The financial industry has been testing LLMs for two years. Most pilots die at the production stage for the same reason: they were treated as chatbots.
A chatbot is an interface. Fund operations are an orchestration problem across sources of truth — fund bylaws, accounting systems, positions, investor communications — under fiduciary liability constraints. It is not the same problem.
Three limitations of a wrapper around an LLM
- 1.Undetected hallucination: a general-purpose model can invent a CVM clause. For an analyst, that is an annoyance. For a fiduciary administrator, it is regulatory risk.
- 2.No traceable source: answers without citations do not survive an audit. Who decided? Based on which document? When did the document change?
- 3.No context-aware permissions: a single bot does not respect RLS, tenant segregation, or mandate boundaries. It sees everything, it says everything.
What changes with a multi-agent architecture
Instead of one model answering anything, a set of specialized agents — each with its own domain, its own tools, and its own sources — coordinates the answer.
- ●A Compliance Agent reads regulation, CVM official letters, and ANBIMA updates.
- ●A Risk Agent queries positions in real time and applies quantitative models.
- ●An Operations Agent calls internal systems and triggers workflows.
- ●An orchestrator decides which agent answers — and when two or more need to be coordinated.
RAG done well is not just vector search
Retrieval-augmented generation (RAG) is where most financial systems fail. Indexing documents into a vector database is the trivial step. The hard part is keeping the index coherent when fund bylaws are versioned, CVM instructions are revoked, and circular letters change interpretations without altering the original rule.
A mature financial RAG needs temporal versioning: 'what was the rule in January 2025?' is a legitimate question — and a different one from 'what is the rule now?'.
Governance as a foundation, not a layer
Permissions, auditing, and traceability are not features you bolt on later. In systems for regulated funds, they are design constraints that shape the architecture from the first line of code.
The Arkar Assistant inherits permissions from the System: if a user cannot see fund X, neither can the agent. Every interaction is logged with context, source, and decision. Logs go to the same audit table that records any human action.
That is the standard separating experimental AI from operational AI in the funds industry.