Private Finance Chatbot.
Answers from policies, procedures, reports and client documents in seconds — inside your perimeter, with a source on every answer.
Self-hosted LLM · Hybrid retrieval · Citations by design · Audit log
Financial firms sit on the most answer-rich document bases in business — compliance manuals, product policies, procedures, past reports, client files — and hold the least permission to send any of it to a third-party AI. Per-seat copilots ask you to trust a vendor's data terms. Most compliance teams, reasonably, decline.
So we built the version that doesn't require the trust: a knowledge assistant that deploys entirely inside the firm's perimeter. We build and operate it ourselves as a production-grade reference system, running against a realistic corpus — public regulatory handbooks, public filings, and a synthetic set of internal policies and client documents modelled on real formats. Everything on this page describes the system as it runs, not as we imagine it.
The knowledge is in the building. Finding it is the job nobody was hired for.
McKinsey Global Institute's estimate is the famous one: knowledge workers spend roughly 19% of their week — close to a full day — searching for and gathering information. Gartner's 2023 survey found 47% of digital workers struggle to find the information they need to do their jobs.
In a regulated firm the failure mode is worse than lost time. When the right policy is hard to find, people answer from memory — and memory is not auditable. The alternative on offer, per-seat AI copilots, means routing regulated documents through someone else's cloud under someone else's terms, at $30–60 per seat, per month, forever.
An assistant that answers questions from the firm's own documents and can prove it.
In-perimeter by architecture, not by policy. The full stack — model, retrieval index, application — runs on infrastructure the firm controls: a single GPU node on-prem or a private cloud instance. No document, query or answer transits a third-party API. There is no data-sharing agreement to negotiate because there is no data sharing.
A citation on every answer. Every response links to the exact source passages — document, section, version — that produced it. If retrieval finds no adequate source, the assistant says so instead of improvising. An answer without a source is a bug, and the system treats it as one.
An audit log by default. Every query, every answer, every cited source, append-only. When compliance asks “who was told what, based on which document,” the answer is a query, not an investigation.
Retrieval that respects the org chart. Permission-aware indexing: the assistant only retrieves from documents the asking user is entitled to read. Hybrid retrieval — keyword and semantic — because policy language is exact and embeddings alone miss exact.
| Layer | What runs | Where |
|---|---|---|
| Generation | Self-hosted open-weight model, sized to the corpus — not to a benchmark | In perimeter |
| Retrieval | Hybrid index (BM25 + embeddings), permission-aware | In perimeter |
| Guardrails | Groundedness check: no source, no answer | In perimeter |
| Audit | Append-only log of queries, answers, citations | In perimeter |
| Evals | Hitbox: retrieval hit-rate and groundedness, scored on every release | Build-time |
The time math — industry benchmarks, attributed:
- ~19% of the knowledge-worker week goes to searching and gathering information — McKinsey Global Institute. For a 100-person firm, that is roughly the working time of 19 people.
- 47% of digital workers report struggling to find the information they need — Gartner, 2023.
- The system's job: turn a multi-minute search-and-skim into a sub-minute cited answer.
The cost math. Public and reported per-seat list prices, August 2026 — our cost model. Per-seat pricing scales with headcount. This system's cost scales with infrastructure — which is to say, barely.
| Option | Price shape | 150 seats, 3 years | Where documents live |
|---|---|---|---|
| Microsoft 365 Copilot | $30/user/mo (list) | ~$162,000 | Vendor cloud, vendor terms |
| ChatGPT Enterprise | ~$60/user/mo (reported) | ~$324,000 | Vendor cloud, vendor terms |
| This system | Fixed build fee + your own infrastructure (typically $400–1,500/mo) | Build fee + ~$14–54k infrastructure | Your perimeter |
True today, by design:
- 100% of answers carry source citations — the system refuses unsourced answers instead of risking them
- 100% of queries and answers land in the append-only audit log
- 0 documents, queries or answers leave the perimeter
We publish measured metrics from our own operations as eval cycles complete.
We started with pure vector retrieval. Policy documents punished us for it within a week — compliance language is exact, and embeddings blur exactness. Hybrid retrieval should have been the day-one default; now it is.
Permission-aware indexing went into this build from the start because we'd retrofitted it once before, elsewhere. Retrofitting permissions onto a live index is the most expensive way to build them. Design the access model before the index.
And we burned time on a larger embedding model than the corpus needed. Eval first, size second — the same lesson our GTM agent taught us, learned again.
A knowledge assistant that can’t leak.
- Documents never leave the firm
- Permission-aware retrieval
- A citation on every answer
- Append-only audit log
Your documents already know the answers.
If your firm's knowledge base is large enough to be valuable and regulated enough that it can't leave the building, this architecture exists for you. The Crit maps your corpus, your access model and your break-even against per-seat tools — in two weeks.