How to Automate Bookkeeping with AI in 2026

A good starting point is to separate bookkeeping operations from tax decisions.

AI can automate bookkeeping operations very well because the inputs are structured (bank transactions, invoices, receipts) and the outputs are repetitive (categories, VAT codes, summaries). Tax decisions are harder because they depend on your full situation, your choices (like KOR), and year-specific rules.

Here’s what we reliably automate today.

What AI can automate (high confidence)

1. Transaction categorisation

1. VAT (BTW) suggestion and running totals

1. Receipt OCR and matching

1. Anomaly detection

1. Exports and handover packs

What AI can’t fully automate (yet)

1. Tax strategy and planning

1. Complex edge cases

1. Audits and representation

1. BV structures and payroll

A practical way to think about it: if the output is a repeatable classification (category, VAT code, attachment link), AI is strong. If the output is a choice with long-term consequences, keep a human checkpoint.

If you remember one rule: let AI automate the routine work, but keep human checkpoints wherever a wrong decision creates tax risk.

You can set up AI bookkeeping in a single afternoon if you follow a clear sequence. The order matters because each step reduces ambiguity for the next one.

Here’s the 7-step setup we use when onboarding businesses into Fiscal Agent.

Step 1: Decide your scope and cut-over date

Pick a clean cut-over. The easiest is the start of a quarter so your VAT totals align. If you’re switching mid-quarter, do it anyway, but write down:

• The cut-over date


• Which bank accounts are in scope (business account, VAT reserve account, savings)


• Which payment providers are in scope (Mollie, Stripe, PayPal)


• Which invoice tools are in scope (Moneybird, Exact, Jortt)

Step 2: Connect your bank via PSD2 (or import a CSV)

For automation you want a PSD2 bank feed. PSD2 access is consent-based and typically read-only for bookkeeping. You authenticate with the bank and grant access; you do not share passwords.

Two practical details:

• Many banks require you to renew consent periodically (often around a 90-day window).


• PSD2 covers bank accounts, but payment providers and credit cards may require separate imports.

Step 3: Import 1–3 months of history and tag the obvious patterns

Don’t start from a blank slate. Import at least one full month of transactions (three months is better). This gives the system vendor repetition to learn from.

If you have 150 transactions/month, three months gives 450 examples. That’s enough to establish patterns like:

• recurring subscriptions (Adobe, Google Workspace)


• telecom (KPN, Vodafone)


• fuel/transport (if you drive for work)


• payment providers (Mollie/Stripe payouts)

Step 4: Configure your VAT profile (BTW, KOR, filing frequency)

AI bookkeeping must know your VAT posture:

• Are you a normal VAT entrepreneur or are you VAT-exempt?


• Are you on KOR (Kleineondernemersregeling) or considering it? (KOR turnover threshold is €20,000/year.)


• Do you file monthly, quarterly, or annually?

Also configure the VAT “edge cases” you expect:

• foreign SaaS subscriptions (reverse charge patterns)


• subcontractor invoices with “btw verlegd”


• EU clients (VAT ID validation and evidence requirements)

Most automation fails because receipts are missing. Make receipt capture easier than doing nothing:

• photograph receipts immediately


• forward supplier invoices to a dedicated email address


• store originals (PDF/photo) and keep the link between document and transaction

The operational rule we use: if a cost has VAT implications, it needs evidence attached in the system. If it’s a recurring subscription, attach the invoice once and let the system match future payments.

Step 6: Review the first month and teach the system

The goal is not to correct every line. The goal is to correct patterns.

We recommend a structured review:

• Sort by vendor and fix the top recurring vendors first


• Check VAT handling on foreign suppliers (reverse charge patterns)


• Flag private/business splits (phone, car, home office items)


• Look at the “unknown / low confidence” queue and resolve it

Step 7: Generate your first BTW overview and run a quarter-to-date check

Before you rely on automation, run a quarter-to-date sanity check. A good AI bookkeeping system should give you:

• output VAT total (VAT on your invoices)


• input VAT total (VAT on your costs)


• net payable or refundable


• a list of exceptions that need attention (reverse charge, missing invoices, foreign VAT)

Once the system is running, the maintenance habit is simple:

• 10–15 minute review twice per month (exceptions only)


• quarter-end exception review one week before the BTW deadline


• export pack for your accountant (if you use one)

A practical addition that prevents a lot of VAT mistakes is to explicitly configure the VAT rates and evidence expectations you’ll see. In the Netherlands, the most common VAT rates are 21% (standard), 9% (reduced), and 0% in specific cross-border situations (with conditions). Your AI bookkeeping setup should be able to tag costs and revenue with the right VAT code, not just a category.

For invoices and receipts, we also recommend you validate a minimum set of fields during onboarding. A VAT invoice typically needs, at a minimum: an invoice date, a unique/sequential invoice number, supplier name and address, a description of what was supplied, the VAT rate, the VAT amount, and totals. If you do EU B2B work, VAT ID evidence and correct wording becomes more important.

AI can’t magically create evidence. What it can do is enforce the workflow: if a transaction should have an invoice attached, it gets flagged until you attach it. That one rule is what turns automation into audit readiness.

“AI bookkeeping” is only valuable if you can trust it. The right way to evaluate trust is to measure accuracy the same way you’d measure a human bookkeeper: sample, verify, and track error types.

Our production benchmark

Across thousands of real Dutch transactions processed through Fiscal Agent, we see:

• 95%+ categorisation accuracy (transactions end up in the correct category after review)


• 97.8% average categorisation confidence (model certainty on its proposed category)

Accuracy vs confidence (and why both matter)

• Accuracy is hindsight: after you review, did the transaction land in the correct category/VAT treatment?


• Confidence is foresight: how sure is the system before you review?

How to measure accuracy in your own setup

Use a simple approach for the first month:

1. Take a sample of 100 transactions.


2. For each transaction, verify the category against the underlying evidence (invoice/receipt, vendor type, business purpose).


3. Track three metrics:

Then categorise your mistakes. Most errors fall into predictable buckets:

• unclear vendor descriptions (payment provider references)


• mixed-use expenses (phone, car)


• missing evidence (no invoice attached)


• cross-border VAT logic that requires confirmation

People assume manual work is “correct by default”. It isn’t. When we onboard freelancers who have been doing their own bookkeeping, we often find a small set of recurring issues:

• subscriptions booked under random categories


• missing receipts (so VAT can’t be reclaimed even when allowed)


• foreign supplier VAT handled inconsistently

How AI improves over time

AI bookkeeping improves with feedback loops:

• vendor memory: once “KPN” is telecom, it stays telecom


• category confirmation: your corrections become rules for similar transactions


• anomaly patterns: if your normal fuel spend is €0 and you suddenly have €400, it gets flagged

Edge cases that deserve a dedicated review

When accuracy drops, it’s usually not because the model forgot what ‘Adobe’ is. It’s because the input is ambiguous. The biggest sources of ambiguity we see:

• Payment provider payouts (Mollie/Stripe): the bank line is a net payout; detail lives in the provider statement


• Refunds and chargebacks: a negative amount can be revenue correction, a supplier refund, or a personal transfer


• Foreign VAT on travel: hotel/restaurant receipts abroad often include foreign VAT that has different reclaim rules


• Transfers between your own accounts: these should not inflate revenue or costs

A simple accuracy scorecard for the first 30 days

We recommend tracking these four numbers for your first month:

1. % transactions auto-categorised with high confidence


2. % transactions in the exception queue


3. # corrections needed after review


4. # transactions missing evidence (receipt/invoice)

Most businesses don’t want to replace everything. They want automation that fits into their current workflow.

AI bookkeeping can integrate in three practical ways:

1. As the primary bookkeeping layer

1. As an automation layer on top of existing tools

1. As a review and exception layer

What “integration” should mean in practice:

• categories map cleanly between systems (chart of accounts)


• VAT codes map cleanly (standard rate, reduced rate, reverse charge, exempt)


• exports preserve dates, amounts, and references


• attachments stay linked (PDFs/photos) so you keep audit readiness

If you’re evaluating a provider, ask what export formats they support, whether they support multiple bank accounts, and whether they preserve links to evidence. Exports without attachments force you back into manual admin.

What this looks like with Moneybird

Moneybird is often the system people keep for invoicing. In that setup, we typically automate everything that happens after an invoice is sent:

• import bank transactions (PSD2/CSV)


• match incoming payments to invoices (by amount/reference)


• categorise costs and attach receipts


• generate a quarter-to-date BTW overview

What this looks like with Exact Online

Exact Online is common when you already have an accountant involved. The most common approach is not to replace Exact—it’s to reduce the hours spent inside it. We prepare categorised transactions and attachments, then export them in a format your accountant can import and review.

CSV vs API: what matters

API integrations are convenient, but CSV is still the universal fallback. What matters is not the transport—it’s the completeness of the data. Any export should include: date, amount, counterparty/description, category, VAT code, and a link or bundle of attachments.

Automation usually fails because people treat it like a one-time setup instead of a workflow. Here are the mistakes we see most often.

Mistake 1: Never reviewing AI suggestions

AI should reduce review, not eliminate it. If you never look at exceptions, you’ll accumulate small mistakes. The fix is simple: a short recurring review habit (10–15 minutes, twice a month).

Mistake 2: Mixing personal and business accounts

If you run everything through one account, categorisation becomes guesswork. Even with AI, you’re forcing the system to infer intent. The best practice is:

• one business account for business income/expenses


• a separate account for tax reserves (optional, but helpful)

A transaction without evidence is a future problem. Set up a receipt workflow that is easier than doing nothing:

• take a photo immediately


• forward invoices to a single inbox


• match receipts to transactions weekly

KOR changes your entire VAT posture. If you opt in and you keep reclaiming VAT, you create a compliance problem. If you opt out and you should have opted in, you might be leaving cash on the table. Automation can track the €20,000 turnover threshold, but you still need to decide.

Mistake 5: Treating reverse charge as “no VAT”

Foreign suppliers and “btw verlegd” invoices are common. They need consistent handling and evidence. The right workflow is: flag, review, and store the invoice.

Mistake 6: Forgetting payment provider reconciliation

If you use Mollie/Stripe/PayPal, the bank deposit is often a net payout after fees and refunds. If you book only the bank line, you lose detail. The fix is to import provider statements (often CSV) so revenue, fees, and refunds are visible.

Mistake 7: Starting without a clean cut-over

If you switch tools mid-year without a plan, you create reconciliation work. Start at the beginning of a quarter when possible. If not, document the cut-over date and export the old system for your 7-year records.

Mistake 8: Forgetting private/business splits

Phone, car, and home office items often require a private/business split. AI can flag likely mixed-use expenses, but you need to confirm the percentage and keep the reasoning consistent.

If you avoid these mistakes, automation becomes stable: the system learns your patterns, exceptions stay small, and quarter-end becomes a review exercise instead of a rebuild.

Mistake 9: Letting VAT rates drift

Even if you don’t think about VAT often, your books still need consistent VAT coding. A common failure mode is that reduced-rate (9%) receipts or reverse-charge items get booked as “no VAT” because nobody checked the invoice. The fix is to make VAT part of the exception review: any transaction with foreign suppliers, any unusually high cost, and any category that often carries VAT should be checked for the VAT code and evidence.

Mistake 10: Not reconciling transfers and taxes

Transfers between your own accounts should not become revenue or costs. Taxes (income tax prepayments, VAT payments) should be clearly labelled so your reports remain interpretable. If your system can’t distinguish “payment to Belastingdienst” from “supplier cost”, you’ll keep fighting your own reports.

A lightweight fix is a monthly checklist:

• review exception queue


• attach missing invoices/receipts


• confirm payment provider payouts are reconciled


• confirm transfers and tax payments are classified correctly

You can estimate ROI without a complicated model. Use three buckets: time, cash, and risk reduction.

1) Time savings

Start with your baseline. Many freelancers spend 16–24 hours/month on bookkeeping when they include receipts and quarter-end VAT work. If AI reduces that to 4–8 hours/month, your time saved is 8–20 hours.

A simple formula:

• Time value per month = hours saved × your hourly value

• 12 hours saved × €75/hour = €900/month of time value.

Compare your current bookkeeping cost to an AI subscription.

Example:

• Accountant €500/month → €6,000/year


• AI bookkeeping €99/month → €1,188/year


• Cash difference: €4,812/year

This is harder to price, but it’s real:

• Missing receipts often means you miss reclaimable VAT.


• Running quarter-to-date totals reduces the chance of deadline surprises.


• An audit trail reduces the time cost if Belastingdienst asks questions.

The practical takeaway: if automation saves even 8 hours/month, it often pays for itself before you account for cash savings. If it saves 12–16 hours/month, it becomes one of the highest ROI operational improvements a freelancer can make.

Two ROI examples (so you can sanity check your numbers)

1) Freelancer / consultant

• 150 transactions/month


• Manual bookkeeping time: 18 hours/month


• With AI: 6 hours/month (exception review + receipts)


• Time saved: 12 hours/month


• Hourly value: €75/hour

2) Small webshop / agency

• 800 transactions/month (payment provider payouts + supplier invoices)


• Manual bookkeeping time: 45 hours/month


• With AI: 15 hours/month (exceptions + payment provider reconciliation)


• Time saved: 30 hours/month


• Hourly value: €50/hour

These aren’t promises; they’re realistic ranges we see when the receipt workflow is in place. Without receipts, time savings collapse because you’re still hunting evidence.

If you want to turn this into a one-line spreadsheet, use:

• (hours_manual - hours_ai) * hourly_value + (current_monthly_cost - ai_monthly_cost)

For many Dutch businesses, the best setup is hybrid.

• Use AI for continuous bookkeeping: bank feeds, receipts, categorisation, VAT totals, anomaly flags.


• Use a human accountant for what humans are still best at: annual filings, complex edge cases, and tax planning.

• you operate a BV


• you have employees or payroll obligations


• you have significant cross-border VAT exposure


• you want proactive planning (not just compliance)

• VAT overview and exception list


• profit-and-loss summary


• attachments linked to transactions

A hybrid workflow that works in practice

If you keep an accountant, agree on responsibilities upfront:

• AI bookkeeping handles categorisation, receipts, and quarter-to-date VAT totals


• You review exceptions (private/business splits, missing invoices)


• Your accountant reviews quarter-end VAT and prepares the annual return

• VAT overview (payable/refundable)


• exception list (reverse charge, missing evidence, unusual items)


• profit-and-loss summary


• attachment bundle or links tied to transactions

If you don’t use an accountant, the same pack is still useful: it’s your own quarterly sanity check before you file.