Our client, one of the largest automobile manufacturers in the world, wanted to automate their finance and accounting processes. JIFFY.ai’s AI-based Invoice Processing HyperApp was leveraged by the customer to overcome the challenges of handling a multi-country roll out. Our HyperApp ensured that the customer was able to achieve 85% straight through processing over a 12-week period across a volume of 150,000 invoices per month for 5,000 suppliers.
Case Study
Intelligent Invoice Processing for a Top Automaker
JIFFY.ai’s AI-based Invoice Processing HyperApp was leveraged by the customer to overcome the challenges of handling a multi-country roll out.
Summary
Challenge
Automating invoice processing in a centralized manner was becoming extremely complex for our client due to huge monthly volumes, varying formats and business rules specific to different supplier types. The client was depending on manual execution, making the process time-consuming and error prone. With increasing volumes, the F&A team was unable to meet the processing timelines. Implementing RPA did not yield results as expected, since most RPA solutions are rule-based, and are unable to handle frequent changes in invoice templates. New vendors, new invoice formats, changes in supplier status etc, made the initial RPA based approach completely unsuccessful.
Every time a new vendor was added to the system or an invoice with a new format was to be processed, or an invoice was delivered in a different format, or an existing supplier became a top supplier for the company, the invoice template had to be updated. This meant extensive changes to the invoice template. This was a continuous process and a maintenance nightmare. Since the customer worked with over 6000 suppliers, the range of invoice formats to be covered were huge and they were looking to integrate an automation solution to process at least 70% of the invoices, but this was complicated due to new formats coming in regularly.
Implementing RPA did not yield results as expected, since most RPA solutions are rule-based and do not fare well with complex processes or unstructured data.
Our approach
JIFFY.ai helped the client address this issue by:
Leveraging JIFFY.ai’s AI based document extraction engine to automatically understand suppliers and their invoice formats without manually training each format.
Leveraging the human in the loop interface to correct errors while processing invoices and automatically improving the overall straight through processing using ML models.
Easily configuring validation rules for specific suppliers and geographies.
Deriving deep insights into suppliers, payments, and cash flows and improving business efficiency and experience.
The JIFFY.ai team studied the customer’s top suppliers, who contributed to the majority of invoices being processed. A combination of rule-based automation and ML models were applied to process all the invoice types using JIFFY.ai.
Solution
How this worked
Creation of training data
To begin with, the JIFFY.ai bot was fed with 12 months of historical invoices. Additionally, it read the invoices that were registered by the operational resources into the ERP system. The bot would automatically map the invoice PDFs to the data in the ERP system. In this manner, it would auto-create the training data for creating the ML model.
Creation of ML model
The bot used the “labelled” training data to create the ML model. It would automatically tune the hyper parameters and choose the right algorithm for predictions. For each field in the invoice which needed to be extracted, a model would be auto-created by the bot.
Applying the ML model
The JIFFY.ai bot began to apply the Machine Learning model it had created to each invoice for processing.
Validations, errors and exceptions
After validating the data, invoices were processed without any delays. Whenever an error or exception was encountered, the bot flagged it up for manual intervention. In all, 15 exceptions were handled, which is equal to 90% of the exception coverage. It also self-learned how such errors and exceptions were being dealt with by a manual operator, and when it occurred again, it could decide by itself how to proceed, and process it all on its own.
Results
