Coupa’s Deep Learning and How It Applies to Spend Optimisation

At Coupa Inspire last week, we attended the ‘AI – Beyond the Marketing Fluff’ session with Paddy Lawton, Coupa’s General Manager of Spend Analytics. AI has been discussed a lot this year at all of the conferences we’ve attended, it’s clearly a hot topic, and it’s full of buzzwords. That’s why we were interested in this presentation, because it promised to go beyond that and the evolution of artificial intelligence in its long journey towards procurement benefits, and instead give a practical, as opposed to a theoretical, look at machine learning, and how it is realistically used today in spend management. We weren’t disappointed – Paddy can get quite ‘techie’ but he doesn’t forget to apply his knowledge to what the audience really wants to know – how can this software actually help me?

He began by explaining how we should envisage artificial intelligence: classifying it into machine learning and deep learning. To cut a long story short – machine learning he describes as rules-based, and once the machine knows the rules, it can perform against them: see the key word (in a line item, invoice, etc) and apply the rules to understand how it should be classified, normalising company name variations along the way. Deep learning comes when the system is trained to understand whole sentences from contexts and has long- and short-term memory. So it also knows the subsidies of those companies and can put them in the right place, for example. It is seeing through the fog and using developed intelligence to think like the human brain. It achieves this by ingesting data, and more data, and more data. The more data it absorbs – the cleverer it gets – whether that be from invoices, contracts, POs, T&E, etc.

But how can something that understands numbers, read words. That was what was really interesting. By vector representation, he says, all words are represented by vectors. Using Word2vec from Google, even if the machine hasn’t seen something before – has no rules for it – so long as the words are similar, or have a short distance between them, it can infer the meaning. So to get to the word Queen it might apply - King minus Woman equals Queen. It can understand a sentence and even predict the next word, because of the context.

So the data, from whatever format you put in, from whatever source, from whatever people have spent money on, is structured and cleansed into a format that finance and procurement can actually use. The Spend360 acquisition has meant Coupa software will be able to help users make good decisions by augmenting their taxonomy, however different each is, by classifying correctly. And not just Coupa data, all your systems’ data, both direct and indirect, cracking masses of classification problems. The value is, if you have loads of portfolio firms, it allows you to see what they’re all buying, across disparate amounts of data – something that traditional classification methods, or manual ones, would be too slow to reap any value. You would never get to see the long tail.

Imagine having the power accurately to categorise a million invoices in about a week, normalise millions of rows of data, and find every bit of spend that has ever gone through the general ledger accounts, no matter the payment source. With that visibility you can take advantage of economies of scale, reduce supplier numbers, control costs and more effectively manage the supply chain because you can see what purchases are made, from whom, where, when and why, globally and from any language or currency. It can make a massive difference getting close to 100% spend visibility – the more you can see, the more you can optimise. (A bit sales-pitchy we concede – but nevertheless, impressive!)

Discuss this:

Your email address will not be published. Required fields are marked *