It’s a Monday morning and you receive an urgent email from a client. They’ve received a letter from HMRC saying their company accounts are late and threatening a huge fine.
After a quick check you see everything was filed on time, so you ask for a copy of the letter. The client sends through a photo taken from a mobile phone and, as you suspect, it’s a reminder from Companies House that the Confirmation Statement is due.
This is a fairly typical problem faced by accountants, with clients giving incomplete, jumbled information using a mixture of spoken word, writing and pictures. It may be annoying, but at least it means you can never be replaced by a robot.
Or could you?
Artificial Intelligence (AI) is currently one of the hottest areas in technology and aims to do all the things you just did to help your client. So what exactly is AI, and how will it affect the accountancy profession in the future?
It’s all about cats
If you ever attend a conference on AI, it’s only a matter of time before someone mentions cats. The reason is humans can instantly recognise a cat, regardless of colour, breed or the angle it’s viewed from. Computers are good at many things, but spotting cats isn’t one of them.
To give computers this key skill, scientists looked at how the human brain processes and understands images. The information is translated into data and passes through various layers of neurons, with each layer taking the results of the previous layer and processing it further before passing it on to the next. Once finished, the brain can match the shapes pulled from the image to objects it recognises from the past.
Techies replicate this process using multi-layered mathematical models (or artificial neural networks), with several layers of nodes each performing small calculations. They then feed a few thousand images into the model, already knowing which of those images contain a cat. They use statistical techniques to set the model’s weights and parameters to make the model as reliable at spotting cats as possible.
More test data can then be fed into the multi-layered (or “deep”) model to see how well it works. The computer scientists check the results and recalculate the weights to make the model work better. This is the concept of “deep learning”.
Artificial neural networks and deep learning are two key elements of AI, as they mimic the analytical process of the brain and, like us, can learn from experience.
Speaking the lingo
So computers can think like humans. But can they speak like us?
Analysing human speech must be every scientist’s worst nightmare. We structure sentences in different ways (often with misspellings and bad grammar) and use slang, abbreviations, sarcasm and words with ambiguous meaning. We’ve been talking to people all our lives, yet there are still occasions when we misunderstand what others are trying to say.
Natural Language Processing (NLP) tries to extract information and meaning from speech, based not only on the meaning of words but also where they fall in the sentence and the words surrounding them. The rise of computing and the Internet has made this process easier, as we have access to large samples of real-life sentences (an example of “big data”) that can be loaded into a database and analysed.
NLP now works well. In a recent test at Stanford University, an AI model was given extracts from Wikipedia and asked questions on the content (e.g. “what is the capital of Kenya?”). The AI model scored higher marks than humans given the same test.
So how can AI technology help your client?
The theory sounds good, but how could AI technology be used to understand and respond to your client’s confused query?
Let’s run the initial email through a real-life NLP program. In my test, the program pulled out “letter” as the most important keyword, followed by “company accounts”, “HMRC” and “fine”. It recognised “fine” to mean a financial penalty rather than as a positive description (e.g. “fine weather”).
So this would give your AI software a starting point. Just like humans, it can then decide what to do based on its training and experience. This would consist of some pre-set rules and training data initially loaded into the software (for example, an accountancy firm’s email history of past queries and responses) plus any emails the AI software had processed since its initial training.
It may first spot the word “fine” and, based on a rule, check whether anything is overdue or was filed late. When the answer is no, it may search for previous emails with the same keywords and find that responses typically requested a copy of the letter. It could then construct a simple email and send it to the client.
Once a copy of the letter is received, the software can use an artificial neural network to analyse the image and spot the text. It runs the extracted content through an NLP program to determine the main keywords, and again checks them against its database. It concludes it’s a reminder to file a Confirmation Statement.
Are you scheduled for obsolescence?
Will the advance of AI technology mean the world no longer needs accountants?
Some implementations of AI technology are starting to be rolled out (and make a positive impact) in areas such as healthcare. But the good news is we are not yet at a point where AI software can think, make decisions and communicate with clients without them even realising they’re speaking to a computer - the Holy Grail of AI developers.
That said, never before have we had such huge amounts of cheap computing power, storage, bandwidth and test data available. The tech titans of the US and China are working hard to lead the field, and investors are falling over each other to pour money into AI startups.
Over the next five or ten years, AI technology will continue to develop quickly and will ultimately spread into areas such as accountancy. So don’t be too surprised if, one far off day, the annoying AI robot from HR hands you a P45.
About Matt Bailey
Founder of Azura Cloud Systems (Gbooks), the leading cloud-based tax and accounts system.