Owners do not need another chart. They need to know whether payroll clears, which invoices are late, and which jobs looked profitable only because someone forgot the supplier bill. An AI copilot can draft that explanation. It can also fabricate a confident one. This article is the line Codex Thrill draws for Codex Finance and for any finance assistant built on a client’s books.
The questions worth automating
The useful questions are narrow. What cash is expected to land this month from invoices already sent? Which customers are outside the terms you agreed? Which jobs have costs recorded against them that push the margin under the quote? A copilot that answers those from the ledger saves a Sunday evening. A copilot that “strategises” without the ledger is a chatbot with a finance costume.
Write the ten questions the owner actually asked last month. If the data to answer them is not in the invoices, bills and payouts you can export, do not add a model yet. Add the missing field.
Ground every figure
Every number in the answer should point at a transaction, an invoice or a payout. If the model cannot point, it should say the books do not contain that fact. This is slower to demo than a fluent paragraph, and it is the only version a finance lead will trust twice. Codex Finance is being built around that rule: the copilot explains, and the source sits next to the sentence.
Rounding and currency need the same discipline as a till. A total that does not match the export it came from will end the experiment on day two.
- Cite the invoice, bill or payout
- Say when the data is missing
- Match the export to the cent or halala
- Keep the draft labelled as a draft
What the copilot must not do
It should not move money, approve a payment, or file a tax return. Those are actions with a legal person behind them. It should not email a customer a chase that a human has not read. It should not invent a forecast by extending a line when last quarter included a one-off. A forecast can be offered as a scenario with the assumption written in the first sentence, or not offered at all.
It should not see records the user is not allowed to see. An owner and a bookkeeper are not the same role. The assistant inherits the permission of the person asking.
Where it sits in the stack
The system of record stays the accounting tool or the SQL database you already trust. The copilot is a reader with a careful mouth. NestJS or a similar API can expose the questions. The model never gets a raw connection it can use to update a row. For a wider operating picture — jobs, customers, today’s tasks — a product like AI Business Manager may sit beside the finance copilot. They should share definitions of “customer” and “overdue”, or the owner will get two answers.
Start with one export and one question. “Which open invoices are older than your payment terms?” is a better first release than a general chat box over the whole drive.
Review, retention and a person
Someone who understands the books should sample answers every week at the start: ten questions, checked back to the source. Keep the prompt, the records retrieved, and the answer for long enough to investigate a bad one, and no longer than your retention policy allows. Finance data is not a corpus you casually send to an unmanaged endpoint. Know the region, the subprocessors and whether the vendor trains on your prompts. If you cannot answer that, you are not ready to connect the ledger.
The win is a shorter path from “I think we are fine” to the three invoices that say otherwise. It is not a synthetic finance officer.
Conclusion
Use AI in small-business finance to explain records you already have, with the transaction in view and a person still approving anything that moves money or contacts a customer. That is the scope of Codex Finance while it is in development. Anything broader is a different product and a different risk review.
Frequently asked questions
Can an AI copilot file VAT or income tax?
It should not. Tax filing stays with the person and the accountant. The copilot can only explain the records it was given.
What is a safe first question to automate?
List open invoices older than the company’s own payment terms, each linked to the invoice. It is easy to check and immediately useful.
Should the model connect directly to the database?
No. Give it a read API that returns the rows the question needs. Do not give it a connection that can update the ledger.
Is Codex Finance available to install?
No. Codex Finance is in development. This article describes the rules it is being built to follow.
