Why advisers need a morning pulse, not another chatbot

Key takeaways

• KritiFin is our AI workspace for financial advisers: it reads the documents an adviser already has and turns them into a ranked morning view, source-cited answers, and pre-meeting briefs.
• The design bet: advisers don't need another chat box. They need to know who to contact, what to review, and what to do next — before the inbox decides for them.
• Language models extract; deterministic rules rank. The priority list is explainable on purpose.
• Every answer is grounded in the workspace's own documents and cited. Anything the AI drafts waits for human approval.

The chatbot trap

Bolt a chatbot onto adviser data and you get a system that is exactly as useful as the questions it gets asked. The adviser has to already suspect that a review is overdue, or that a client mentioned a pension transfer in passing four months ago, before the tool can help. But the work that actually slips — the annual review that quietly passed twelve months, the follow-up promised in a meeting note three weeks back — slips precisely because nobody thought to ask.

So KritiFin starts from the other end. Chat is one capability inside the workspace, not the workspace.

Priorities before questions

Each morning the dashboard ranks what deserves attention: reviews past their twelve-month mark, follow-up commitments that have aged beyond their promise date, and tasks generated by playbooks — annual review, onboarding, protection review — applied across the book.

The ranking is deliberately boring: alerts sort by trigger date, then severity. No model decides what matters today. The language model's job is upstream — reading uploaded fact-finds, meeting notes, and transcripts, and extracting the clients, dates, and commitments that the rules then rank. When a priority surfaces, an adviser (or a compliance officer) can trace exactly why it's there.

That split — models extract, rules decide — is the same instinct as caging the agent inside the workflow: put the probabilistic part where it's strong, and keep the part that must be explainable deterministic.

One upload, two paths

Every document ingested takes two paths from a single upload. Path one is structured: extract the clients, dates, and commitments, and feed the priority engine. Path two is semantic: chunk the text with overlap, embed it, and index it in an org-scoped vector store for retrieval.

The first path powers the morning view; the second powers everything the adviser asks later. One upload, two kinds of leverage — and no separate "please re-upload your documents for search" step.

Answers that show their work

When the adviser does ask, the Copilot grounds its answer in two sources at once: structured client records, and semantic search over the ingested documents — scoped to the whole book or to one client. Excerpts arrive numbered, answers cite them inline, and the sources render alongside the reply.

Below a similarity floor, excerpts are dropped rather than stretched — "I don't have that in your documents" beats confident invention, the same discipline as production RAG that doesn't hallucinate. Retrieved documents are treated as untrusted input: sanitised, delimited, and never handed tool access.

And one more boundary: AI-drafted emails and review notes land in an approval queue backed by an append-only audit log. Nothing goes out on the model's authority.

Boundaries kept on purpose

Tenancy is enforced four times over — database row-level security, org-scoped queries, an org-scoped cache, and org-filtered vector search — so one workspace's documents can never leak into another's answers. Client documents are used to serve that workspace, never to train models.

The workflows are shaped for UK advice: Consumer Duty–style review notes, and a vulnerability-signal scan that flags language for a human to review. Signals for the adviser, not determinations by the machine — in regulated work, that line is the product.

FAQ

Is KritiFin a chatbot? No. Chat is one surface. The core is a priority engine over the adviser's book — reviews due, ageing follow-ups, meeting briefs — with an AI Copilot that answers cited questions when asked.

How are answers grounded? Uploaded fact-finds, notes, and records are indexed into structured data and org-scoped embeddings. Answers cite numbered sources, and low-confidence excerpts are dropped rather than guessed.

Does KritiFin train on client data? No. Documents are used only to serve that workspace, never for model training. AI drafts wait for adviser approval, and every AI action is audit-logged.

How do I try it? kritifin.obliviox.in — new workspaces start free with a bundle of AI credits, no card required.

---

KritiFin is built by oblivioX — engineering intelligent systems. If you're building grounded AI over sensitive documents, the door is open.