Every conversation about AI in accounting eventually focuses on the models: which algorithm, which architecture, which training approach. But the factor that determines whether an AI system actually works in practice is not the model. It is the data.
Garbage in, garbage out is one of the oldest principles in computing. AI does not repeal this law. It amplifies it.
The Scale of the Problem
Financial data is messier than most people realise. Even in well-run businesses, the raw data flowing into accounting systems contains significant quality issues.
Duplicate vendors. The same supplier appears as "Malta Post," "MaltaPost," "MALTA POST Ltd," and "M. Post" across different bank transactions and invoices. To a human, these are obviously the same entity. To a data system, they are four different vendors.
Inconsistent naming. An employee expenses a meal at "Il-Horza Restaurant" one month and "Horza Rest." the next. A bank feed might record it as "POS IL-HORZA VALLETTA MT." Three different strings for the same transaction type at the same merchant.
Missing fields. An invoice without a VAT number. A bank transaction without a reference. A receipt without a date. Each missing field reduces the system's ability to categorise, match, and reconcile.
Wrong currencies. A payment recorded in euros when it was actually in pounds. A conversion rate applied to the wrong date. Currency mismatches cascade through financial statements.
Mismatched dates. An invoice dated 31 December but the payment does not clear until 3 January. The bank statement shows one date; the supplier's records show another. Which date is used for accounting purposes depends on the accounting standard, the type of transaction, and the jurisdiction.
Truncated descriptions. Bank feeds often truncate merchant names and references. "AMAZON EU SARL LUXEMBOURG LU" becomes "AMAZON EU S" or just "AMAZON." The lost characters might have contained the information needed for accurate categorisation.
Why AI Models Struggle with Dirty Data
Machine learning models learn patterns from data. If the data contains inconsistent patterns, the model learns inconsistency.
Consider a transaction categorisation model. If the training data contains "Malta Post" classified as "Postage & Courier" in 80% of cases, "Office Supplies" in 15% (because someone included the cost of packaging materials), and "Miscellaneous" in 5% (because someone gave up trying to categorise it), the model inherits this confusion.
The model does not know that "Office Supplies" was wrong. It does not know that "Miscellaneous" was lazy. It treats all the training labels as equally valid and builds a probabilistic representation of the category distribution.
At scale, these quality issues compound. A model trained on millions of transactions from thousands of businesses can absorb and average out individual errors. But systematic biases, categories that are consistently misapplied across the industry, become baked into the model.
Data Cleaning and Normalisation
Before any AI model touches financial data, that data needs to be cleaned and normalised. This is not glamorous work, but it is essential.
Vendor normalisation involves identifying that "Malta Post," "MaltaPost," and "MALTA POST Ltd" are the same entity and mapping them to a single canonical name. This can itself be partially automated using fuzzy matching algorithms and entity resolution techniques, but it requires a maintained master list and human oversight for edge cases.
Description standardisation means parsing the often cryptic strings that come from bank feeds and extracting meaningful information: merchant name, location, transaction type, and reference number. Companies like Plaid and MX have built entire businesses around enriching raw banking data into standardised, meaningful records.
Deduplication identifies and merges duplicate records. This includes both exact duplicates (the same transaction imported twice) and semantic duplicates (the same economic event recorded from two different sources, such as a bank transaction and a matching invoice payment).
Gap filling addresses missing fields. If an invoice is missing a VAT number, can it be inferred from the vendor master data? If a transaction date is missing, can the bank posting date be used as a reasonable proxy?
Master Data Management
At the heart of data quality is master data management (MDM): maintaining authoritative, clean, and consistent reference data for the entities your accounting system tracks.
The two most critical master data sets in accounting are:
Vendor master. A single, clean list of every supplier and counterparty, with their correct legal name, VAT number, payment terms, and default accounting category. Every transaction involving that vendor maps to a single master record.
Chart of accounts. A structured, unambiguous list of accounting categories with clear definitions and rules for what belongs in each. If there is no clear definition of what qualifies as "marketing expense" versus "business development" versus "client entertainment," the categorisation will be inconsistent no matter how good the model is.
Maintaining master data requires ongoing effort. New vendors appear. Existing vendors change their names (through mergers, rebranding, or simply updating their bank details). Categories need to evolve as the business changes. This is not a one-time cleanup; it is a continuous discipline.
The Feedback Loop
The good news is that AI and data quality can form a virtuous cycle.
AI models are excellent at identifying data quality issues. They can flag potential duplicates, highlight inconsistent categorisations, detect outliers that may indicate data entry errors, and identify transactions that do not match expected patterns.
When humans review and correct these flagged items, the corrections serve dual purposes: they fix the immediate data quality issue, and they provide new training data that improves the model's future performance.
Over time, this feedback loop produces both cleaner data and more accurate models. But it requires the human review step. Without it, the model's errors propagate uncorrected, and data quality degrades rather than improves.
The Uncomfortable Truth
Here is the reality that AI vendors often gloss over: the best model in the world cannot compensate for fundamentally messy records.
If a business has three years of accounting data where personal and business expenses are mixed, vendor names are inconsistent, receipts are missing, and categories were assigned arbitrarily, no amount of AI sophistication will produce accurate financial statements from that data.
The AI might produce financial statements that look plausible. The totals might seem reasonable. But "plausible" and "accurate" are not the same thing, and in accounting, only accuracy matters.
What Good Data Quality Looks Like
For a self-employed professional, good data quality is achievable and does not require sophisticated systems:
One bank account for business, one for personal. The single most impactful step for data quality is not mixing personal and business transactions. This eliminates the largest source of categorisation ambiguity.
Consistent record-keeping. Use the same process for recording expenses every time. Whether that is photographing receipts immediately, entering transactions daily, or using an app that syncs in real time, consistency matters more than the specific method.
Timely processing. The longer a transaction sits unreconciled, the harder it is to remember its purpose. A charge from three days ago is easy to categorise. A charge from three months ago might be a mystery.
Complete documentation. Every transaction should have supporting documentation. Every invoice should be stored. Every receipt should be captured. When the AI (or the auditor) asks "what was this payment for?" the answer should be immediately available.
The Bottom Line
The AI revolution in accounting is real, but it rests on a foundation of data quality. Firms and individuals that invest in clean, consistent, complete financial records will benefit enormously from AI-powered automation. Those that do not will find that AI simply processes their mess faster.
The unsexy truth is that data quality discipline, not model architecture, is the single biggest determinant of whether AI delivers value in accounting.
Michael Cutajar, CPA — Founder of Accora.