The structural limit of classic RAG
RAG (Retrieval-Augmented Generation) brought large language models into finance teams, but its operating model is primitive: it takes the question, finds the most similar passages in a vector index, stuffs them into the prompt and generates an answer. That works for an internal FAQ. It breaks the moment the question is analytical.
Take a real due diligence on a company listed on Euronext Milan. The question is not "what is EBITDA?" but "how did adjusted margin move over the last three financial years, net of the disclosed non-recurring items, and what do the notes to the accounts say about the reclassifications?". A RAG chatbot pulls three or four semantically close chunks, often from the same document, and composes a plausible answer. The failure is twofold: it does not know it needs to read the half-year report, the consolidated IFRS statements and the results press release together, and it tends to fill gaps with figures that look right but are fabricated. In finance, a plausible wrong number is worse than no answer at all.
What a research agent does differently
An agent, in the spirit of platforms like Hebbia and Rogo, does not merely retrieve: it plans. It decomposes the question into sub-tasks, decides which documents are required, iterates its search across thousands of files and verifies intermediate findings before synthesising. On an M&A data room with twelve thousand documents, the difference is categorical.
A typical Amaril flow on a due diligence request looks like this. First, planning: the agent works out that it needs three years of IFRS statements, key contracts in the data room, term sheets and filings lodged with Consob. Second, targeted iterative search: it queries each source separately, refining its queries based on what it finds rather than stopping at the first match. Third, anchored extraction: every extracted figure stays tied to its source page and paragraph. Fourth, verified synthesis: the agent cross-checks the evidence and flags contradictions (for example a covenant in the credit file that conflicts with what the management presentation claims) instead of smoothing them over.
This is what separates a productivity tool from an analysis tool. RAG answers the question you asked. The agent answers the question you should have asked, and shows you where it looked.
Source citation: the constraint that makes AI usable in finance
No analyst signs an investment memo on numbers they cannot trace. No credit committee approves a summary that says "leverage is high" without pointing to the line in the accounts. That is why Amaril is built around a non-negotiable constraint: every claim, every figure, every cited clause links back to its exact source, page and paragraph, clickable and verifiable.
The consequences are practical. When the agent produces an IC memo or a credit memo, the analyst does not re-read three thousand pages: they check the citations. When a comparable analysis lines up multiples for a panel of European peers (one issuer on Deutsche Borse, one on Euronext Paris, one on Borsa Italiana), each multiple points to the filing it was extracted from, with its euro currency and its reference date. When an earnings summary for a bank speaks to CRR/CRD requirements, or Solvency II metrics for an insurer, the number is anchored, not regenerated.
Above all: the agent is designed to say "this is not in the documents" rather than invent. In a domain governed by MAR on market abuse and MiFID II, a hallucination is not an annoyance, it is a compliance risk. The discipline of abstaining matters more than the fluency of the answer.
Architecture, perimeter and trust
Analytical power is worthless if the data leaves your control perimeter. Amaril runs on EU cloud, with end-to-end encryption and a zero-retention policy: data room documents do not train models and do not persist beyond the session. For firms operating under Banca d'Italia, BaFin, AMF or EBA guidelines, on-premise deployment is available, with the full agentic pipeline inside the client's own infrastructure. GDPR, SFDR and DORA are not boxes ticked after the fact: they are design constraints.
Takeaway
Classic RAG is a decent conversational search engine. For due diligence, comparable analysis and investment or credit memos you need something else: an agent that plans the work, genuinely searches thousands of documents, anchors every number to its source and abstains when the data is not there. A chatbot's fluency impresses in a demo. An agent's traceability holds up in committee, in front of the regulator and at closing. In finance, the source is not a detail: it is the line between an opinion and an analysis.
