Brain-Dead Xero Machine Learning by Design, Part 4 of 5

Filed in Accounting, Business by on January 15, 2018
Accounting professionals (Certified Public Accountants, Chartered Accounts, and professional bookkeepers) should especially relate to this Brain-Dead Xero Machine Learning by Design post.

Xero uses “a tuned machine learning model for every customer” (company). As you will see, this is a terrible accounting Machine Learning approach. Unfortunately, it is the approach Xero must use.

Intuit CEO Brad Smith recently said, “… if a small business works with an accountant, their odds of success go up 89%.” He did not say this about Machine Learning, but it is critical to Brain-Dead Xero Machine Learning. Accounting professionals may have more than 5,000 client companies. Barring business differences, we each try to use the same accounting programs and the same basic charts of accounts for most companies. This tends to let us use one tax program, with minimal changes. It also lets us compare companies to spot mistakes, analyze industry operating or financial ratios and provide combined financial statements.

We also try to assign transactions to accounts consistently. Consistency is one of the most important international accounting standards. Auditors must report on inconsistent material deviations from accepted accounting principles. However, Brain-Dead Xero Machine Learning by Design uses a separate tuned machine learning model for each company, which only leads to consistent account assignments in each company. The Xero separate Machine Learning model for each company results in grossly excessive duplicate work and countless account assignment differences between companies.

Accounting professionals, lenders, and investors have long wanted consistency between company financial statements, so they can better compare companies to each other or standards. Lenders, in particular, are increasingly using automated account analysis to decide on credit. Xero may rightly say that manual or Machine Learning corrections can provide consistency between companies. It also may say that inconsistencies between companies will be no worse than they are now. However, for the first time in 5,000 years, new technology gives us the ability to quickly and easily automate consistency between companies. Despite this, Xero has no master (global) Machine Learning rules and does not let us copy manual and Machine Learning rules between companies. It rejected a Xero add-on that did this four years ago. To the contrary, QuickBooks Online has long done this, quickly and easily.

Xero also may rightly say that many accounting professionals will not care much about the level of consistency between company Machine Learning account assignments. However, separate Xero Machine Learning models for each company will make its many accounting professionals make each of their countless manual or Machine Learning rule changes repeatedly, for each of their many companies. That is, they must correct each change to each company for telephone, for utilities and for each of many auto expenses. This will continue indefinitely, as it does now. Each change will involve a manual transaction change (or the same manual change to many instances of one payee), as well as a hopefully accurate automatic Xero Machine Learning rule changes. It will be faster, easier and more accurate if accounting professionals work at a global level, or quickly, easily, and selectively copy rules between companies.

QuickBooks Online, unlike Xero, also has long let us copy rules to and from Excel. This lets us instantly sort rules by description, account and other fields, so we can see and correct many more errors quickly (vs reviewing one rule at a time in Xero). Sorts also let us drop complete duplicates, which we will otherwise have if we combine rules from many companies. Excel also is very good at catching errors because we can see many company rules at once, knowing they affect many clients, and can easily review the work of assistants.

Xero wrote about new users using Machine Learning without realizing that these users are most hurt by separate EMPTY Brain-Dead Machine Learning for each company. New users and accounting professionals should not use Machine Learning, with 80% accuracy after four passes through a few income accounts (not far more complex payments). To avoid this, users should start as we do now, with a simple standard chart of accounts. Google has 632,000 links for charts of accounts. Xero and QuickBooks Online also have charts of accounts. These average less than 150 accounts each. Unfortunately, Brain-dead Xero Machine Learning writes about the 10 million account CODES it considered in all Xero companies https://www.xero.com/blog/2017/03/teaching-robot-accounting/. This may be one reason why Xero Machine Learning seems slow.

Xero Machine Learning would be far less Brain-Dead if it disregarded account codes. Here is why. A million Xero users may assign phone expense to countless account codes, but most have all phone expense in Telephone & Utilities, Telephone Expense or Utilities accounts. This also applies to most accounts and payees. Therefore, Machine Learning should focus on key parts of account names, like phone or utilities, not account codes. However, an item description speed problem is stopping Xero from beiing fast. The Xero CTO has this Official Xero Reply a threads on Global Seach and Replace), “… released Xero Search (2015)… going to be a journey to get it right… investigating line item description search… we turned it off because it seriously bloated our index size.” This major Xero search problem is critical to Brain Dead Xero Machine Learning.

There is a critical reason why Xero Machine Learning should concentrate on account description and payee names: it is much easier to use Machine Learning rules of one company with other companies, despite differences in account CODES. My Xero keyboard macro add-on was slower than most conventional programs, but it did this four years ago (not for sale, see Part 5). It saved ridiculous amounts of time, while vastly increasing accuracy and consistency, compared to creating a new set of rules for each company. We also have long been able to do this with QuickBooks Online.

We should soon see massive extra benefits from combined QuickBooks Online Machine Learning rules: standard accounting professional, local, regional and national rules. Why should the hardest part of accounting program use relate to install and account assignments, training, and supervision? Professional accountants should be eager to publish, sell or give away carefully coded and branded (or even copyrighted) complete sets of Machine Learning rules to new or potential clients. QuickBooks Online and other accounting program sellers should be especially eager to bundle these with companion charts of accounts. This should save ridiculous amounts of user and professional time, while immeasurably increasing accuracy.

Carrying this only a bit further, QuickBooks Online already uses a series of questions to customize its chart of accounts. It would be trivial to use this approach to create many immensely powerful custom Machine Learning rules, quickly and easily. Professional accountants also could create these simple scripts to ask about telephone, electric or water companies and change each corresponding Machine Learning rule. Other questions can ask many other questions about common payees. It also would be very easy to have questions work with dropdowns, keyed to a local user country, state, city or zip code. Alternatively, the programs could use Excel to easily create or modify custom rules. Of course, accounting professionals could easily use Excel to review changes for new, converting or Machine Learning upgrade users.

There is one more amazing way that QuickBooks Online users can quickly make all clients better at account assignments. It relates to MCC codes. These Merchant Category Codes “are numbers that classify businesses by what they sell or the service they provide. These four-digit codes are assigned by payment card organizations—Visa, MasterCard, American Express and Discover—when the business first starts accepting these methods of payment.” There are about a thousand such codes. They appear as MCC= on more than 60% of imports from banks and credit card companies (about 88% by dollar value). My old Xero add-on used an Excel table of codes, with accounts for each code. Xero rules then assigned corresponding accounts to each entry. Unfortunately, Xero (unlike QuickBooks Online) still cannot import Excel tables for Machine Learning.

Xero and QuickBooks Online also do not yet distribute standard Machine Learning rule sets, much less offer scripts to easily customize them. They also both lack a transaction field, to show which rule assigned each transaction to an account. Such a field would be great for debugging.

As a one-time programmer, programmer supervisor and the only two-time winner of the #1 QuickBooks Top Tester award, I am beginning to create this for QuickBooks Online. Any professional accountant or programmer can do this, so please contact me if you are interested.

For more soon, please subscribe to the QuickBooks-Blog.

For related posts, please click on:
1. Horribly Brain-Dead Xero Machine Learning, programmers are misdirected, slow and incompetent.)
2.
Brain-Dead Xero Machine Learning will persist, Xero culture denies problems instead of fixing them
3.
Brain-Dead Xero Machine Learning ridiculously incomplete
4. Brain-Dead Xero Machine Learning by Design (this post)

About the Author ()

Fastest, easiest, accurate, low-cost QuickBooks! Intuit - QuickBooks CEO, "You're fantastic Mike! Absolutely fantastic!" Tax fighting CPA.

Comments are closed.

Skip to toolbar