Frequently asked questions.

SableAlpha builds Augustine, payment controls for the finance teams paying the most bills. Written rules check every proposed payment against your books before it goes out, an authorized person approves it, and every check and approval is kept on the record. A look back over last year's bills finds money already lost: bills entered twice, payments made twice, credits never used.

The finance team that pays the bills: the controller, CFO or accounts payable lead at a company that pays many vendors and keeps its books in QuickBooks Online or Xero.

Payment mistakes and fake bank-change emails are usually found at month end, if at all, and AI is starting to draft payment runs at small companies with nobody independent checking it. The controls that larger companies buy are priced and staffed for them, not for a company with one bookkeeper.

It covers the controls part: every payment checked against written rules, an authorized approver, and a record of who decided what and why, which is what an auditor asks to see. It does not track licenses, renewals or changes in regulation.

Built: nine rules that check each line of a payment run, five checks that look back over a year of bills and payments, a review sheet added to an Excel workbook, one approved Slack message, one approved comment on a Jira ticket, reading QuickBooks Online and Xero file exports, and documents you upload, indexed separately for each company. Live connections to accounting systems are not open yet, and no real company's books have run through Augustine yet.

A chatbot gets a pasted question, writes paragraphs, and is never checked. Augustine uses a commercial model of that kind to write the draft, and what SableAlpha built is everything around it: picking the exact record, the checklist of facts, the rule check against your books, the approval on every change, and the log. The model can be swapped.

Keep it: those give your team a capable model with company controls, and with connectors they can read your tools. They do not check an answer against your books with written rules before a change is made, and they do not hold that change for an authorized approver and log both. Augustine adds that check.

Rules written by people in plain code, which read the checklist of facts and your records, give the same verdict on the same input every time, and can be read and corrected by a person.

A figure the rules cannot find in your records is blocked, and what cannot be fixed goes to a person with the reason attached.

It holds a payment inside its own review: a held payment stays out of the approved list it hands back to whoever releases payments. It does not connect to your bank, and it cannot stop a payment someone makes directly in QuickBooks, Xero or the bank.

As current as the export you give it. Every run shows the date the export was made, and a bill paid after that still looks unpaid, so export again before approving if payments went out since. Checks run when you bring a payment run or an export; it is not continuous monitoring.

It needs the vendor bank-detail change history, from your accounting system's vendor change log or audit trail; a bills export alone does not carry it. Without that history the bank check does not run and says so, and the payment is not marked as passed.

It is shown as not checked, with the reason, and never counted as passed. Every run shows how many records were imported, checked, excluded and still unresolved.

From your own records: every recommendation shows the figures and the records it rests on.

We gave the same AI model 510 practice finance questions in September 2026, once on its own and once inside Augustine. On its own it got about 1 in 4 right (24%), and with Augustine more than 8 in 10 (85%). It went to make a payment or change a record on a wrong answer 112 times on its own and 37 times with Augustine, which is down by two thirds and not zero. AI is billed by how much text it reads and writes, and each right answer took less than half as much, while the cost of asking a question stayed about the same. This is practice data, and no real company's books have run through Augustine yet.

Ten hosted AI models answered the same 510 practice questions, among them Claude Sonnet 5, Gemini 3.8 Flash, GPT-6 Astra and Claude Fable 5.1, and every one got more answers right and made fewer wrong payments and edits inside Augustine than on its own. Each result compares a model with itself, so it says nothing about one model against another, and we make no such claim.

Neither: it checks payments and flags what it finds, an authorized person approves every change, and it gives no investment or market advice.

Book a demo: a founder walks you through Augustine on practice data, shows the numbers behind each recommendation, and answers your questions.

A QuickBooks or Xero export, a few documents, and two hours over two weeks from the person who pays the bills.

Fourteen days, ending on the stated date or earlier whenever you ask, after which we remove your documents by hand and confirm in writing when it is done.

Pricing is not set: we are recruiting design partners and agree terms with each one in writing. Ask for a custom quote on the pricing page. There are no public plans, accounts, or checkout on this site, and no fees are charged through it.

The support page covers setup and connections, data and privacy requests, and security reports, and a named person replies to every request.

Into a workspace for your company, and connection secrets never reach the browser. The security page sets out who can open it and what we keep. You own your data and the outputs, as the terms and the privacy policy set out.

No, we never train models on your data. Each AI model provider handles a request under its own data terms, as the security page explains.

The look-back reads it in your browser and never uploads it. We keep only the run's totals and the status of each finding, so you can see what was found and what was recovered. If you connect QuickBooks Online instead, we read the transactions for that one check and do not store them.

No, nothing is written to a system of yours without an authorized person approving it, and each decision is recorded with the person's name, the time and any reason given.

Good-faith reports are welcome by email as described on the security page, and no bounty is promised.