Confidence Scoring on Every Categorisation

The Silent Miscategorisation Problem

Here's a question most sole traders never think to ask: when your bookkeeping software auto-categorises a transaction, how sure is it?

The answer, with most tools, is that you have no idea. Xero categorises a transaction as "Office expenses." QuickBooks assigns it to "Travel." FreeAgent labels it "Cost of sales." They all do it with the same quiet confidence, whether they're 99% certain or flipping a coin.

This is fine for the straightforward cases. Your monthly mobile phone bill is always "Telephone." Your regular fuel purchase is always "Travel costs." These are easy, and any system gets them right.

But what about the ambiguous ones? The Amazon purchase that could be office supplies or personal. The supermarket transaction that might be stock or your weekly shop. The payment to a person's name that could be a subcontractor, a consultant, or a gift. When the software guesses these, it doesn't tell you it's guessing. It just categorises and moves on.

Over a year, these silent miscategorisations add up. A few hundred pounds in the wrong category here, a personal expense counted as business there. Your records drift from reality, and you don't know which entries to trust.

Accounted's confidence scoring solves this completely. Every single categorisation Penny makes comes with a percentage score telling you how certain she is.

What Confidence Scoring Is

Confidence scoring is a percentage attached to every transaction categorisation. It tells you, in plain terms, how sure Penny is that she's got the category right.

The score is calculated by Accounted's three-tier reasoning engine, which analyses multiple signals: merchant recognition, your personal transaction history, the amount, the timing, and what other sole traders in your industry categorise similar transactions as.

The score isn't arbitrary. It's a calibrated probability. When Penny says she's 95% confident, she's right about 95% of the time. When she says 70%, she's right about 70% of the time. This calibration is tested and maintained -- it's not just a number to make you feel good.

The Three Confidence Bands

What happens with each transaction depends on its confidence score:

High Confidence: 95% and Above

These transactions are categorised automatically. No notification, no WhatsApp message, no action needed from you.

Your monthly phone bill? 99% confidence. Your regular fuel purchase from the same station? 98% confidence. Your Adobe Creative Cloud subscription on the first of every month? 99% confidence.

These are the transactions where the merchant is well-known, the amount is consistent, and the category is unambiguous. For most sole traders, 80-85% of transactions fall into this band from the first month, rising to 90%+ after a few months of learning.

The time saving is enormous. If you have 50 transactions a month and 85% are high confidence, that's 42-43 transactions you don't need to look at. They're processed accurately in the background while you get on with your actual work.

Medium Confidence: 70-94%

These transactions are categorised, but Penny flags them for your quick review via WhatsApp:

Penny: I've categorised a £27.50 payment to WH Smith as "Office supplies" (confidence: 81%). Is that right?

You: Yes

Penny: Great, confirmed.

Or:

You: No, that was a birthday card -- personal

Penny: Understood -- removed from business expenses. I'll remember that WH Smith purchases aren't always business.

One tap to confirm. One message to correct. Three seconds either way.

Medium confidence transactions typically represent 10-15% of your total. They're the ones where you use a merchant for both business and personal, where the amount is unusual, or where the merchant is unfamiliar.

The key insight: Penny tells you she's not fully sure. She offers her best answer and asks you to verify. This is exactly what a good human bookkeeper would do.

Low Confidence: Below 70%

When Penny genuinely doesn't know, she asks directly:

Penny: I saw a payment of £85.00 to "Northern Services" on 24 Feb. I'm not sure about this one. What was it for?

1. Professional services
2. Repairs and maintenance
3. Cleaning
4. Something else

No guess. No assumption. Just an honest question.

Low confidence transactions typically make up 3-5% of the total. These are the genuinely ambiguous ones -- unfamiliar merchants, unusual amounts, transactions that don't match any known pattern.

Why This Matters for Your Tax Return

Your bookkeeping records are the foundation of your tax return. Under Making Tax Digital, they're also the foundation of your quarterly submissions to HMRC. If your categories are wrong, your tax figures are wrong.

Consider what happens with silent miscategorisation:

• A £500 personal expense counted as business: Your taxable profit is understated by £500. At 20% basic rate, that's £100 of underpaid tax. At 40% higher rate, it's £200.
• A £300 business expense missed or categorised as personal: Your taxable profit is overstated by £300. You're paying £60-120 more tax than you should.
• Over a full year: If just 5% of transactions are silently miscategorised with an average error of £100 each, that's meaningful money -- either overpaid or underpaid.

With confidence scoring, silent miscategorisation doesn't happen. Every uncertain transaction is flagged. Your records are accurate not because the AI is infallible, but because it's transparent about its limitations.

The Learning Effect

Every time you confirm or correct a categorisation, Penny's confidence model updates. This creates a virtuous cycle:

1. Penny categorises with an initial confidence score
2. You confirm or correct the medium and low confidence items
3. Penny's model updates -- future similar transactions get higher confidence
4. Fewer items need your review over time

In practical terms:

| Month | High Confidence (auto) | Medium (flagged) | Low (asked) | |-------|----------------------|------------------|-------------| | Month 1 | 80% | 15% | 5% | | Month 3 | 88% | 9% | 3% | | Month 6 | 92% | 6% | 2% | | Month 12 | 95%+ | 4% | 1% |

By month 12, you're reviewing perhaps two or three transactions a week. The rest are handled automatically with high confidence. Your bookkeeping takes minutes per month, and every categorisation is either verified by you or confident enough to trust.

Confidence Scoring in Practice

Let's walk through a realistic week for a self-employed electrician:

Monday: 3 transactions processed

• British Gas commercial account: £145.00 -- "Subcontractor costs" (98% confidence) -- auto-categorised
• Screwfix: £67.82 -- "Stock and materials" (99% confidence) -- auto-categorised
• Costa Coffee: £4.50 -- "Subsistence" (72% confidence) -- Penny asks: "Was this a business meeting or personal?"

Wednesday: 2 transactions processed

• Direct debit to EE: £35.00 -- "Telephone" (99% confidence) -- auto-categorised
• Payment received from client: £850.00 -- "Turnover" (99% confidence) -- auto-categorised

Friday: 2 transactions processed

• Amazon.co.uk: £34.99 -- not enough information (55% confidence) -- Penny asks what it was for
• Shell fuel: £65.00 -- "Travel costs" (96% confidence) -- auto-categorised

Total time spent on bookkeeping that week: about 30 seconds answering two WhatsApp messages. Seven transactions processed accurately with full confidence transparency.

No Other Tool Does This

We've examined every major bookkeeping tool available in the UK. Xero, QuickBooks, FreeAgent, Sage, Coconut -- none offer transaction-level confidence scoring. They all auto-categorise, but none tell you how confident the system is about each individual decision.

Some offer accuracy percentages at the system level ("95% auto-categorisation rate"), but that's a different thing entirely. A system-level accuracy rate tells you that, on average, 95% of transactions are correct. It doesn't tell you which 5% are wrong. You'd still need to check everything to find the errors.

Confidence scoring tells you exactly which transactions need attention and which ones you can trust. That's a fundamental difference in how you interact with your bookkeeping -- and in how accurate your records end up being.

The Connection to MTD Compliance

Under Making Tax Digital, your quarterly submissions must be based on accurate digital records. Confidence scoring ensures those records are accurate by preventing uncertain categorisations from going unreviewed.

When Penny prepares your quarterly submission, she's working with data that's been confidence-scored and verified. The uncertain items have been checked by you. The confident items have been validated by the model. The result is a submission you can trust.

The importance of transparency in AI systems is well-documented. The UK government's framework for automated decision-making emphasises that users should understand when and why AI is making decisions on their behalf. Confidence scoring is our implementation of that principle.

Read about how Penny uses confidence scoring as part of her categorisation process, or how the underlying reasoning engine generates confidence scores.

See It for Yourself

The best way to understand confidence scoring is to experience it. Sign up for Accounted, connect your bank, and watch Penny categorise your first batch of transactions. You'll see the scores, understand the system, and appreciate why transparency matters.

Explore our pricing or read about all our features -- confidence scoring is included with every plan.

For further reading on AI transparency in financial services, see the FCA's guidance on AI in regulated industries.