AURELIA FSSC · PRODUCT OVERVIEW

AI does more of the finance work. Nothing it does becomes harder to overturn.

Aurelia FSSC is an AI-native financial shared services platform built for the two kinds of company whose finance breaks first: AI and model-API businesses, and quantitative funds. It runs the record-to-report close, reconciles counterparty statements, keeps an immutable ledger, forecasts cash, and supports spend and pricing decisions — as an outsourced operating platform rather than a reporting tool.

Open the live console →

Models propose. A deterministic kernel decides.

Most AI finance software makes the same bet: let the model decide, then audit afterwards. In accounting that bet is backwards — by the time an error is found it has already flowed into a reported figure, and the cost of discovering it late is far higher than the cost of stopping it early.

Aurelia inverts it. A model may classify, extract, match and draft; it may never adjudicate. Its proposal becomes an ordinary command carrying the same validation, the same role checks and the same audit chain as a person's, and a deterministic policy kernel rules on every side effect before it happens. Payment release, bank detail changes, material postings, period sign-off and tax submission stay with configured human authority — permanently, by construction rather than by policy document.

This is not a stance in a deck. It is a boundary in the code: the proposer has no write path, data above a configured classification never leaves the deployment, and every authorisation decision is hash-chained so any reader can recompute the whole trail.

What it answers

对账与归因 · RECONCILIATION

When the business system and the ledger disagree

A counterparty statement — supplier or customer — is matched deterministically against our own records, first by reference and then by amount. Whatever fails to match becomes a set of signed differences that conserve exactly: the differences always sum to our total minus theirs, which is what makes the residue trustworthy rather than a rounding story. A model then proposes a reason code for each unexplained difference, a person rules on it, and every journal can open its full chain back through the command that authorised it, the policy verdict that permitted it, and the document or lease it came from.

References that cannot be resolved are reported as unlinked rather than invented.

利润 · PROFITABILITY

Revenue is up and margin is down — say why

Ledger legs carry registry-governed channel, product-line and customer dimensions, so gross margin is computed per dimension from the same postings that feed the trial balance. When a dimension inverts — revenue rising while margin falls — the platform decomposes the change into volume, price and cost effects that sum to the margin delta exactly, with no residual bucket to hide an approximation. The attribution of a primary driver is where a model contributes: it proposes, a controller confirms, and the score is derived from what the human decided.

Period costs are excluded by account classification, so a channel margin stays a gross margin rather than drifting into an opex report.

预算 · BUDGET CONTROL

A budget that is a gate, not a month-end report

Budgets are master data scoped by book, business unit, account and period, and their consumption is derived from real postings rather than a parallel spreadsheet. A spend event naming a budgeted account is evaluated at the moment it arrives: if the budget is exhausted, the event comes back requiring approval, with the budget identifier and the remaining amount attached as evidence for whoever approves it.

Spend into a channel already selling below its margin floor trips a second, independent control.

资金 · TREASURY

Cash, rolled forward day by day

A certified opening balance is rolled across a 7, 30 or 90 day horizon over dated flows drawn from frozen lease amortisation schedules and funding maturities. Each day's closing balance is the previous day's plus that day's net, and the response names the record every input came from so the whole curve can be recomputed by the reader. Flows in another currency are excluded and said so — never silently converted at a rate nobody chose.

Rescheduling a payment queue is reversible and permitted; moving funds always requires two people.

定价与投放 · DECISIONS

Saying no to a price cut, with the arithmetic attached

A spend proposal is assembled by the server from the channel's measured margin, the governing budget's remaining balance and the marginal contribution the spend would produce; a second identity approves or rejects it, and the proposer cannot approve their own. A pricing probe answers a proposed discount across elasticity scenarios and reports the volume increase needed to break even — or refuses to answer at all when the account carries no unit counts, because unit economics without units would be invented.

A negative contribution is information for the decider, never an automatic refusal.

How it is built

Trust: Every request resolves its tenant from the hostname, its identity from the deployment's directory, and its scope from a tenant, legal entity and accounting book trio. Roles come from the server, never the browser. Each authorisation decision, including every refusal, is persisted and hash-chained.
Command pipeline: Writes are accepted as commands with mandatory idempotency keys and materialised by handlers that enforce segregation of duties, legal state transitions and arithmetic re-proof. A packet computed by a route is recomputed before it lands, so a tampered body is refused rather than trusted.
Ledger: Double-entry with exact eight-decimal integer arithmetic, dual-currency conservation, and correction only by reversal — original entries are never mutated. Four deterministic controls stand between a draft and a posting: the chart of accounts, cost centres, dimension registration and the open period.
Business domains: Lease accounting under ASC 842, document intelligence, reconciliation, close orchestration with dependency-aware tasks and SLA-driven service cases, profitability, budget, treasury and decision support — each posting through the same controlled path a person's entry takes.
Skills: A skill is a control boundary plus a replaceable proposer. Three exist today — reconciliation break classification, invoice field extraction and margin attribution — sharing one template: the proposer has no write path, its proposal becomes an ordinary command, and accuracy is derived at read time from the human ruling rather than self-reported.

中文摘要

Aurelia FSSC 是面向大模型公司与量化基金的 AI 原生财务共享服务（FSSC）运营平台。它承接月末关账、往来对账、总账与凭证、维度毛利分析、预算控制、资金预测与投放定价决策，把财务共享中心从"报表工具"变成"可审计的运营平台"。

核心主张是一条写进代码的边界：模型只能提议，裁决永远由确定性内核做出。AI 负责可被人复核打分的判断——对账差异归因、发票字段提取、毛利驱动归因；付款放行、银行信息变更、重大分录、关账签核与税务申报永远需要人的授权。每一次放行与拒绝都进入哈希链证据，任何人可以重新算一遍。

典型问题的答案：业务系统与财务系统对不上时，用守恒的差异集加全链溯源定位；收入涨而利润降时，把变动精确拆成量、价、成本三个效应；业务要加大投放时，用真实毛利快照与边际贡献给出带算式的判断。

Where the boundaries honestly are

The platform states its own maturity rather than implying it. The close, ledger, lease subledger, reconciliation, profitability, budget, treasury, decision support and evidence layers are live and covered by an automated test suite that runs on every change.

Procure-to-pay and order-to-cash exist as working slices — statement import, payment planning, invoice-to-journal — rather than complete subledgers; the full order, matching and terms models are declared roadmap, not shipped capability. Domain packs for marketplace settlement, fund NAV and usage-based billing publish their contracts and activation gates while the console labels them planned. A platform that overstates what it has built is a poor choice for the one function whose job is to be exact.

See it running →