Data and connections
Counterparties' financial statements, the analytical sheet and the analyst panel. Language models are called through a single gateway and run locally on the client's server.
Financial statements and reports without manual retyping
Implementation in progress. AI reads counterparties' financial statements, checks their arithmetic and feeds the analytical sheet, and the analyst approves the result instead of retyping the data by hand.
Results
- minutes*
- to read a statement instead of 15-30 min of manual retyping
- min. 90%*
- of documents processed automatically, the rest go to an analyst
- approx. 200*
- documents per month in the process we are automating
Work in progress. These are results verified at the test stage or the targets we are aiming for. We will confirm them on completion.
How it works
What the AI does
Identifying type and period
balance sheet · fiscal year 2025 · comparative column 2024
Scan without a text layer: OCR
4 pages · stamp and signature on page 2
Reading tables0 / 42
line items into the balance sheet schema
Checking control totals
assets = equity and liabilities · current assets: PLN 500 difference
Checking period consistency
opening 2025 = closing 2024
Handing over to the analyst
1 discrepancy · original next to the data
Data for the analytical sheet
Counterparty A · balance sheet 2025
closing 2024 = opening 2025
Non-current assets12,480,000
Inventories2,140,500
Short-term receivables4,862,800
Cash1,212,500
Total current assets8,215,300
Total assets20,695,300
Total equity and liabilities20,695,300
Line items add up to PLN 8,215,800, the document says PLN 8,215,300. PLN 500 difference in receivables: there is a stamp over the row.Totals and periods match after the analyst’s correction. Only approved data goes into the sheet.
ApproveCorrectReject
Saved to the analytical sheet
01
Challenge
The financial analysis team retyped data by hand from clients' financial statements, in many categories and layouts, into a sheet where it calculates ratios to assess financial standing: 15-30 minutes per document. Documents arrive in different layouts and for different periods, often as scans with stamps and signatures.
Implementation is in progress. The figures for this case study are results from the test stage or targets we will confirm after acceptance.
02
Solution
The solution recognizes the format and period of the statement, reads the tables, extracts the data into a fixed schema and checks control totals and consistency between periods. Documents with discrepancies go to a panel where the analyst sees the original next to the data and approves, corrects or rejects the result, and only accepted data goes into the counterparty's sheet.
03
How we built it
The pipeline checks whether a file has a text layer or is a scan; scans go through OCR with Polish language support, and the hardest ones through an additional OCR language model. A language model extracts the data into the schema for the document type, long documents page by page, and the confidence score comes from arithmetic checks, not from the model's self-assessment. Models with open licenses run on the client's server and are called only through a single gateway. During the build phase the model received only anonymized text; today it runs locally on the client's server and documents do not leave the client's network. Every operator decision stays in the change history, and we measure quality on a set of statements verified by hand by the client's analysis team.
Tesseract OCROpen-weight modelsLiteLLMOllamaPythonPostgreSQLDocker
Workflow
How the solution works
- 01
Reading the document
The solution recognises the format and period of the statement, also on scans, and extracts the data into a fixed schema.
- 02
Arithmetic checks
Confidence in the result comes from checksums and consistency between periods, not from the model rating itself.
- 03
The analyst's decision
Documents with discrepancies go to a panel where the analyst approves, corrects or rejects the result.
The human role
The analyst decides what goes into the counterparty sheet. Every decision stays in the change history.
Lessons
What made the difference.
- 01
Reads scans and tables
Recognizes the statement layout, period and fiscal year, also on scans with stamps and in tables with comparative columns.
- 02
Arithmetic instead of self-assessment
Confidence in the result comes from checks of totals and period consistency, not from the model's own claim, and every discrepancy goes to an analyst.
- 03
Data stays in the company
During the build phase the model saw only anonymized text. Now the whole solution runs on the client's server with a local model.
Document extractionData validationLocal modelOperator control
Company of the Future
How this builds the Company of the Future
A project delivers results on its own, but it also leaves organized context behind in the company. Every next AI solution builds on it. This project added to the AI Engine:
- Context Lakehouse
Data from counterparties’ financial statements in one schema, ready for calculating financial ratios.
- Skills
Checks of totals and period consistency written down as rules: the confidence of a reading comes from them, not from the model rating itself.
- Decisions
Every analyst decision (approve, correct, reject) stays in the change history.
Value
Where value is created
Time saved
The analyst checks the extracted data instead of retyping it by hand.
New possibilities
Counterparty data in one schema, ready for calculating financial ratios.
What comes next
The next steps are testing with the client's analysis team and launch in the production environment.
Most of our projects are covered by NDAs, so we only share part of them, without client names or details, as inspiration. We are happy to tell you more about our experience in a conversation.
Let’s talk about the work you want to improve.
Tell us how a process works today. Together, we will explore where AI Engine can help and which module to start with.
Book a callOr write to: hello@wearefuture.ai
