Automation
Finance
AI
Accounting
Automated Transaction Categorisation: How Machine Learning Saves You Hours
Discover how automated transaction categorisation works in Dutch bookkeeping software. Machine learning classifies transactions, handles BTW codes, and helps Dutch BV founders save dozens of hours.
•
18 mins
Intro
Dutch entrepreneurs spend an average of 8 hours per week on administrative tasks. A significant portion of that time is spent on repetitive bookkeeping work: reviewing transactions, assigning ledger accounts, checking BTW treatments, and preparing records for quarterly tax filings. For many founders, these tasks are necessary but add little value to the business itself.
This is exactly where machine learning has transformed modern bookkeeping. Instead of manually reviewing every incoming bank transaction, intelligent bookkeeping systems can automatically categorize the majority of transactions based on patterns they have learned from historical data. The result is not futuristic automation that replaces humans, but practical automation that removes repetitive work while allowing founders and accountants to focus on higher-value decisions. Based on the briefing provided by the user.
The Problem: How Many Hours Do Dutch Entrepreneurs Lose to Manual Bookkeeping?
For every transaction that enters a Dutch business bank account, someone must decide which grootboekrekening it belongs to and which BTW treatment applies. That sounds simple until you realize how many transactions even a small company processes.
A typical BV processing 200 transactions per month must make 200 separate bookkeeping decisions. Every payment, invoice, subscription charge, bank fee, and expense requires categorization before it can be used in financial reporting, VAT filings, or VPB calculations.
For a founder billing clients at €100 per hour, those administrative hours quickly become expensive. Eight hours per week equals more than €40,000 per year in opportunity cost.
This is also changing the role of an accountant or bookkeeper. Rather than spending hours manually entering transactions, professionals increasingly focus on review, compliance, and strategic advice while automation handles routine categorization.
Automated transaction categorization removes most of the repetitive decision-making by suggesting or automatically assigning categories based on previous behaviour.
How Machine Learning Categorization Actually Works
The phrase "machine learning" appears everywhere in bookkeeping software marketing, but few founders understand what actually happens behind the scenes.
The process can be broken down into four practical steps.
Step 1: Feature Extraction
When a new transaction arrives, the system reads multiple signals:
Transaction description
Counterparty name
IBAN information
Transaction amount
Transaction date
Existing invoice matches
Step 2: Pattern Matching
The model compares these signals against previously categorized transactions.
Research from the University of Twente found that combining transaction descriptions with transaction amounts produces significantly higher accuracy than relying on either factor alone. A payment of €29.99 to Adobe carries both textual and numerical clues that help identify the correct category.
Step 3: Confidence Scoring
The system calculates probabilities for every possible ledger account and BTW code.
For example:
Category | Confidence |
Software subscriptions | 96% |
Office expenses | 3% |
Marketing expenses | 1% |
If confidence exceeds a predefined threshold, often between 80% and 95%, the transaction can be categorised automatically.
Step 4: Human Review
Transactions that fall below the confidence threshold are placed into a review queue.
Every correction teaches the model something new. Over time, accuracy improves significantly.
For entrepreneurs who are starting a company in the Netherlands, this means the system becomes smarter as the business grows and processes more transactions.
Most established Dutch bookkeeping platforms report auto-categorization rates between 70% and 90% after several months of learning.
The Dutch Context: BTW Codes, Grootboekrekeningen, and Why Generic AI Falls Short
Dutch bookkeeping has unique requirements that many international bookkeeping systems struggle to handle correctly.
Every transaction requires two decisions:
Which grootboekrekening applies?
Which BTW treatment applies?
The Dutch BTW system includes:
BTW Category | Rate |
Standard rate | 21% |
Reduced rate | 9% |
Zero rate | 0% |
Exempt | N/A |
A miscategorized transaction does not simply affect bookkeeping accuracy. It can directly affect quarterly VAT returns and lead to corrections, additional tax assessments, or even tax interest.
This becomes especially important when preparing to file VAT returns.
Dutch-trained models generally perform better than generic international systems because they recognize common transaction descriptions such as:
Albert Heijn
Jumbo
NS Reizen
KPN
Odido
The quality of bank transaction data also matters. PSD2 feeds from ABN AMRO, ING, and Rabobank tend to provide more consistent transaction descriptions, improving categorization accuracy.
The Accuracy Curve: Why New BVs See Different Results Than Established Businesses
One of the most important realities of machine learning categorization is that accuracy improves with experience.
Stage | Transaction History | Typical Auto-Categorization Rate | Main Limitation |
New administration (0–3 months) | Under 100 transactions | 40–60% | Limited historical data |
Early stage (3–6 months) | 100–500 transactions | 60–75% | Limited supplier patterns |
Established (6–18 months) | 500–2,000 transactions | 75–88% | Most recurring transactions learned |
Mature (18+ months) | 2,000+ transactions | 85–92% | Mainly new suppliers remain |
The practical lesson for founders is simple.
Do not expect perfection during the first few months. Every correction improves future performance. Consistently correcting mistakes trains the system. Inconsistent corrections confuse it.
The first 90 days are often the most important learning period.
Gemengde Kosten and the DGA Mixed-Use Problem
Not every transaction should be automated.
Certain categories require human judgement because Dutch tax law depends on circumstances that software cannot always determine.
Examples include:
Restaurant expenses
Entertainment expenses
Personal services
Home office costs
Mixed-use assets
Vehicle expenses
For example, restaurant meals create two separate challenges:
Restaurant BTW is generally not reclaimable.
Mixed-use expense rules affect deductibility for VPB purposes.
If an AI model treats a restaurant bill as a fully deductible business expense, it may create both a BTW error and a VPB error.
These decisions directly affect how much tax you pay.
The safest approach is to automatically route categories such as restaurants, entertainment, and personal services to human review regardless of the confidence score.
PSD2 Bank Connections: The Foundation That Makes Automation Possible
Automated categorization only works when accurate transaction data enters the system.
PSD2, the European open banking framework, makes this possible.
Before PSD2, entrepreneurs had to download bank files manually and upload them into bookkeeping software. Today, authorized software can receive transaction data directly through secure APIs.
The process looks like this:
Connect bank account.
Grant PSD2 authorization.
Transactions flow automatically into bookkeeping software.
Machine learning categorizes transactions.
Accountant reviews exceptions.
Banks with particularly strong PSD2 data quality include:
ABN AMRO
ING
Rabobank
Bunq
Revolut
Whether you operate as a BV or sole trader, maintaining connected accounts improves bookkeeping efficiency dramatically.
Importantly, PSD2 uses OAuth authentication. Bank credentials are never shared with the bookkeeping platform and permissions can be revoked at any time.
A Worked Example: Time Saved for a Typical Dutch BV
Abstract percentages are useful, but a practical example is easier to understand.
Business Profile
Single-DGA IT consulting BV:
Monthly revenue: €15,000
180 monthly transactions
60 client payments
45 supplier invoices
30 software subscriptions
25 miscellaneous expenses
20 payroll transactions
Before ML Categorization
Metric | Value |
Average time per transaction | 2 minutes |
Monthly transactions | 180 |
Monthly bookkeeping time | 6 hours |
Annual bookkeeping time | 72 hours |
Opportunity cost (€100/hour) | €7,200 |
After ML Categorization (85% Accuracy)
Metric | Value |
Manual reviews required | 27 transactions |
Review time per transaction | 90 seconds |
Monthly bookkeeping time | 40 minutes |
Annual bookkeeping time | 8 hours |
Opportunity cost | €800 |
Result
Metric | Value |
Time saved | 64 hours |
Opportunity cost saved | €6,400 |
Software cost | €360–€720 |
Net annual benefit | €5,700–€6,000 |
This assumes an established administration with at least 18 months of transaction history. New BVs should expect lower accuracy initially, with improvements over time.
What to Set Up and What to Monitor: A Practical Configuration Guide
Getting the most value from machine learning categorization requires proper setup.
1. Connect Every Relevant Bank Account
Include:
Main current account
Savings accounts
Business credit cards
2. Configure Your Chart of Accounts First
The model can only categorize into accounts that already exist.
The standard Dutch chart of accounts should be configured before importing transactions. This also supports accurate financial reporting and your future balance sheet Netherlands.
3. Configure BTW Codes Correctly
Each ledger account should have a default BTW treatment.
4. Create Rules for Recurring Suppliers
Examples:
Internet provider
Rent
Microsoft
Google
Adobe
5. Review Daily During the First Three Months
Small corrections today create large efficiency gains later.
6. Audit Auto-Categorized Transactions Monthly
Even at 85% accuracy, periodic spot checks remain valuable.
7. Flag Mixed-Use Categories
Restaurant, entertainment, and personal-service transactions should always receive human review.
The Human-in-the-Loop: What Still Needs Professional Oversight
The most accurate description of machine learning bookkeeping is not that it replaces accountants.
It changes how accountants spend their time.
Tasks that automation handles well:
Subscription payments
Payroll postings
Supplier invoices
Client payment matching
Bank fees
Tasks that still require judgement:
New suppliers
Asset purchases
Depreciation schedules
Year-end accruals
Intercompany transactions
Special BTW treatments
For DGA founders, matters such as VPB planning, salary decisions, and DGA salary vs dividend remain human decisions.
This is where the strongest modern platforms differentiate themselves. Rather than replacing accountants, they combine AI-driven categorization with professional oversight.
Neno follows exactly this approach: machine learning handles repetitive categorization while certified Dutch accountants monitor exceptions, compliance, and strategic decisions.
Save Time Without Sacrificing Accuracy
Let Automation Handle the Repetition
The biggest benefit of automated categorization is not technological sophistication. It is getting hours of your week back.
By combining PSD2 bank feeds, machine learning categorization, and professional oversight, modern bookkeeping systems can automate the majority of repetitive bookkeeping work while maintaining compliance with Dutch tax rules.
Whether you want to incorporate your BV, improve your bookkeeping and payroll, or simply spend less time categorizing transactions manually, automation can significantly reduce the administrative burden of running a business.
Book a demo and discover how Neno combines AI-driven bookkeeping with embedded financial expertise.
FAQs: Automated Transaction Categorization for Dutch Entrepreneurs
What is automated transaction categorization?
It is the process of automatically assigning incoming bank transactions to ledger accounts and tax categories using machine learning and predefined rules.
How accurate is machine learning categorization?
Established Dutch administrations typically achieve auto-categorization rates between 80% and 90%.
Does ML categorization handle Dutch BTW codes correctly?
Yes, but mixed-use costs and unusual transactions still require human review.
How does the system learn from corrections?
Every manual correction becomes training data that improves future categorization decisions.
Can I trust automated categorization for my VPB return?
For recurring transactions, generally yes. However, year-end reviews and professional oversight remain important.
What is a PSD2 bank connection?
A secure API connection that allows bookkeeping software to access transaction data directly from a bank.
Is PSD2 secure?
Yes. PSD2 uses regulated authentication systems and does not require sharing banking passwords.
How long does it take before accuracy improves?
Most businesses see substantial improvements after three to six months of consistent use.
Do I still need a bookkeeper?
Usually yes. Automation reduces manual work but does not replace tax expertise, compliance reviews, and strategic advice.
Which Dutch platforms support machine learning categorization?
Popular examples include Exact Online, Moneybird, Jortt, Informer, AFAS, and Neno.
Written by
Nick Knuppe
CEO & Founder
