AI Bookkeeping Approval Workbench

• [QuickBooks Accounting AI](https://quickbooks.intuit.com/ai-accounting/) includes categorization, anomaly detection, bank feeds, reporting, and human control inside a full accounting suite. Its published U.S. Simple Start list price is about $38/month, subject to promotions and change.
• [Bonnie](https://mybonnie.ai/pricing) offers bank/Plaid or file imports, up to 3,000 AI-categorized transactions per month, P&L, invoices, and document storage for $15/month.
• [TraceEntry](https://traceentry.com/pricing) charges $5/500 transactions, $15/2,000, and $30/5,000 for AI categorization, custom charts of accounts, review, and CSV/Excel export.
• [Accounting Workbench](https://www.accountingworkbench.com/) serves accounting practices with PDF, Excel, and invoice extraction, categorization, review, and exportable double-entry journals.
• [Xenon Connect](https://www.xenonconnect.com/) says more than 1,400 accounting and bookkeeping firms use its Xero, QuickBooks, FreeAgent, and Sage health checks. Error detection is not an empty market either.
• [Magpie](https://github.com/kyle-visner/magpie) is a CLI/domain engine, not a hosted SaaS or substitute for professional judgment. It is evidence about method, not a product that can simply be wrapped. Its AGPL license is another reason not to copy its code into a closed hosted MVP.
• Entry wedge: do not replace QuickBooks/Xero, connect bank accounts, or become the ledger. Accept drafts from this product or any agent/LLM/CSV flow, and make original suggestion, deterministic failure, low confidence, human edit, approver/time, and export state the product.
• **SERP occupancy:** Head terms are visibly occupied by active SaaS vendors, major brands, and review publishers. Narrow terms such as AI bookkeeping review, bookkeeping approval workflow, and AI bookkeeping audit trail still need DataForSEO and a reproducible Top 10. If those are also fully occupied, acquisition must begin with bookkeeper outbound and demos.

MVP Checklist

1. 1.Define the v1 accounting scope, supported CSV/CoA schema, disclaimer, and synthetic fixtures.
2. 2.Build browser CSV parsing, field mapping, locale handling, sensitive-field selection, and data-quality checks.
3. 3.Build the normalized transaction schema and vendor deduplication.
4. 4.Add structured LLM output with batching, timeout, retry, and cost/row limits.
5. 5.Build the deterministic validation engine and unit tests.
6. 6.Build the exception table, evidence drawer, approve/edit/reject, and audit events.
7. 7.Build generic, QuickBooks-style, and Xero-style exports plus the sample report.
8. 8.Build Home, Methodology, About, FAQ, and Privacy.
9. 9.Add the free 50-row limit, `$49` Payment Link, lead form, and anonymous funnel.
10. 10.Run build, functional, SEO, security, and deployment-readiness checks using at least 20 synthetic or de-identified real-like files.

Don't Build

• ✕Do not expand features without validation.
• ✕Do not add Plaid, bank feeds, a full ledger, or a complex backend.
• ✕Do not build custom login, membership, orders, subscription billing, or admin first; the `$49` Payment Link and lead form are mandatory.
• ✕Do not sacrifice launch speed for completeness.
• ✕Do not remove real structured model drafts, deterministic rules, human approval, or exports to make the project “lighter.”
• ✕Do not auto-post to any accounting system.
• ✕Do not let the model decide whether entries balance or rules pass.
• ✕Do not claim accounting, tax, or legal advice.