Service
Finance automation and AI — from statement to report without retyping
Registry filings, ledgers and time records are read by machines; conclusions and accountability stay with a person. I build pipelines that turn a system export into analysis, control and a report — with a language model where reading documents matters, and with code where arithmetic does.
What I automate — from practice
| Task | How it works | Effect |
|---|---|---|
| Reading registry filings | A parser for XML e-filings: balance sheet, P&L, cash flows, notes — loaded into one structure and reconciled (assets = liabilities, result = result) | Ratio analysis of a dozen companies with one method in minutes instead of days; market screening |
| Ledger control for an accounting office | Around a hundred rules (VAT, deadlines, counterparties against the official register, e-invoicing, central-bank rates, account balances) on the accounting-system export; several hundred automated tests | Errors caught before month-end, not after an audit; a dashboard for the office owner |
| Management report from the model | A script reads the model (Excel), builds an HTML dashboard, a PDF report and a presentation; closing checks stop publication when the numbers do not reconcile | The same number in three documents; Polish and French versions from one source |
| Time records → client cost | A monthly export from the ticketing system, a star schema, allocation of personnel cost to clients and departments | Client profitability visible monthly; cost per counterparty allocated from 16 % to 44 % |
| A glossary that answers | A base of 176 financial terms with semantic search (RAG) — answers cite the source, not “the model’s memory” | One place where the board and the accountants understand the same word the same way |
What I never hand to the model
- Arithmetic. A language model does not compute — it guesses numbers. Every multiplication lives in code or a formula, with a test.
- Normalisation adjustments and valuation method weights — decisions someone answers for at the table.
- Forecast assumptions and the final conclusion for the board. The model may propose; a person signs.
- Client data for training anything. Materials stay local, on an encrypted drive.
How I work
- 1A map of the process and the dataA week to see how the report is made today: where the export comes from, who retypes, where the hours vanish and where the discrepancies arise. Usually 70 % of the time is three activities — and those go to the machine first.
- 2A data gate and zonesData enters through one gate (CSV/JSON/XML), with validation and a log. Sensitive data (names, personal IDs, salaries) stays in a zone the model cannot reach; analysis receives identifiers. Not an option — a condition for me taking the deployment on at all.
- 3Rules and tests before the modelEverything that can be written as a rule (segments sum to the P&L, assets equal liabilities, the central-bank rate, a counterparty on the official register) is a rule with a test. The language model gets what remains: reading contracts, classifying descriptions, extracting fields from documents — always with a human sample check.
- 4A tool that outlives its authorDocumentation, a register of pitfalls (what went wrong and why), refresh instructions and team training. Quality checks built into the tool: a document that checks itself before anyone opens it.
What you get
- A working pipeline: from the system export to the report — run by your team, not by me.
- Control rules with tests and a dashboard showing what failed today.
- Documentation, a register of pitfalls, instructions and training — plus a written boundary: what the machine does, what a person does.
- A payback account: how many hours a month return to the team and how many errors never reach the report.
Own projects: the analysis engine, the ledger controller, FinGloss
Frequently asked questions
Does my data go to an external model?
Only what must, and only after anonymisation: the language model gets a text to classify, not the client base. Where a client requires it I use a locally run model — slower, but nothing leaves the company. The zone rule: sensitive data never passes the gate into analysis.
Where to start in a small company?
With the one report that eats the most hours today — usually the monthly result or time allocation to clients. Export, account map, model, dashboard script. Two to three weeks; the effect shows at the second month-end. Only then control rules and the language model.
What tools?
Python for processing and rules, Excel on live formulas as the model, HTML/PDF for reports, SQL and BI where data sits in a database, language models (Claude, also local ones) for reading documents. I do not sell any platform — I pick what your team can maintain.
Will AI replace the accountant or the controller?
It will not replace them — it will change what they do. Retyping, reconciling and hunting errors disappear; understanding the business and accountability remain. In an accounting office with control rules the accountant works on exceptions, not on every entry. That is a gain, not a threat — provided the boundary between machine and person is written down.
Let’s talk about your situation
Write a few sentences about the company and the problem. I reply within two working days and the first conversation is free.
Write to me