Finance & Accounting with AI Copilot

The problem

Language models are fluent even about numbers they have never seen. In accounting, one invented figure in a well-written paragraph is worse than no answer at all — and an AI that can post entries is an AI that can move the books.

What I built

• A deterministic general ledger: multi-company with role-based access, chart of accounts, fiscal calendar, journal lifecycle (draft → approve → post → reverse), opening balances, multi-currency foundations and an append-only audit trail.
• Trial balance, journal, general ledger and account-statement reports; receivables and payables with ageing; cash and bank reconciliation; budgets with budget-versus-actual; deterministic cash-flow forecasting and management dashboards.
• A Finance Copilot whose numbers come from typed tools over the same reporting engine the screens use, while a deterministic claim validator checks its narrative — one regeneration, then the narrative is withheld and the reader gets the table.
• The AI can only propose a draft entry; a person confirms it, and approving or posting do not even exist in the AI layer.
• Preview-then-confirm imports from CSV/XLSX files, SQL and HTTP sources, and an Azure Document Intelligence invoice-intake adapter — every path ends at a draft.
• A SQL Server test gate that must execute in CI rather than skip: 165 engine-specific tests, including concurrency and locking.

Architecture

1. • Blazor screens

     28 routes · Turkish and English

   • Finance Copilot

     Azure OpenAI function calling

2. • Application services

     every call authorized against the active company

   • Typed tools + fact ledger

     a claim validator checks the narrative

3. • Ledger engine

     journals · reports · open items · bank · budgets · forecasts

4. • SQL Server

     EF Core migrations · append-only audit

Key decisions

• Numbers from tools, words from the model, verification from code

  A narrative whose figures cannot be traced to a tool result is regenerated once and then withheld. No model verifies another model on this path.

• Absent, not denied

  Approve, post and reverse do not exist in the AI and import layers, so no setting or prompt can reach them — and a source scan fails the build if one appears.

• A forecast is arithmetic, not a stored guess

  Cash-flow forecasts are recomputed from stated inputs and never stored, and every figure says whether it is actual, bank-asserted, budget or forecast.