements document, nothing to do with Visa. Technically responsive. Safely off the record of anything Visa-specific. This reads as: I can't go this way, let me show you what I can. The audience didn't read it that way. They pressed again, and again, and eventually the explicit refusal came: not comfortable having that conversation here.
Two moves, not one. The practiced version of this starts soft because an explicit refusal is a last resort, not a first response. You redirect to what can be said and hope the room reads it. When the room doesn't and keeps pressing, you name the wall. Still without hostility. Still without making the questioner feel foolish. But the wall gets named.
What does that wall tell you? The refusal suggests the issue is live, sensitive, or competitively important. The regulatory exposure makes public discussion inadvisable. The answer may exist — it is just not in this corridor.
The founders in the audience experienced the explicit refusal as a slightly awkward non-answer. They had already missed the first move entirely.
The money line.
Same room. Same words. Different positions in the maze.
Three terms from the evening: hallucination, human in the loop, trust.
To a founder building a fintech AI product, hallucination is a technical problem. An open one — nobody in that room claimed to have solved it — but a technical problem nonetheless. You mitigate it. You run shadow systems and parallel validation layers and human checkpoints. You treat it as an engineering challenge, the kind that feels like it should eventually yield to enough clever effort. Whether it actually will is still unknown.
To the Senior Director at Visa, hallucination is a liability exposure. A regulatory conversation. Not shareable in public. Not now, possibly not ever. Which is precisely why, when pressed on it directly, he was not comfortable having that conversation here.
The Nium CTO named it most directly — not in the rapid-fire segment, but earlier, when someone asked how long until humans weren't needed in the loop at all. His answer came out of something specific: agents had been used internally to close transaction monitoring alerts. They closed too many by mistake. The team had to go back and reopen them. If a regulator came asking, they would have to account for it. That is the version of the hallucination problem that lives at a regulated institution. Not an engineering challenge. A liability already incurred.
Then, in the rapid-fire segment near the end, when asked what subject line would get him to respond to a cold email — the question was not quite "tell us what product to build so we can sell it to you," but it was not far from it — his answer was:
"Somebody can provide me certainty for probabilistic agents. Yes, I would definitely respond to it."
Not a hint about acquisition targets. Not a product roadmap. A statement of his actual unsolved problem, framed as an invitation rather than a requirement. He took a fishing question and turned it into something real.
To a founder in the audience hearing that: an interesting market insight about an unsolved problem.
To the practitioners on stage: a description of the thing keeping them up at night, finally said out loud in a form that was safe to say.
They were not having the same conversation, even though they were using the same words.
The vocabulary gap is not a communication failure. It is a diagnostic signal. When the same word carries completely different weight depending on who is using it, you are looking at an asymmetry in experience. The person for whom hallucination carries regulatory weight has been somewhere the other person hasn't been yet. That asymmetry is visible in every deflection, every careful hedge, every "it's going to take quite a while." The founders asking the direct questions haven't crossed that line. The panelists crossed it a long time ago and stopped trying to explain what's on the other side of it.
The question underneath the question.
The Nium CTO's answer is the model — not because the question was good, but because of what he did with it.
The audience was fishing all evening. Asking panelists who are Senior Directors and CTOs at regulated financial institutions to reveal their internal approaches to problems they haven't solved yet. Asking, in effect, for the moat. And when that didn't land, asking in a different format. And when that didn't land, trying again.
There is a simpler way to say what was happening. These are not students. These are adults raising money, signing contracts, building companies in regulated industries. The room has an architecture, and that architecture is not hidden. It follows a logic that anyone willing to pay close enough attention can learn to read. In a setting where people have competing interests in what gets said, the direct question -- what is your secret sauce? -- is not bold. It signals that you have not yet read the room. The code is not secret. It is just not handed to you. It rewards attention.
Asking someone for their secret sauce in public is not bold. It is not scrappy. It signals that you have not yet understood the game being played. And that signal is visible to everyone in the room who has.
The Nium CTO's answer worked because he didn't answer the question that was asked. He answered the question underneath it — what do I actually need, stated as a problem, open to whoever has the solution. That is the version of the question that belongs in this room. It does not demand. It does not fish. It goes around the wall instead of into it.
The intelligence you came for is not on stage. It never was. It is in the conversation that happens after the room changes — in the hallway, in a follow-up email, in a private meeting where the composition is different and the four filters don't all apply at once.
The wrong read.
This is what it looks like from where VCs sit.
The portfolio founders attending panels like this one are doing exactly what they should be doing: getting into rooms, listening carefully, trying to learn how the industry actually works. The effort is real. The intent is right.
And some of them are leaving with the wrong read.
The tell is subtle. It shows up in how they describe the event afterward — frustrated that nobody really answered the good questions, convinced that the practitioners are holding something back. That frustration is a signal. It means the founder is still navigating as though the space is open, still expecting the stage to return what only the hallway can give.
The read that actually fits: public panels return general principles, relationship surfaces, and market orientation. They identify who in the room is sophisticated enough to have a private conversation with afterward. The real intelligence lives in that conversation. The panel is not where information lives. It is where you find out who has it.
Most VCs already know this. It is part of how they operate. Whether their portfolio founders have the same read is a different question. And whether anyone has given it to them explicitly is, more often than not, the answer.
The code is not secret.
It is just not written down. It is transmitted the way all institutional knowledge is: by being in enough rooms to learn their architecture, by asking the question that hits a wall and then figuring out why, by watching the adult in the room and noticing where they don't go.
The adult in the room is not guarding the answers. They just know which corridors go through.
Simon Carney is Founder and Principal UX Strategist at Consortium UX, a boutique UX strategy consultancy serving VC-backed founders in complex, regulated markets. He has spent thirty years diagnosing the gap between what founders think is wrong and what's actually wrong. This is Part 1 of The Performance Trap series.
Comments