Following all the rules of AI

26.02.2019 - The rule-based approach provides great solutions for optimising business processes in particular. For example, for correctly recognising and assigning commercial documents. Using pre-defined principles, the system identifies text-based content, compares it with existing data and is then able to automatically classify and correctly assign the incoming mail. Finding the recipient on an invoice is a good example of this. To do this, tangro uses an effective procedure, that firstly automatically recognises the VAT registration number and then compares this with the SAP-ERP system to discover which supplier this number belongs to.

Proven algorithms

The system also compares other ERP data, which can also be linked to the supplier, with the content on the document to discover, for example, which elements of the text in the document relate to invoice line items. In all of these actions, the software applies proven algorithms which ensure that all relevant data in the document is reliably recognised and that the data is correctly processed in the SAP-ERP system. This doesn't need a procedure which acts autonomously because rules can be used to clearly state what a VAT registration number looks like. Taking the example of Germany, this number is always produced following the fixed rule of the country code according to ISO 3166, followed by 9 digits.

No transparency

You could now argue that self-learning AI systems offer the benefit of being able to adapt independently to changing conditions. But these applications bring with them the disadvantage that ultimately it is very hard to know which decisions have enabled the system to come to the conclusion it has reached. Sticking with the example of the VAT registration number: instances may arise where a self-learning text recognition system always looks for this number where the page has a slight yellow tinge simply because in the training data the registration number appeared in a part of the page which had discoloured to yellow. If the system now defines the registration number in all subsequent documents based on this criterion, there will be no way of avoiding mistakes. However, it would be very hard to rectify this mistake because the criteria upon which the decision was taken are not at all transparent. This is compounded by the fact that it is often hard to provide a large amount of quality training data in the world of business.

The processing of inbound documents is a good example of this. Rule-based expert systems provided everything this application requires. Maximum recognition accuracy of the data required with one hundred percent traceability. But this doesn't mean that self-learning systems won’t be used in our inboxes as the technology behind AI continues to advance and we make more use of intelligent machines.  And not vice versa as Hollywood would like us to believe.