This Is Your Accounting Ledger On AI

Your ledger is about to get a lot smarter.

If you built an accounting ledger from scratch tomorrow - something built around the AI development of the past 6 months - it would look very different to incumbent tools.

Whether incumbents adapt to the AI developments of the last 6 months, or a new AI-first crop of ledger tools will emerge remains to be seen. But one way or another change is on the way, and it can't come soon enough.

The current crop of ledger apps rely on rules-based classifications. The secret to making an accounting file sing today is a sweet set of bank rules.

And for all the promises ledgers apps have made around using machine learning and proprietary data to make classification easier than ever, you still make us tell you the Chevron purchase is a Fuel Expense.

But thanks to a vastly improved embeddings model released by OpenAI in January, all of this is about to change.

"Vector similarity search" uses embeddings to determine the semantic similarity of two pieces of text. "feline friends say" is more similar to "meow" than "a quarterback throws a football."

A basic accounting example here from a great Digits write-up:

1. Take a bank statement description, a purchase from Starbucks
2. Embed the text
3. Compare the embedding to all past embeddings, ie transactions that have already been classified
4. See how the most similar embeddings were classified
5. Voila, if within a certain distance, auto-classify the Starbucks transaction

The result is a clustering of transactions according to the similarity of the text embedding:

In the past, machine learning models required thousands of examples to produce a semi-accurate prediction. Today, classifying a couple Starbucks transactions to Office Expense (I'm kidding (am I kidding?)) is all it takes to classify the next Starbucks transaction.

It means no more bank rules.

It means managing exceptions instead of each transaction.

It means the classification model is the accounting file's unique fingerprint.

It means classification models are now transferrable between clients.

It means the new accounting file with your standard chart of accounts can classify activity from transaction zero, thanks to the model you created in the other construction company's file.

It means classifying a new type of transaction in company A makes the classification model behind similar company B even smarter.

It means you simply connect a bank account, and thousands of transactions are classified nearly instantly.

What's at the core of an accounting system now, if it's not transactional classification? I'm not sure, but probably something better.

The proprietary data sets of incumbents don't take into account three things:

1. Different accountants will classify the same transaction differently
2. Chart of accounts are not standardized across users
3. Oftentimes you need more context than just the bank statement line to make a correct classification

And it's why vector similarity search wipes out that proprietary data moat.

Vector similarity search isn't new - but it's newly cost-effective, and now more technically accessible to developers. Long-term it's something everyone will implement in some capacity, and will be the secret sauce behind the best practice management systems.

Ultimately the more context we can include in the text embedding, the more accurate the classification will be. Receipt, invoices, purchase confirmations.

If you're a developer looking to build the best auto-classification model of tomorrow, give us the best reason to put our context on your platform today, and you're in the driver's seat to take a big step closer to fully-automated bookkeeping.