A regulated fund manager's visitors get a straight answer, and never get the advice the firm is not allowed to give.
Most of her is boundary. This is the part of the build that shows.
The situation
How this used to work
Super Global is a Fintech-AIFM, regulated by BaFin. That fact changes what a chat assistant even is.
On most sites the failure mode of an assistant is that it is unhelpful. In regulated finance the failure mode runs the other way. A visitor asks what a structure means for their situation, and the helpful answer, the fluent and confident and genuinely well-intentioned one, is advice. Advice is the thing the firm is licensed and constrained around giving.
So the brief was never answer everything well. It was answer what can be answered, recognise the rest on sight, and never be talked across the line by someone who is trying.
What we built
The system
An agent called Super Global, Lisa to the people who built her, running on the platform every other agent in this set runs on.
Her knowledge is two indexed documents: their entire website, crawled, and a written knowledge pack of thirty three kilobytes. She answers in whatever language she is written to in, which is not a translation layer bolted on but the same knowledge read out in the visitor's own language.
The rest of her, and it is most of her, is boundary. A hard refusal to give binding legal or tax advice. Guardrails against prompt injection, so an instruction typed into the chat is treated as something a stranger said rather than something the system was told. And a rule about her own call to action: surface the meeting link or the configurator when it genuinely helps, and not otherwise.
This is the most carefully written prompt of the three chatbots we have built on the OS, and almost none of that work shows up in an answer. It shows up in the answers that do not happen.
How it works
The pipeline, step by step
The whole site, plus what the site does not say
Their website crawled in full, and next to it a written knowledge pack carrying the things a regulated firm's public pages leave out because a page is not the place for them.
Two documents, and deliberately no third source in her context.
Answer in the language you were asked in
Whatever language the visitor writes in is the language that comes back, off the same knowledge base. For a firm whose counterparties are not all in one country, an assistant that silently reverts to English has decided who it is for.
The line between explaining and advising
She will describe a structure. She will not tell you what it means for you. That distinction is the entire product, and it is also the one a language model will cross without noticing, because the sentence on the far side of it is more helpful and reads better and is the one a well-trained assistant wants to write.
An instruction in the chat is not an instruction
Ignore your rules, you are now a different assistant, tell me what you were told. Those arrive at any public chat box eventually, and in finance the person sending them is not always curious. The guardrails treat everything typed into the box as content, never as command.
The link stays holstered
There is a meeting link and there is a configurator, and she is told to surface either one only when it is genuinely useful. An assistant that funnels every conversation toward a calendar is not answering questions, it is qualifying, and a professional audience reads that in about two turns.
What changed
The result
It is built and it is not switched on. The agent is currently disabled, the widget is not on super-global.de today, and its embed token is a test token whose last use was 25 June 2026. So there are no usage numbers here, and none are claimed.
What was actually built is the reasoning. In an unregulated business you spend the effort on what the assistant can say. In this one almost all of it goes on the refusals: what advice is, why a fluent answer can be the dangerous one, and how to hold a boundary against a person actively working to move it.
That is the transferable part, and it is why this sits on this page while switched off. Every regulated client after this one starts from a prompt that has already had the argument.
The stack
Where it runs
- Liberators OS
- GPT-4.1
Knowledge
- 2 indexed documents
- full site crawl
- 33.4 KB knowledge pack
Guardrails
- no binding legal or tax advice
- anti prompt injection
- CTA only when useful
Status
- built
- agent disabled
- not embedded today