Date posted: 30/10/2018 10 min read

How cheap prediction can change accounting

Artificial intelligence may be best understood as “cheap prediction”. And prediction is a useful input to decisions – but people will still be needed to make the decisions.

In Brief

  • Three academics in Canada believe that machine learning of today is good at one thing – offering a way to make cheap predictions.
  • One of the latest machine learning applications is Google’s word processor for Google Docs that automatically corrects your grammar in real time.
  • Forensic accounting and auditing is already making way for algorithm-driven programs that process huge volumes of transactions and can check for fraud or error.
  • “Auditors will eventually veer towards the forensic accounting, accuracy, validation type of role rather than sitting with Excel spreadsheets trying to manually reconcile thousands of transactions,” says Gavin Whyte, chief data scientist at Deloitte Australia.

In the mid-1990s an economist called William Nordhaus had a radical idea for valuing the invention of the light bulb. How much effort, he wondered, would it take to produce a similar amount of light with a wood fire?

The answer: to produce an hour of electric light with a light bulb would require chopping wood for 10 hours a day, for six days. Nordhaus went on to create a price index going back to sesame oil-powered lamps from Babylonian times, showing that the real benefit was a dramatic fall in the cost of artificial light.

Researchers have since used this economic approach for valuing technology to examine the internet's role in lowering the cost of search; transporting, verifying and replicating information; and in tracking behaviour.

And now three academics in Canada have followed this process to cut through the hype around machine learning, the most popular example of artificial intelligence today. "Digging into the technology it became clear that it was a drop in the cost of prediction," says Professor Avi Goldfarb, an econometrician who specialises in the science of quantitative marketing. With Professors Ajay Agrawal and Joshua Gans he has co-authored the book Prediction Machines, the simple economics of artificial intelligence.

Not everyone sees machine learning in such a reductive light. Computer science academics emphasise the potential of AI's ability to learn, and the ramifications this holds for training robots and the like.

"But that's not the AI we have today. It hasn't gone past that one thing – cheap prediction," says Gans, a renowned Australian economist who moved to Toronto in 2010.

Prediction explained

Prediction is the process of filling in missing information. It takes the information (or data) you have, and uses it to generate information you don't have.

Given the great advances in artificial intelligence, calling it "cheap prediction" seems a little underwhelming. It suggests we haven't yet reached a drop in the cost of intelligence; we're only reducing the cost of one part of it. Yet this part is a critical step. Machine learning's probabilistic model mimics our own learning process, a process that developed through millennia of evolution. Prediction, argues another author, Jeff Hawkins, is the basis for human intelligence. "Prediction is not just one of the things your brain does. It is the primary function of the neocortex, and the foundation for intelligence. The cortex is an organ of prediction," Hawkins wrote in his book, On Intelligence.

Prediction Machines explains that machine learning is not on its own a tool to replace professionals; it is merely a tool for improving prediction. And prediction is one of several inputs into the process of decision making, the authors argue. Another is that undervalued input called judgement.

"Prediction facilitates decisions by reducing uncertainty, while judgement assigns value," the authors write. Luckily for accountants, value is a difficult thing for machines to assess. Machine learning may speed up the process of making predictions by categorising and sorting data and spotting patterns. But turning those lessons into business advice, and prioritising them in terms of success, requires analysing a combination of emotional, intellectual and practical considerations.

We've got a lot to do with the AI we currently have and that's going to keep people occupied for the moment.
Joshua Gans Professor of Technical Innovation and Entrepreneurship at the Rotman School of Management

"In economists' parlance, a judgement is the skill used to determine a payoff, utility, reward or profit," the authors explain. "The most significant implication of prediction machines is that they increase the value of judgement." "You can appreciate what else people do to make a decision," Gans says. "They can't just predict things. They also have to know what the trade-offs are, and these things only come from people. Then you start to understand why it's really hard to create a fully automated thing, because we may not understand the nature of decisions that the robot needs to make."

When will AI move beyond cheap prediction to making judgements? Few agree on the timeline for a breakthrough of that magnitude; the predictions range from imminent to almost never. "Someone might switch on a robot AI that works it out itself and just becomes sentient. I'm not a computer scientist so I can't give you a probability, but my feeling is that it's not for a long time," Gans says. "We've got a lot to do with the AI we currently have and that's going to keep people occupied for the moment."

Cheap prediction plus accounting equals better forecasts

A wave of machine learning applications is breaking across the business world. One of the latest is Google's word processor for Google Docs that automatically corrects your grammar in real time. AI is being quickly built into other programs, from SME accounting software to ERP. But Gans cautions against believing everything you hear.

"I wrote the book because I was concerned that people would say, 'Buy my magic AI!' and it would turn out to be not that good," he says. "I don't think there's any need to rush to add it to your operations."

The use case for accountants in practice is more clear-cut.

"Accounting does have data going for it, so it's only a step away from being put to use," Gans says. Forensic accounting and auditing is already making way for algorithm-driven programs that process huge volumes of transactions. These programs can pull up a shortlist of transactions to check for fraud or error (see below).

Blue J Legal, a Toronto company, has partnered with Thomson Reuters to produce a program that predicts how a court will rule in new tax situations by analysing case law. The program was trained by computer scientists and law professors from the University of Toronto to recognise hidden patterns in legal decisions. Blue J Legal provides answers, links to relevant cases, and a tailored explanation of its analysis.

"I don't think it's going to displace the accountants any time soon but I think it will make their job a lot easier," Gans says of the technology.

Will accountants will be replaced by machines? Goldfarb believes this is unlikely. Fifty years ago, accountants spent most of their time doing arithmetic. Training focused on calculating sums without making mistakes. When the spreadsheet arrived it dramatically lowered the cost of doing arithmetic and helped customers make decisions. Before spreadsheets arrived, one would have expected the arrival of such a powerful decision- making tool to reduce the need for accountants. "But the numbers have remained steady," Goldfarb says.

"Most of the tasks that accountants do today they will not be doing in 10 to 15 years from now. That doesn't mean we won't have lots of accountants, because these tools will enable accountants to better serve clients which will open up new opportunities," Goldfarb says.

His advice to accountants worried about the pace of change is to spend more time understanding their clients' needs.

"Increasingly, it's not going to be enough to do the same thing that everyone else does in a standardised way for all clients. Knowing your clients better and knowing how to serve them better is going to be a key part of new opportunities."

And of course, accountants should understand the capability of machine learning and other technologies to improve their clients' businesses. There's no need to learn about the mechanics of machine learning any more than you need to understand binary maths to use a computer, he adds.

It's worth noting that Gans, an economist, isn't rushing to make predictions powered by machine learning. He is loathe to make economic predictions, he says, because economists are "so bad at it". Even though machine learning can use an element of brute force – for instance, when uploading and processing 10,000 pictures of labelled dogs and cats to understand the factors that differentiate them – there is still "a lot of art" to good predictions, Gans adds. Translating Greek to Portuguese or knowing a Saint Bernard from a Dachshund is relatively easy because there are few variables.

Economic variables are far more complex. "I don't know if they are predictable," he says. "There might be too much randomness. These AIs can't break the laws of statistics. They can do better but they can't change them."

Using AI to automate audits

Accountants are already using machine learning software to audit accounting files in minutes. And the task of poring over spreadsheets to match transactions looks like one of the first to fall under the wheels of automation.

Radlee Moller was at a partner retreat in Hawaii when he first realised the opportunity for machine learning.

Managing partner at CA firm CIB Accountants and Advisers, in Parramatta, NSW, Moller was intrigued at claims made by a software company, MindBridge Ai Auditor, that it could automate most of the legwork for auditors.

Moller invited MindBridge CEO Solon Angel to Australia and watched Angel run the "Pepsi challenge". The software took five minutes to audit a file and find the four mistakes within that had taken a five-person team three weeks. It even revealed a fifth, unknown error.

Muller timed the software on other files at CIB. "It took 12 minutes for the biggest file" in the firm, Moller says.

Several months later, Moller had convinced Angel to let him distribute MindBridge to Australian firms.

The startup is barely a year old yet it is already making millions in revenue, has 120 customers – including the Bank of England – and is preparing for an IPO in 2021.

CIB Accountants hasn't dropped its fees for audits, despite the time saved. "I tell clients there's a software cost. We don't pass that on; we wear it," Moller says.

Moller's experience fits the prediction made by Deloitte Australia in a report last year. It identified auditing as the most likely role for automation.

"Auditors will eventually veer towards the forensic accounting, accuracy, validation type of role rather than sitting with Excel spreadsheets trying to manually reconcile thousands of transactions," says Gavin Whyte, chief data scientist at Deloitte Australia.

Whyte has been developing in-house algorithms that replicate MindBridge's smarts. The Big Four firm can customise them for different clients or applications, Whyte says.

Prediction Machines, the simple economics of artificial intelligence, by Ajay Agrawal, Joshua Gans and Avi Goldfarb, is available at the CA ANZ Library both in hardcopy and eBook. To reserve or download visit

Prediction Machines book

Read a recent paper by CA ANZ, Machines can learn, but what will we teach them?

Sholto Macpherson is a technology journalist and editor of, a blog on the latest in accounting technology for firms and SMEs.