I have to confess that I am a gadget head. My enthusiasm for collecting stuff occasionally leads to my abilities to use them falling short. 

As CTO, it’s an enjoyable and important part of my role to encourage research into new technologies. However, the equally important, but more challenging part is to determine which of these cool new tools should enter our “best of breed“ collection and also ensure that we are able leverage them fully.

Artificial Intelligence is one of today’s ‘must have’, collectable technologies. It also provides a highly desirable solution to some rather tricky problems many companies experience with processing unstructured client communications.

Given its complexity, unstructured call centre communications is the perfect testbed to try out some of the latest and greatest AI transformers. Previously we were using an existing partially trained neural net, which we were constantly feeding with a simple “bag of words” (term frequency-inverse document frequency). Although we were getting quite a good accuracy, it all felt a bit, well, low-tech.

So, like any good gadget head, I want to find the latest doodad to try out…

The first point of call was to enhance our pre-processing step with some big gun AI, such as transformers – namely Google’s BERT and XLNet. For those of you who are not familiar, these are the Top Guns of natural language processing, having been trained at significant cost with vast amounts of data. Both tools have shown incredible results applied to text classification and cardinal data extraction.

While we anticipated incredible results, sadly, we got worse outcomes than our simple bag of words – albeit a domain specific model with some careful additional data cleaning and model calibration. We were a little crestfallen.

The reason is that the vast training sets used for training these models are not geared to communications about financial services products. So, however powerful the model, it may not be optimal for the job. As an illustration – a new financial markets analyst may be asked by a client to “call the reverse convert” – this doesn’t make sense as it would either have to be to “put” the convert or it wasn’t actually a “reverse” convertible. The language of financial markets, like many industries, is very specific and arcane and so we need a domain specific model it also takes time to configure them optimally.

The value of the Opsmatix SaaS AI solution is that we are constantly training and optimising models from the shared experience of many similar financial services clients. Over time, the training set becomes far more extensive (and therefore smarter) than any one organisation can access. We also know how to tweak the critically important data pre-processing to maximise model accuracy. The beauty of providing this as a managed service is that clients very quickly get the results they want, without having to spend ages test driving many different solutions on their own.

This approach is enabling us to deliver increased value for every client beyond what would be possible alone. Being Top Gun is very attractive, but everyone needs a wingman – and this is where we come into our own.