How Penny Categorises Transactions Accurately

The Problem with Silent Auto-Categorisation

Every bookkeeping app on the market offers some form of auto-categorisation. Your bank feed comes in, the software looks at the merchant name, and assigns a category. "TESCO STORES" becomes "Stock and materials." "SHELL FORECOURT" becomes "Travel costs." Job done.

Except it's not. Because that Tesco purchase might have been your weekly shop, not business stock. That Shell transaction might have been a personal trip, not a business one. And that £45.00 payment to "J. Smith" could be a subcontractor, a professional service, or a birthday present for your mate John.

Most bookkeeping apps handle this by picking the most likely category and moving on in silence. If you spot the mistake, you fix it. If you don't, your records are wrong -- and you might not find out until your accountant reviews, or worse, until HMRC asks questions.

Penny does things differently. She doesn't just categorise -- she tells you how confident she is about every single decision.

How Penny's Categorisation Engine Works

When a new transaction arrives from your bank feed, Penny doesn't simply match the merchant name against a lookup table. She runs it through Accounted's three-tier reasoning engine, which analyses multiple signals simultaneously:

Signal 1: Merchant Recognition

Penny maintains a comprehensive database of UK merchants mapped to HMRC expense categories. She knows that Screwfix sells building materials, that EE is a phone provider, and that Pret A Manger is a food outlet. This handles the straightforward cases.

Signal 2: Your Transaction History

Over time, Penny learns your specific patterns. If you regularly buy from a merchant and always categorise it the same way, she applies that knowledge to future transactions. Your personal history is more reliable than generic rules because it reflects your actual business.

Signal 3: Amount Patterns

The transaction amount provides surprisingly useful context. A £3.50 transaction at a cafe is probably a coffee (and probably personal). A £35.00 transaction at the same cafe with a note mentioning a client's name is more likely a business meeting. Penny considers these patterns.

Signal 4: Timing Context

When and how frequently a transaction occurs matters. A monthly payment of exactly £49.99 on the first of the month is almost certainly a subscription. A one-off payment at 11pm on a Saturday is more likely personal. Penny factors timing into her assessment.

Signal 5: Industry Norms

Penny knows that electricians regularly buy from Screwfix, that web developers frequently pay for SaaS subscriptions, and that tutors buy educational materials. Your industry context helps disambiguate borderline cases.

All five signals feed into a composite confidence score for every categorisation.

The Confidence Scoring System

This is where Penny genuinely differentiates from other bookkeeping tools. Every categorisation receives a confidence score expressed as a percentage. What happens next depends on that score:

High Confidence (95% and above)

When Penny is 95% or more confident, she categorises automatically without bothering you. These are the clear-cut cases: your monthly phone bill, your regular fuel purchases from the same station, subscription payments that arrive like clockwork.

For most sole traders, around 80-85% of transactions fall into this band. They're categorised instantly and accurately. You can review them at any time, but no action is needed from you.

Medium Confidence (70-94%)

When Penny is fairly sure but not certain, she reaches out via WhatsApp:

Penny: I've categorised a £34.50 payment to Tesco as "Stock and materials" (confidence: 78%). Does that look right, or was this personal?

One tap to confirm. One tap to correct. Three seconds of your time.

These typically represent 10-15% of your transactions. They're the cases where you shop at a merchant for both business and personal, or you've used a new supplier for the first time.

Low Confidence (Below 70%)

When Penny genuinely doesn't know, she doesn't pretend:

Penny: I saw a payment of £125.00 to "R. Davies" on 15 February. I'm not confident about this one. Could you tell me what it was for?

1. Subcontractor payment
2. Professional services
3. Stock and materials
4. Something else

No silent miscategorisation. No wrong category sitting unnoticed in your records. These typically make up just 3-5% of transactions -- the genuinely ambiguous ones where getting it wrong would cause problems.

Why 95% Accuracy Matters

Let's put this in real terms. Say you have 50 business transactions per month. With 95% accuracy:

• 47-48 transactions are categorised correctly without you doing anything
• 2-3 transactions are flagged for your quick review
• 0 transactions are silently wrong

Compare that to a rules-based system with, say, 80% accuracy:

• 40 transactions categorised correctly
• 10 transactions potentially wrong
• 0 flags telling you which ones to check

With the rules-based system, you'd need to review every transaction to find the errors. Or trust the software and hope for the best. Neither is a good option when HMRC expects accurate records for your Making Tax Digital submissions.

The Learning Loop

Penny doesn't just categorise -- she learns. Every time you confirm or correct a categorisation, that feedback trains her understanding of your specific business.

Tell Penny that payments to "R. Davies" are subcontractor costs, and she'll remember next time. Confirm that Tesco purchases on weekday mornings are stock, and she'll apply that pattern going forward. Correct a category, and she adjusts her model.

The result is that Penny gets smarter the longer you use her:

• Month 1: You might review 15-20% of transactions
• Month 3: Closer to 8-10%
• Month 6: Often 5% or less

You're effectively training a bookkeeper who never forgets, never leaves, and improves every month. The ICAEW's guidance on AI in accounting highlights exactly this kind of continuous improvement as the future of automated bookkeeping.

Edge Cases Penny Handles Well

Real-world bookkeeping isn't just about straightforward expenses. Penny handles the awkward cases too:

Mixed-Use Purchases

Bought something that's partly business, partly personal? Tell Penny, and she'll split the transaction accordingly. "That Costco trip was about half business supplies, half personal" is enough for Penny to create the split and categorise each portion correctly.

Refunds and Returns

When a refund appears, Penny traces it back to the original transaction and reverses the categorisation. No orphaned refunds sitting in the wrong category.

Recurring vs One-Off

Penny distinguishes between regular payments and one-offs automatically. A new subscription is identified within two or three payments, and future instances are categorised with high confidence.

VAT Identification

For VAT-registered businesses, Penny identifies the VAT component of transactions where possible, especially when matched with receipt data. This feeds directly into your VAT returns.

How This Compares to Other Tools

Traditional bookkeeping software like Xero and QuickBooks relies primarily on bank rules -- simple "if merchant name contains X, categorise as Y" logic. These rules work well for recurring, predictable transactions but struggle with ambiguity.

Accounted's approach is fundamentally different:

| Feature | Traditional Apps | Accounted (Penny) | |---------|-----------------|-------------------| | Categorisation method | Rules-based | AI with multi-signal analysis | | Confidence visibility | None | Percentage on every transaction | | Uncertain transactions | Silently categorised | Flagged for human review | | Learning from corrections | Limited (manual rules) | Automatic, continuous | | Accuracy over time | Static | Improves with use |

The difference isn't marginal. It's the difference between trusting your records and hoping they're right. Read our detailed comparison with Xero to see the full picture.

What This Means for You

If you're a sole trader, this matters because your bookkeeping records are the foundation of everything: your tax returns, your MTD submissions, your understanding of how your business is performing.

Inaccurate records lead to overpaying tax (frustrating), underpaying tax (dangerous), or simply not knowing where you stand financially. Penny's confidence-scored categorisation means you always know your records are right -- not because the AI is perfect, but because it's honest about what it doesn't know.

That honesty is what makes the 95%+ accuracy figure meaningful. It's not a theoretical number -- it's what our users actually experience, and it improves the longer they use Accounted.

Try It Yourself

The best way to understand how Penny categorises transactions is to experience it. Sign up, connect your bank account, and watch Penny process your first batch of transactions. You'll see the confidence scores, the flagged items, and the learning in action.

View our pricing plans or start your free trial -- no credit card required.