The back office that never guesses.

Nien is the agentic back office for regulated finance, built on neurosymbolic agents. Language models read your messy documents and draft prose; deterministic code — your encoded SOPs and the regulatory instruments — computes every figure and takes every action, under your maker-checker matrix. A named officer signs everything that leaves. It runs read-only, inside your environment.

Two halves with a hard boundary between them. The neural half is language models doing what they are provably good at: reading messy documents and drafting prose. The symbolic half is deterministic code — your encoded SOPs, the regulatory arithmetic, the maker-checker gates — doing everything that must be defensible: computing figures, taking actions, keeping the journal. The boundary is architectural, not a policy: the parts that improvise are never the parts that calculate, act or approve. That is why the same input produces the same output, every run, and why every figure has an answer your inspector will accept.

The library is lender-first today because that is where the sharpest deadlines are. But the engine is not lending-specific: grievance response, regulator correspondence and reconciliation have the same shape at a bank, an NBFC, an MFI or an insurer — deadline-bound paperwork, evidence packs, named sign-off. Each segment's books are built the same way, from that segment's own Directions.

Yes — the live demo at demo.nien.ai plays one full weekly reporting cycle for a fictional NBFC, entirely in your browser: four bureaus' rejection files parsed from raw bytes, 26 encoded SOP rules applied, DQI recomputed to the official RBI worked example, and four artifacts drafted for sign-off. Click any figure and it traces to the exact file, line and byte range.

Neither. You already have a board-approved framework and a core stack; Nien does not replace any of it. It is an overlay that does the recurring paperwork those frameworks generate — the corrections, the packs, the responses — so the framework you already report to your board actually runs on time, with evidence.

No. This is the central design rule: a language model never produces a number. Models read messy documents and draft prose. Every figure — a corrected field, a DQI movement, a settlement variance — comes from deterministic code executing your written SOPs, and each one links back to the source line it was computed from. Same input, same output, every run.

Read-only exports from your LMS and access to the document share where the files already land. No write access to any system of record, no core integration project, no new data leaving your perimeter. A typical install is measured in days, not quarters.

Inside your own environment — your VPC or on-premise. Customer data never leaves your perimeter and is never used to train anything. Access is scoped to the minimum the configured books need, and every access is on the append-only journal.

We are working with a small number of institutions as design partners. You bring one messy, deadline-bound workflow and the SOP behind it; we encode the SOP, run it against your real files, and show you the drafts line by line, inside your environment. The engagement runs against measured outcomes — cycles run clean, hours removed, exceptions caught — and the commercial terms are named in the first conversation, not discovered later.

The discovery sprint runs in days and shows accuracy on your own files. Tailoring runs in weeks, on your site. The gate to production is the shadow-run: Nien runs in parallel with your manual process until the numbers match, cycle after cycle — and only then are procedures published, locked and versioned, with your officers signing the outputs. If it cannot match your manual cell in shadow, it does not go live. That is the opposite of a pilot that lives in a sandbox.

Yes, and non-invasively. Nien works from read-only exports and the document shares where your files already land; connectors come from the book library. There is no middleware program, no write access, and no core migration — your LMS, CBS and accounting system stay exactly as they are. That is also why the InfoSec review is short.

No — the engagement is designed to end. The forward-deployed engineer's job is to make themselves unnecessary: assemble the blades, fill the context graph, pass the shadow-run, train your maker-checker team, and rotate off. What stays behind is locked procedures, your ground truth, and software your own team operates — served by a central platform team, not a resident consultant. Each deployment also hardens the library, which is why tailoring gets shorter with every institution we install.

Per role taken over, as a single monthly line item — the way you budget for staff, not per seat or per record. Each assembled role is priced well under the loaded cost of the work it absorbs. We name the number in the first conversation.

No. Nien holds no licences and takes no regulated role. It is software your institution runs, like your LMS or your document management system. Decisions, sign-offs and regulatory accountability remain with your named officers — which is exactly how your inspector expects it.

The encoded rule changes, with a reviewable diff your compliance team signs off on, and every run after it follows the new rule. No retraining program, no vendor change-request queue. This is the practical payoff of keeping the rules legible instead of burying them in a model.