The numbers sound promising.
PwC estimates that artificial intelligence (AI) could contribute up to $15.7 trillion to the global economy in 2030.
In Singapore, 86% of the 150 CFOs recruitment agency Robert Half surveyed earlier this year indicated that they are implementing the technology or have plans to do so in the next 12 months. The benefits those CFOs expect AI to bring include increased efficiencies and productivity (50%), better decision-making capabilities (46%), and enhanced processes (44%), the company noted.
A late 2018 Gartner survey found that 27% of the more than 400 companies polled planned to deploy AI for their finance functions by 2020.
But what is AI?
Enterprise- and consumer-technology companies attach the AI label to a wide range of products — among them chatbots, facial recognition, and even smartphone games that “predict” how you will look when you grow old. So, what is AI, and what would it allow users such as CFOs and their teams to achieve?
While AI is a broad term, it all comes down to data and getting a learning engine to do a better job, according to Eric Thain, co-chair and president at the Artificial Intelligence Society of Hong Kong and general manager of brand development at HK Express, a Hong Kong-headquartered airline.
As machines learn by trial-and-error and/or pattern recognition, data is the fuel for their learning process, Thain said in a phone interview. “The more the data, the better the learning engine works,” he noted.
In addition, machine learning is the natural progression of data analytics when it comes to AI, he said.
“While descriptive analytics tells us what the situation is, predictive analytics tells us about the future based on past data,” he explained. “And then there’ll be prescriptive analytics that tells us what to do when something happens. When narrowed down, AI and data analytics is machine learning at work.”
In Charting the Future of Accountancy With AI, a report by CPA Australia and the Singapore Management University School of Accountancy, machine learning — a subfield of AI and the driver of most of AI’s recent progress — is defined as the use of techniques that enable computers to learn and continually improve without being explicitly programmed.
“In essence, AI and machine learning (in their current form) consist of techniques which learn to recognise patterns in order to make predictions that facilitate decision-making,” the report said.
What can AI do for the finance function?
One way AI can help finance functions is to automate the extraction of data from documents such as invoices and receipts, according to Rahul Chandra, head of South East Asia and Middle East at Intain Fintech, an AI and blockchain solutions provider based in India, Singapore, and the US.
Context-based learning capabilities allow AI tools to recognise common features on different invoices even when those invoices have varied formats, he said in an interview with FM.
Thus, after successfully recognising and classifying a document such as an invoice, an AI tool can extract data such as invoice numbers, amounts payable, bank account numbers, and company names, even if they appear in different places on different invoices.
“Companies denote invoice numbers in different ways such as INV number, invoice number, or purchase number. An AI tool will eventually learn that they refer to the same thing, and there’s no need for employees to manually extract data from documents,” Chandra said.
How much time is required to train an AI tool to extract the correct data? The more invoices that a finance function uses to train an AI tool, the faster it learns, according to Chandra.
While the time required for AI tools to become highly accurate differs, Chandra said it takes one to four weeks for Intain Fintech’s AI tool to become 90% confident in its data extraction work.
Confidence is the probability with which an AI model predicts what it has extracted is accurate; confidence ranges from 0% to 100% — from not so confident to absolutely sure, according to Chandra.
He added that Intain Fintech’s AI solution allows a user to decide “the level of confidence” that requires human intervention, he said.
“For instance, a user can choose to get an alert if the AI tool only has 70% confidence in document recognition,” he explained.
Fully automated finance processes go beyond data extraction
Successful data extraction from documents is not the same as the complete automation of a finance process.
In the invoice scenario, finance functions need to be aware that making the extracted invoice data available to a company’s existing enterprise resource planning (ERP) system for the accounts payable process requires integration between the AI tool and the ERP system, Chandra pointed out.
Integration usually requires an application programming interface (API), which in simple terms is computer code that allows computer applications to communicate with one another. An AI vendor that provides a document recognition tool might also provide an API.
What about forecasting? It might not be that straightforward
According to Charting the Future of Accountancy With AI, forecasting is a task that AI can help in financial planning and analysis (FP&A).
AI analytics tools can generate new insights into costing, pricing, sales, revenue, and cash flow by analysing historical relationships between internal and external drivers such as product portfolios, competitor behaviour, natural events, macroeconomics, and historical performance, the report stated.
While Gartner also sees interest growing in using AI to improve FP&A, it said that only a few organisations are currently using it successfully.
The reason is that AI is not yet built into most FP&A application suites, said Christopher Iervolino, senior director analyst at Gartner.
But to someone like Thain, who comes from the airline industry, revenues forecasting with AI is nothing new and is doable with clean and integrated data.
“The airline industry — which has an enormous amount of customer data — has used predictive analytics to help come up with pricing and revenue strategies for ten to 20 years already,” he said. “Indeed the industry is one of the cases used in university-level AI teaching.”
Start with examining your data
As data and its quality matter, Chandra and Thain advised finance functions to take the following steps if they want to deploy AI for forecasting.
Digitise data. According to Chandra, if you have not digitised your data, you will have to do so before using AI to get any forecasting done.
“The more data you have the better,” he said. “There’s no point to forecast based on a small amount of data because it won’t be accurate.”
While a scanner helps to digitise a paper document, it takes the optical character recognition (OCR) technology to enable computers to read its data and content.
However, OCR can only give users a raw dump of data, Chandra pointed out. Thus, finance teams need to work with IT teams to extract data points such as dates and the different types of numbers in a document, he added.
Clean data. A task harder than data digitisation is data cleansing. “There’s always junk in data,” said Chandra. “This is why data cleansing is important. You don’t want to forecast with incorrect data.”
Data cleansing involves the detection, correction, or removal of inaccurate data from a database, a spreadsheet, or other sources. All these tasks can be done by data cleansing tools offered by different technology vendors, Chandra said.
Before data cleansing is done, one has to define the rules for the task, he said.
After a finance team has come up with the rules, it asks the IT function to implement them. An alternative is to use a tool that allows users to directly apply rules without involving IT, Chandra said.
Integrate data. Data cleansing and data integration combined are 70% of the battle, said Thain. Finance functions need to make sure that data points and data sets are integrated — especially when some of the data needed is captured differently because it comes from other functions — before AI tools can provide the desired analysis and insights, he said. An AI solution provider can help with data integration, he added.
Advice to CFOs
Know what you want to do with AI. The most important question that CFOs need to ask before deployment is what they want AI to do, Chandra said. “You start this by mapping your processes,” he said. “Once you have a correct process map, you start to see the pain points. Some of them might not require you to fix with AI at all.”
Act now. As AI deployment is a big step, CFOs should act now, Thain said.
“Adoption is only half the battle while the other half is about changing the way you and others work and an organisation’s culture. So start talking to people who are knowledgeable about AI,” he said.
After that, CFOs need to have a top-down strategic vision and communicate with the organisation about that properly.
Stay away from a big-bang type of transformation. As AI deployment plus data cleansing is a long and tedious process, one can start by taking baby steps, Chandra noted.
“If there are ten steps, you can start with two steps first and see what benefits you can reap before moving to the next one” he said.
- “Skills That Help Accounting Professionals Succeed Alongside AI”
- “Ethical Implications of Artificial Intelligence”
— Teresa Leung is a freelance writer based in Hong Kong. To comment on this article or to suggest an idea for another article, contact Alexis See Tho, an FMmagazine associate editor, at Alexis.SeeTho@aicpa-cima.com.