Pack de IA para Finanzas y Contabilidad

What's in this pack

This is the stack for someone who actually has to close the month — a small-business accountant, a one-person FP&A team at a Series A startup, a CFO who still has to touch the books between board meetings. Not a Twitter list of 30 AI finance startups. Every pick here does one specific job in the close cycle: turn paper into data, get it into something auditable, categorize and analyze, model the future, and report it.

The whole pack is built around one principle: AI accelerates the work, but the audit trail is non-negotiable. That's why the ledger pick is plain-text (hledger), not a black-box SaaS. That's why categorization runs through a self-hosted app you can inspect (TaxHacker), not someone else's API. And it's why the spreadsheet picks let an agent edit your existing workbook instead of replacing it.

Install in this order (ingestion → categorization → analysis → forecasting → reporting)

1. Zerox — Zero-Shot PDF OCR for AI Pipelines — start here. Receipts, vendor invoices, and bank statements arrive as PDFs and JPGs. Zerox uses vision models to extract structured data in one shot, no template training. This is the front door.
2. Marker — Convert PDF to Markdown with High Accuracy — Zerox is great for single receipts; Marker shines on the long stuff (multi-page statements, 10-Ks, audit reports). Markdown out, tables preserved. Pair them.
3. Claude Official Skill: XLSX — Spreadsheet Operations — once your data is structured, you need a workbook. The official Claude skill reads/writes .xlsx cleanly, including formulas and formatting. This is the bridge between "raw extract" and "thing you can show your boss."
4. Excel MCP Server — Edit .xlsx Without Excel — MCP exposure of the same capability to your local agent. Lets Claude Code or Cursor open your monthly close workbook, append rows, recompute totals, and save — without you ever leaving the chat. Critical for repetitive period-end work.
5. TaxHacker — Self-Hosted AI Accounting for Receipts and Invoices — self-hosted app that scans receipts, categorizes expenses with an LLM, and tracks finances with multi-currency. Custom prompts mean you can tune it to your chart of accounts. The categorization layer that doesn't ship your data to a vendor.
6. Hledger — Robust Plain-Text Accounting Tool — the audit backbone. Every categorized transaction lands in a plain-text journal you can git diff, review, and roll back. Double-entry, reconciliation, balance sheets, P&L — all from one file you actually own.
7. Invoice Ninja — Open Source Invoicing and Payments — the AR side. Send invoices, track payments, reconcile receivables back into hledger. Self-hosted so client data stays on your server.
8. Claude Code Agent: Quant Analyst — financial modeling specialist subagent. Use it for valuation, scenario analysis, sensitivity tables, and DCFs. Once your data is clean in XLSX/hledger, this is the agent you point at it to do the FP&A heavy lifting.
9. Claude Code Agent: Report Generator — monthly board deck, variance commentary, narrative around the numbers. The Quant Analyst gives you the model; the Report Generator gives you the slide.
10. Dash — Data Apps and Dashboards in Pure Python — for the always-on view. Build a P&L dashboard, cash runway tracker, or AR aging report once in pure Python; refresh it on a schedule. Replaces "send me the latest sheet" forever.

How they fit together

   Paper / PDF / JPG
         │
   ┌─────┴──────┐
   ▼            ▼
 Zerox       Marker
 (1-page)  (multi-page,
            statements)
   │            │
   └─────┬──────┘
         ▼
  structured rows / markdown
         │
   ┌─────┴──────────┐
   ▼                ▼
 Claude XLSX   TaxHacker
 (workbook)    (categorize
   │            + GL coding)
 Excel MCP          │
 (agent edits)      ▼
   │            Hledger
   └─────┬──────(audit trail)
         ▼              ▲
   Invoice Ninja ───────┘
   (AR side feeds back)
         │
         ▼
   ┌──── Quant Analyst ────┐
   │  (modeling, sensitivity)│
   └────┬───────────────────┘
        ▼
   Report Generator      Dash
   (board deck text)  (live dashboard)

The critical join is Hledger + the agents: as long as the source of truth is one plain-text file, you can let agents read it, model on top, and write reports — without ever giving them write access to your books. Read-only AI on auditable data is the safe shape.

Tradeoffs you'll hit

• AI categorization vs rule-based — TaxHacker (or any LLM categorizer) will guess your chart of accounts brilliantly 90% of the time and quietly miscode 10%. Rule-based (hledger's built-in auto-postings, regex rules) is boring but reproducible. Best practice: rules first for the predictable 80%, LLM only for the long tail. Always review.
• Monthly close vs continuous close — automation tempts you toward "close every day." Resist for the first quarter. Run your first three closes on a monthly cadence so you learn where the categorization fails. Then go continuous, with anomaly checks instead of human review.
• Self-hosted vs SaaS for sensitive data — every tool in this pack except the Claude agents can be self-hosted. That's deliberate. Vendor invoices and payroll data are exactly the categories you don't want flowing to a SaaS you can't audit. Pay the 1-2 hours of Docker setup; save yourself the SOC 2 conversation later.
• Plain-text accounting vs QuickBooks/Xero — hledger is uncompromising and ugly. QuickBooks is friendly and a walled garden. If your business is <20 transactions/week and you can read text, hledger wins on auditability. If you have employees who need to enter invoices, stay on QuickBooks and let the AI agents read its exports — don't make staff learn Ledger syntax.

Common pitfalls

• AI hallucinating numbers — the single most dangerous failure mode in finance AI. An LLM that summarizes a P&L will invent line items that look plausible. Mitigation: never let the agent produce the number; let it produce the query (SQL, hledger command, formula) and run it deterministically. Then explain what it returns.
• Lack of audit trail — if you can't answer "why did this entry hit this account on this date" three months later, the audit will be miserable. Commit your hledger journal to git, tag each monthly close, and keep the OCR source PDFs in the same repo. Boring infrastructure that saves your year-end.
• Letting an agent write to the GL directly — don't. Agents propose journal entries to a staging file; a human (you) reviews and commits to the main journal. The latency cost is 5 minutes/week; the cost of an agent silently doubling an accrual is six hours of forensic work.
• OCR confidence without verification — Zerox will happily return $1,200.00 when the receipt says $12.00. Always reconcile the OCR'd total against the credit card statement before posting. Two-way match is not optional.
• Forgetting the legal layer — this pack is editorial guidance for a workflow. It does not replace your CPA, your tax advisor, or jurisdiction-specific bookkeeping rules. The tools make the work faster; they don't make you compliant. Talk to a human professional before any tax-relevant decision.