NoumenaiFOR FINANCE WORK

Noumenai engineers AI that does real finance work.

The work you can't afford to get wrong — done, evidenced, and ready for audit.

one defensible number

Finance is not like the rest of the enterprise.

Everywhere else, an AI assistant can be roughly right and still useful. In finance it cannot. Every number has to tie to a source. Every entry has to survive an audit. Every output carries the name of the person who signed it. A plausible-sounding answer that is wrong is not a small failure here — it is the failure.

A

Looks right, isn't right.

The number that reads plausibly and the number that ties to source are not the same number. Advisory AI can afford the gap. A signed one cannot.

B

No trail, no trust.

If the work cannot show its evidence — cited, reversible, attributable — it cannot be posted, filed, or defended. Provenance is a first-class output, not a log line.

The Method

02Two agents, one gate

One method, under everything we build.

Whatever the process, the agent works the same way. It never acts on trust. It acts on evidence, and only after a check.

Two agents. Two judgments. One gate.
Row 01 · Sources
INVOICE
SUPPORTING DOCS
SYSTEM OF RECORD
FIRM HISTORY
Evidence ↓
Row 02 · The Split
Agent 01 · proposes
— Accounting —
ACCOUNTPROFESSIONAL FEES · expense
TAXINPUT VAT · deductible
CONFIDENCE96%
BASIS247 prior entries · FIRM HISTORY
— Analytical —
COST CENTREOPERATIONS   70%
COST CENTRESALES   30%
CONFIDENCE91%
BASISthe firm's own convention
Agent 02 · re-derives independently
— Accounting —
ACCOUNTPROFESSIONAL FEES · expense ✓ agrees
TAXINPUT VAT · deductible ✓ agrees
CONFIDENCE95%
— Analytical —
COST CENTREOPERATIONS   100% ✗ split differs
CONFIDENCE64%
BASISengagement letter · SUPPORTING DOCS
Row 03 · The Gate
Gate · Deterministic
ACCOUNTINGagreement   →   96% · ELIGIBLE
ANALYTICALconflict   →   discounted · HUMAN REVIEW
The entry inherits its weakest element
→ Accounting cleared · Allocation raised for review
Row 04 · The Human
Accountant confirms the split → posted, with its evidence and its trail

An example: a supplier invoice, classified and allocated. The method is the same whether the work is a reconciliation, a VAT return, or a cash forecast — propose, check, confirm, then act.

IN PRODUCTION · ANONYMISED DATA

A real supplier invoice, cross-referenced against two supporting documents, composed into a single proposed entry. Anonymised data.

01

Merges evidence across documents

Three sources, one entry.

02

Confidence per field

Not blind extraction.

03

Understands the document

Recognises the transaction type, the tariffs, the sessions.

04

Reasons from the rules

Splits across sub-accounts by acquisition value, per the tax code. Not a rule-table.

05

Institutional memory

Supplier confirmed against 247 prior entries by the firm itself.

06

Knows what it doesn't know

Flags partial deductibility for human verification rather than assuming.

07

Analytical allocation

Splits the net across cost centres, ready.

Into the Ledger

03In production

Past the demo. Into the ledger.

Noumenai systems are running in production today. They read live systems of record. They propose work against real source documents — merging evidence across several where they must. A second agent independently re-derives the answer and checks it. A person confirms. Then they write back to the systems of record, with the evidence tied to them and the trail intact.

Most AI in finance
Noumenai
drafts and suggests
posts the entry
summarises
files the return, writes back
stops at the demo
into the ledger, with the trail intact

ERP-agnostic by architecture — every system the agent touches connects through one governed tool layer. Live today against a mid-market European ERP.

See the range of work it carries

Sooner

04Time-to-knowing

Not just correct. Sooner.

A CFO does not lie awake worrying about journal entries. They lie awake about not knowing something soon enough — the margin quietly eroding, the cash gap in week nine, the provision that has been wrong for a quarter.

The agent doesn't out-think your team. It compresses the time between a problem existing and you knowing about it — from months to days. Not smarter. Sooner.

See it interrogate the numbers

247
prior entries behind one classification
96%
confidence, composed not declared
62%
of allocation patterns resolve to one centre
1,900
allocation patterns discovered in one client's history

Anonymised data from real client ledgers.

The Range

05Coloured honestly

Eight process families. Coloured honestly — proven, in production, or buildable.

PROVEN

Record to Report

IN PRODUCTION

Order to Cash

IN PRODUCTION

Procure to Pay

IN PRODUCTION

Treasury & Cash

IN PRODUCTION

Planning & Analysis

IN PRODUCTION

Tax & Compliance

BUILDABLE

Investment & Funds

BUILDABLE

Controls, Audit & Risk

Earned Trust

Trust ladder

Autonomy is earned, never switched on.

Systems climb a trust ladder — Shadow, Eligible, Human-enabled, Auto-post. Promotion is measured on how often a human accepts a proposal unchanged. Demotion is automatic on any correction.

How trust is earned

If a finance process is costing you people, time, or confidence — and you need whatever replaces it to be defensible — that is the conversation we want to have.

Tell us the process. We'll tell you honestly whether an agent can carry it yet.