Overnight compliance work, done on a machine that lives in your building. Every conclusion signed by a person, every morning.
In development · shown by invitation
The situation
A generation ago, a handful of small banks led the giants into digital banking — first to the phone, first to the customer where the customer actually was. The lead came from conviction, not headcount.
The back office never got that revolution. Behind every mission-driven bank is a compliance operation built for a different era: alert queues triaged by hand, loan files checked page by page, call reports assembled from spreadsheets, every regulation change absorbed as overtime. The work is honest and endless — and it is the most expensive thing a small bank does that no customer ever sees.
Every hour it consumes is an hour not spent where the bank’s heart is: in the neighborhood, at the branch, in the loan that changes a family’s arithmetic. That is the real cost of the old back office. Not the salaries — the mission displaced.
The machine
smallbanc runs on a sealed appliance that sits on your premises, on your network, behind your locks. Borrower data never leaves the building. There is no cloud tenancy, no per-seat meter, no third party holding what your customers trusted you with.
Inside it, working models handle the procedural passes banking is full of — reading files, comparing them against the rules that govern them, drafting what a person needs in order to decide. The machine works through the night. It produces drafts, never decisions.
Local is not the promise; bounded is. The device can read what it is shown and write only to a queue. It files nothing, sends nothing, touches no payment rail, and holds no credential to your production systems. What it cannot reach, it cannot compromise.
One night, one file
A set of loan files, an alert queue, a reporting period — whatever tomorrow was going to cost the team. The device is shown the documents and the current edition of the rules that govern them.
A model drafts each item. A deterministic checker — no model, no judgment, same result every time — verifies the form: every field present, every date in period, every figure carrying its citation. Then the item is laid beside the governing text itself, claim against source, and the machine stops. It is never asked whether it is right.
The queue is sized to human attention, each item shown beside its source with the differences marked. Approval is one item at a time, by a named person, on the record. Nothing becomes real until someone whose name is on it decides it should.
Every pass records the rule edition it ran against, the model version that drafted, the checks that passed, and the person who signed. When the examiner asks who decided and on what basis, the answer is a record, not a reconstruction.
The work it can hold
Most of a bank’s back office is the same motion repeated: fetch the file, compare it to the rule, note what a person needs to see. Everything below is that motion — drafted overnight, signed in the morning.
Underwriting approval is not on this list, and will not be. The machine spreads, compares, and drafts the memo; approving, pricing, and declining are the banker’s — see the line below.
The line we hold
Credit decisions — approving, pricing, declining. Suspicious-activity determinations. Fair-lending judgments. Adverse action. Anything where the wrong word is an enforcement matter rather than a correction — these are decisions, and the machine drafts, it does not decide. They are named here, in advance, because a tool that claims to do everything should not be trusted with anything.
Declined by design · printed before the first demo, not discovered after the first incident
The vendors will promise full automation. The honest ceiling is the differentiator — and the thing your examiner will read first.
Covenants
On your premises, under your policies. No borrower record ever transits a vendor cloud, trains a vendor model, or sits in a jurisdiction you didn’t choose.
Rules carry effective dates; so does the machine. Every pass cites the edition it ran against. A change in the rules is a new, dated configuration — reviewed and adopted by you, never pushed in the night.
Human review is the architecture, not a checkbox. Queues are sized to attention. Approval is per item. The system is built so that rubber-stamping is harder than reading.
Checking happens against the published rule and the original document — deterministically, or by a person. No model is ever asked to grade its own work.
The audit trail is not a feature added for the exam. It is what the method produces by existing.
Why this shape
The giants will buy this capability from giants, and it will arrive as a cloud, a meter, and a terms-of-service update. A community bank deserves the other architecture: a machine it owns, a method it can read, a ceiling stated in writing, and a bill that doesn’t grow with success.
The banks that led once didn’t lead by outspending anyone. They led by seeing what the customer’s life actually required and building that, plainly. The back office is waiting for the same eye.
The invitation
smallbanc is in development with a small number of institutions whose mission makes the arithmetic matter most. No demo reel, no procurement funnel — a conversation, a file of your choosing, and a morning queue you can judge for yourself.
Write to us