This article by M-Files COO Miika Mäkitalo appeared in the February 2019 edition of the German-language magazine Digitus and is reprinted here in English translation.
In many discussions on artificial intelligence, unfortunately, the cart is often put before the horse. Now more than ever, we must adopt more pragmatism in the use of AI in information management.
Hardly any other topic currently generates as much hype as the use of artificial intelligence (AI). Depending on your individual perspective, the picture of how AI systems will replace cohorts of qualified employees across enterprises will be painted either as a horror film or as a dream. It is believed that automation will replace the so-called knowledge workers as it has done with blue-collar workers. But this conception is far from reality — by miles.
We like to remember the 1990s fondly, when AI started to change the world. It was then that expert systems were believed to be able to replace experts and other skilled knowledge workers on a large scale. In practice, this has largely failed. AI-based expert systems usually only work well where highly repetitive, cognitive tasks in a closely framed setting have to be performed. In the routine evaluation of X-ray images, machines are better than humans today; as universal general practitioners, however, they fail. As a rule, AI is good at routine tasks and bad at exceptions.
Even in the interim, AI-supported systems will only be able to take over the complete management of processes in very narrowly defined challenges. An enormous investment in the form of domain know-how, training and fine tuning is required in each case, and so these systems will be used primarily on selected use cases.
Focus on the Essentials
The much greater potential for AI lies in supporting the large mass of knowledge workers who perform tasks that require greater bandwidth. If you ask CIOs where the biggest challenges lie in information management, you will hear variations of the same problem: The amount of unstructured information is growing too fast to be processed by users with today’s means. This, in turn, places demands on IT to provide more automation.
If information technology could increase the productivity of project staff or sales development staff by as little as 10-20% by making it easier for workers to handle information, this would have a significant impact on the flexibility, agility and performance of companies. So, it’s about support and not about substitution.
Approximately 85% of all information is unstructured and is still largely handled through manual processes. The simple process of saving a file makes this clear: The user must examine the file, implicitly classify it and assign it to other business objects and then — hopefully correctly — manually store it. The same applies to documents, emails, chats and all the other forms in which information reaches us today. The handling of unstructured information is still largely manual and that’s why knowledge workers are drowning in a flood of information. Against the background of an estimated doubling of the volume of information every 18 months and the lack of qualified personnel everywhere, this is a massive problem.
This is exactly where AI must start today and improve the possibilities for automated processing of unstructured data. For this purpose, IT systems must be enabled to understand three things for each information object using AI methods: First, what is the piece of information (classification)? Second, how does it relate to other information (contextualization) and, most difficultly, what does it mean (meaning)? To ensure that these findings are not volatile, they must be structured as metadata and documented in a machine-readable form.
Hard Working Helpers in the Background
The first two steps of understanding — classification and contextualization — can already be largely automated with AI today. Self-learning classifiers can suggest the appropriate classification for documents by using machine learning to constantly learn from the behavior of users. This facilitates introduction of and flexibility for changes in the document classes.
Various AI methods are used for contextualization. Text analytics and pattern matching are used to search and extract already known metadata tags such as customer names, project numbers or other data with known formats such as tax numbers, dates, personnel numbers, etc. This procedure is used intensively to create the connection to structured data from ERP or CRM systems and thus establishes context. This is supplemented by the recognition of topics (subjects) using Natural Language Processing (NLP). The first step is to determine the meaning of the content of a document and then to extract metadata entries matching the topic.
These functions are ideally integrated into modern content services platforms, which not only bring together as many repositories as possible, but also provide uniform structures for both metadata and service levels. This includes versioning, access control, or mobile use. On these platforms, classification and contextualization services can enrich, enhance and unlock the vast amount of existing enterprise information. In concrete terms, this means, for example, that all relevant documents relating to a customer are identified in various sources such as network folders, SharePoint or document management systems and offered to the user in the CRM in a 360-degree view of the customer for processing.
In addition, classification and contextualization allow the automation of policies for the correct handling of information. This enables sensitive content such as personal data, confidential project information or internal research results to be automatically recognized and assigned to the relevant protective measures.
In summary, AI can provide significantly more transparency in information management today with classification and contextualization and can help to automate the implementation of important information governance policies. This means that it can significantly support the user to avoid manual errors.
Intelligent Assistants are the Future
Soon, the use of AI will be much more deeply integrated into daily work processes. The term Intelligent Assistants (IA) has established itself as a play on words based on Artificial Intelligence (AI). These assistants will help the user with anticipatory analyses and intelligent suggestions. In order to do so, AI must master the third step of understanding — recognizing the meaning of the content.
This concept can be illustrated with a simple scenario: An incoming email is classified by the intelligent assistant as a request to send a document and the document deadline is assigned. The wizard determines the documents in question, suggests them to the user as a selection and then creates the reply email.
Although this example may seem banal at first glance, at second glance it illustrates the enormous challenges towards AI that needs to be solved. To understand the meaning of this email, the system has to understand simple things like “slides mean a PowerPoint file,” but also more sophisticated concepts like “requesting documents” or that you should receive a presentation at least one day before the meeting if you want to use it in that meeting.
In contrast to specialized expert systems, intelligent assistants are confronted with a much wider range of questions, exceptions and error possibilities. Therefore, the aim here is not to provide a fully automated reply to the mail, but to support the user in processing the email by means of preparatory measures.
The charm of these assistants lies in the fact that they make a wide range of users more efficient – millions of times every day. In addition, they require a lower degree of maturity during the introduction because the user is available to correct errors and they can learn from their mistakes with machine learning. In this way, they automatically adapt to the individual requirements of companies and users alike.
There is no question that there are standardized tasks that can be automated with AI — or even with traditional IT. However, the small, daily efficiencies of intelligent assistants offer the greatest potential to a mass user base. What is most fundamental are the three levels of understanding — classification, contextualization and meaning. Software manufacturers are particularly in demand to integrate more AI into their solutions. Content services platforms offer powerful methods as services with which the handling of unstructured information can be significantly improved. We need a new, more intelligent approach to dealing with information.