Articles 17th Nov, 2026

AI for financial data extraction: from family offices to banks and energy trading companies

A family office receives contract notes, dividend statements and portfolio reports from several custodian banks, each in its own layout. An energy trading company sees the same deal arrive as a broker recap, a counterparty confirmation and later an invoice. A bank collects financial statements from hundreds of borrowers. The figures these organisations need…

_Summary

  • The need is shared, the documents are not. Family offices, trading companies, banks and finance teams all need documents turned into structured data, but each has its own document types and target systems.
  • Domain knowledge drives accuracy. General-purpose models lose accuracy on tasks that need business context. Small language models trained on domain content are built to close that gap.
  • Small language models are built for this. Compact SLMs trained on one category of content, such as financial documents, are faster, cheaper and easier to run on your own infrastructure.
  • Extraction is a means, not an end. The value comes when validated data feeds the workflows that depend on it.

A family office receives contract notes, dividend statements and portfolio reports from several custodian banks, each in its own layout. An energy trading company sees the same deal arrive as a broker recap, a counterparty confirmation and later an invoice. A bank collects financial statements from hundreds of borrowers. The figures these organisations need already exist. They are simply locked in documents that someone has to read and re-type before any system can use them.

AI can now do most of this reading. But in financial operations, reading is not enough: each value has to be correct, mean the right thing and arrive in the right system. That is where small language models (SLMs) trained on financial content make the difference.

What do different organisations need from financial data extraction?

The task is the same everywhere: read the document, capture each value with its meaning, and deliver it in a consistent structure. What changes is the documents and where the data has to go.

OrganisationTypical documentsWhere the data needs to go
Family officesContract notes, dividend and interest statements, account statements, portfolio reports from several custodian banksPortfolio management system, consolidated reporting, compliance checks
Energy trading companiesBroker recaps, counterparty confirmations, master agreements, invoices, nomination and shipping documentsTrading and risk management system, reconciliation, settlement, REMIT reporting
BanksBorrower financial statements, loan and collateral documents, counterparties’ sustainability reportsCredit analysis, risk systems, regulatory reporting such as the Green Asset Ratio
Asset managersBroker and custodian statements, fund reports, corporate action noticesReconciliation, performance and client reporting
Corporate finance and treasuryBank confirmations, invoices, supplier contractsERP, cash management, audit files

In energy trading, for example, BCG describes AI that reads a broker message, extracts the key terms, prepares the trade entry and drafts the confirmation, with humans keeping approval rights. Deloitte lists extracting and validating key terms from contracts, confirmations, invoices and shipping documents among the core back-office uses of AI in commodity trading.

Different organisations, same bottleneck: the step between the document and the system.

What does this look like in a family office?

This is the use case we work on with family offices. A family office typically holds assets at several custodian banks, and each bank sends its own stream of documents: contract notes for every trade, dividend and interest statements, account statements, and investment and performance reports. Today, a team reads these documents and enters the figures into the portfolio management system by hand.

With DYDON AI, the documents go straight into the platform, whichever bank they come from and whatever their format. The AI recognises the document type and interprets each value in context, so a withholding tax line is not mistaken for a fee and a dividend is not booked as interest. The results are organised by client, depot and transaction type in one consistent structure. The team then confirms each figure against the exact line it came from in the source document. Once validated, the data flows through a single API into the portfolio management system, client reporting and internal compliance checks.

All of this runs in an isolated environment on Swiss or EU servers, and client data is never used to train the AI. We describe the full workflow in AI for family offices in 2026.

The result: no more re-typing of bank documents, and one validated dataset that feeds every workflow in the office.

What are small language models, and why do they fit financial data?

Small language models (SLMs) use the same transformer architecture as large language models (LLMs), but with far fewer parameters: usually up to around ten billion, against hundreds of billions for the largest LLMs. They are typically derived from larger models through knowledge distillation, pruning and quantization, then fine-tuned on task-specific data, such as thousands of labelled custody statements.

In financial data extraction, OCR and vision models first turn a PDF or scan into text. The SLM then classifies the document and returns each value into a fixed schema (ISIN, quantity, price, value date), and rule-based checks verify the result before it moves on.

NVIDIA researchers argue in Small Language Models are the Future of Agentic AI that SLMs are powerful enough, better suited and more economical for the narrow, repetitive tasks that make up most AI workflows, with larger models called in only when needed.

Gartner expects organisations to use small, task-specific models three times more than general-purpose LLMs by 2027. And in a study on financial transaction data, a small purpose-built model matched the accuracy of an 8-billion-parameter general model at about a tenth of the cost.

The caveat: without task-specific training, larger models still extract financial data better. SLMs earn their accuracy through targeted training and the pipeline around them, and in return they run on your own infrastructure at a predictable cost.

Small does not mean limited. It means the right language model for each type of document.

How does DYDON AI handle financial data extraction?

DYDON AI’s platform for Regulatory Intelligence captures and structures data from thousands of documents and puts it to work across workflows, always under human control. For financial operations, that means trade confirmations, custody statements, corporate actions, settlement instructions and portfolio statements from banks, brokers and custodians, in any format, become verified data in your back office.

The flow takes four steps:

  1. Receive. Documents from all custodian banks and brokers arrive in one channel, in bulk or via API.
  2. Read and understand. The AI recognises the document and transaction type and assigns each value by its meaning, not only its position on the page. Account, ISIN, quantity, price, fees, currency, trade and settlement date, counterparty: every key figure is extracted and structured by client, account and category.
  3. Your team validates. Each field is checked before it enters your system, and every value can be traced back to its source line in the original document. No entry reaches the target system without your team’s approval.
  4. Transfer. Approved data flows via API into your portfolio management or accounting software, as well as into reporting and compliance workflows.

Accuracy comes from domain depth: the platform is calibrated on real bank documents and draws on securities accounting expertise. [Confirm whether to describe DYDON’s models explicitly as small language models (SLMs).] Volumes can grow across banks, clients and accounts without the workload growing with them. A multi-family office service provider in the DACH region already uses it this way: its accounting team now reviews and approves entries instead of retyping them.

One platform, many needs. The same foundation serves very different use cases, from answering regulatory questionnaires for a bank to extracting financial data for a family office. In every case, the platform reads the documents, extracts the relevant data and puts it to work. Validated data can feed regulatory reports and questionnaire answers, flow into portfolio management or accounting systems, or support operations and back-office processes. Everything comes from one central data collector, so the same information is never extracted twice.

Everything runs in client-isolated infrastructure, with a hosting model aligned to each client’s security requirements, and client data is never used to train the AI. See how it works for family offices on our dedicated page.

Let’s talk about your use case

Which documents arrive in your organisation every day, and where do their numbers need to go? Our free AI potential analysis maps the document work that can be automated, where human validation should stay, and what a secure, API-connected setup could look like for you.

Get your AI potential analysis →