Classification: a key step for automation

Classification is a key step for automation. And automation is the “Holy Grail” for companies. Robotic process automation (RPA) technologies have taken by storm the ecosystem of simple automations. They can handle a large amount of processes that can be automated and can be tailored for specific workflows. But they reach a limit when there is some type of discerning to do, some decision to be taken, for example a document which has a typography in the spelling of a company. A human can easily point out it’s a spelling mistake, but for rule-based software this is not an easy task, especially above some volume. Or an image of a car with a scratch, should this be flagged as a small issue by an insurance company and therefore be paid right away, or as a medium issue that needs the review of an expert, therefore more overhead costs? These decisions have an impact on business, and AI/ML is pushing the boundaries in automation by correctly classifying digital files, like images, text, audio and video.

Document classification

Classifying documents is a crucial step in many business workflows. Depending on the content of a document, different steps are taken. For example, a telephone bill is sent during an online loan application. A human reviews the invoice, checks address, bank account and dates. Many small typos are present in many documents, simple errors that a human immediately would notice and would move on, but rule-based systems crash with them. Another problem is the format of the documents. It's just impossible to design a rule-based system everytime a department releases a new version of a document. How many invoice formats have you seen? As a human, you see them as all “more or less” the same, but for rule-based software that’s just too hard. NLP technologies can understand the text, and the document, and models are able to find, understand and process the content of those documents. In the example we are explaining, a user would send a telephone bill and the NLP engine is able to extract all relevant fields from different types of document formats, and depending on the content, flag the document as “approved”, “not approved”, or “needs review by an expert”. This speeds up the loan application process providing a much better customer experience, and saving thousands of man hours of internal admin work.

Audio and image defect classification?

Is this a defect on this product? What is that weird sound in my car, is there something broken? These are questions that are being answered by AI/ML models. Companies like Facebook, Google and Instagram have contributed to the astonishing development of image processing. They have billions of images in their databases, which allow them to train huge models that keep getting better at detecting events of interest inside the image. One of the most widespread uses in business outside social media and digital marketing for image processing is defect detection in manufacturing processes. AI/ML models process images of products and classify them as “OK”, “NOTOK” and “needs review”. Furthermore, the real power of these tools are that they can be used in edge devices. This means, the computing is done at device level (smart camera, tablets, phones, IoT devices), which provides a totally new set of possibilities and reduces the costs because companies only pay for the computing they need. Traditional computer vision systems are very precise in specific inspection tasks, but they require a big hardware investment and usually they can’t be trained for new inspection tasks. On the other hand, edge AI inspection systems like a smart camera are much less costly and if they are open source, they can be trained for new inspection tasks in a matter of minutes/hours.

Intent and behavior classification

Intent recognition is widely used in chatbots and in other NLP applications. NLP models can detect what is the intent of the text, what is the intention of a given text. Depending on the intent of the text, the workflow of a given business process will go one or other way. And on top of intent recognition, sentiment analysis models are also used. In combination with both, AI/ML models can not only understand what the user wants, but also a customer sentiment score, which can be an important piece of information. For example, a user is using a chatbot of a food delivering company. The NLP understands the order has not arrived yet, and so the chatbot activates the standard process in those cases. The sentiment analysis score shows the customer is not happy at all, so the chatbot activates a $3 voucher for the next purchase. This allows to automate more sophisticated customer interactions which otherwise would only be possible using live agents in customer centers, which would increase the structural costs and would reduce the profitability of the business.