47 Audit AI · Year-end statements
I built 47 Audit AI to take the mechanical part out of year-end preparation. It reads a trial balance and last year's signed report, maps every account to a statement heading, checks the arithmetic, and only then assembles the report from the figures a reviewer approved. AI understands, code calculates, rules ask, humans approve.
The app sits behind a sign-in and is invite only for now.
Why I built it
Before I built software I worked as a financial auditor. A large share of every engagement was the same mechanical work: mapping a trial balance to statement headings, carrying last year's figures across by hand, and re-checking totals that should never have been typed twice. I wanted a tool that does that part, shows its working, and refuses to finish until the numbers agree.
The application
Redrawn live from the application’s own layout with sample figures, not customer records.
Every engagement on one line: client, year end, the stage it has reached and whether it is approved.
47 Audit AI
Audit projects
4
Created from real source files
Approved statements
1
Explicit reviewer approvals
Reports ready
1
Approved in-app audit reports
Each trial-balance row matched to a statement heading, with the suggestion's source and confidence still on screen.
47 Audit AI
312 of 326 rows mapped · 14 still need a heading · unresolved rows block the next stage
Financial position, profit or loss, equity, cash flow and the PPE note, built from the mapping rather than retyped.
47 Audit AI
Assembled from the approved version only. The tables are bound to the figures that were signed off.
47 Audit AI
Al Noor Trading W.L.L.
Financial statements for the year ended 31 December 2025
How it works
A trial balance in .xlsx, .xls or .csv, and last year's signed report as PDF or DOCX. The reviewer chooses the worksheet, the header row and the amount columns; nothing is picked silently.
Rules and last year's report suggest a heading for every account, and each suggestion keeps its source and confidence. Capital, related parties, loans and tax are never auto-confirmed.
The statements are built together and must agree: balance, profit or loss, cross-foots. A failed control blocks approval, and changing a mapping afterwards voids it.
Approval releases the report stage: a versioned DOCX and PDF with the financial tables locked to the approved figures.
AI understands. Code calculates. Rules ask. Humans approve.
What I decided
A language model reads documents and suggests classifications through schema-validated structured output. Every figure is exact decimal arithmetic (decimal.js, never floating point), and AI output never enters a calculation.
The audit opinion is an explicit choice a person records: unqualified, qualified, adverse, disclaimer or emphasis of matter. Export is refused until it is made. The tool prepares; it does not audit.
The original trial-balance rows are never edited. Adjustments are separate records, so every statement line can be traced back to what the client actually sent.
OCR is slow enough to outrun a serverless function, so a scanned prior-year report is read page by page in the visitor's browser with Web Workers, and the OCR files are served from the app's own origin. A client's document never touches a third-party CDN.
Built withNext.jsTypeScriptSupabase (Postgres, RLS, storage)OpenAI structured outputsdecimal.jsdocx + PDF exportTesseract OCRVitest
Questions about it
I am happy to walk through the design, the data model or the decisions above. Send me a message.
The other two tools