For this new episode of “Let’s give the the Product Department a say”, we had the chance to talk with Yoam HAZOT, an Artificial Intelligence engineer at DIMO Software. Here he explains both his role and that of the group’s “artificial intelligence team”, the different types of artificial intelligence, various related projects, and the methodology used. He then goes on to provide more details regarding some of Notilus’ AI projects, especially the OCR, a technological process that automatically recognizes key fields of an expense report in a paper receipt.
What is artificial intelligence?
AI is a set of techniques and theories that allow machines to solve problems that require some form of intelligence. The term “artificial intelligence” is employed quite broadly today and is increasingly cited. In the field of learning, we mainly find Machine Learning and Deep Learning.
Machine Learning corresponds to a category of algorithms designed to pick up meaningful patterns from an original set of data to make predictions on other data.
Deep Learning, a sub-category of Machine Learning, is based on artificial neural networks with a structure comparable to that of the human brain.
These types of algorithms are purposely designed to mimic human reasoning.
With our OCR, all fields such as VAT, currency or amount are automatically detected. With the electronic archiving, your expense receipts are stored in a secure cloud, no more paper.
How is DIMO Software’s AI team organized?
Today, DIMO Software’s Artificial Intelligence team consists of two engineers, including Yoam HAZOT, who devote themselves to extracting predictors from one set of data that apply to another set of data. In this process of generating useful data for our end-users, they are used to wearing different hats, such as the data analyst’s hat when it comes to data analysis and processing, the data scientist’s hat when they work on the design and development of predictive models, and finally the data engineer’s hat when they focus on the industrialization of these models.
The AI team is transversal and works for all departments within DIMO Software. Apart from the flagship project (automatic identification of the elements that make up an expense report for the Notilus solution), they have also produced YellowBox CRM in collaboration with customer relationship management, CashOnTime, a debt collection and accounting lettering software, and CMMS for DIMO Maint.
But how does a project come to fruition? Below are the different stages:
Product teams express their needs.
“Our first task is to ask ourselves if the solution to the problem truly requires artificial intelligence, which often is not the case. We have to evaluate whether the problem is worthy of such investments since the implementation and development of an AI can be quite costly in terms of time and energy, while sometimes not being the most appropriate solution,” says Yoam.
The project is validated and POC (proof of concept) begins.
If the AI team finds the project interesting and AI necessary, a POC is launched to determine its feasibility. Several questions must then be addressed: is there enough data available? Is there enough time? It is not uncommon for a project to be abandoned at this stage.
“For example, we once worked on a speech recognition project based on the latest technological progress in the field. The results were remarkable, but only in English because we didn’t have enough French audio recordings to get a satisfactory result,” explains Yoam.
- Standards are set for the project’s performance before entering production.
Fortunately for the AI team, the project is often feasible. They then move on to the development phase of the project, on the condition that the end product must achieve the best possible performance in its field. This last part is a “sine qua non” criteria that must be met to enter the development phase.
After everything is done, the product teams only have to integrate the result into the corresponding solutions.
Notilus’ “OCR Project”, the pride and joy of the AI team
The AI team was launched four years ago with a flagship project based on OCR (Optical Character Recognition) technology.
OCR is used to obtain information from an image (mostly the text inside). For Notilus, the goal was to send an image (namely an expense receipt) and extract the key elements of the expense from it: the amount, the date, the currency, the VAT, the country, and the nature of the expense, which often isn’t indicated on the receipt and needs to be deduced.
The AI team developed both Machine Learning and Deep Learning algorithms to work together and reliably identify the key elements in the image.
“It gets quickly technical, but all you need to know is that there are algorithms behind every detection. We’ve implemented multiple predictors that work simultaneously so that there are as few errors as possible, with one algorithm covering for the potential error of another,” says Yoam.
An OCR that sets itself apart from the competition
Today, our solution can analyze receipts from several countries, notably France, Switzerland, England, Spain, Germany, the United States, Canada, Mexico, Italy, and Portugal. It also boasts several other features that set it apart from the competition:
We have set up confidence scores for each of the fields our solution is set to recognize. This allows us to measure the performance of our predictors and pinpoint among the different algorithms the most accurate one so as to guarantee a very high recognition rate.
Spectacular processing speed
Our AI team often has to study other solutions on the market to see how Notilus’ OCR compares to the competition.
- Impressive recognition rates
Any player in the market can pretend they have the best OCR. What we have, however, on our side, is a database filled with some 7.9 million tickets that have been sent to us by users over the years. On average, we boast a recognition rate superior to 90% for each field, and in some cases, we are closer to 95%.
“It should be noted that the 90% recognition rate per field is required before deployment to production. For example, we are currently working on restaurant or hotel tips and making sure that on all the validated tickets, we have a very high recognition rate. You would have to imagine that with about 14,000 tickets a day, every single percent point is crucial, both for our users and for us.”
Credential control to ensure a strong recognition rate
Unfortunately, we can’t train our algorithms directly on the expense data that is being continuously sent in. To ensure a high recognition rate, it is necessary to first control what has been logged in by the user.
“There are many users who submit receipts but fill in a value that is not the one that can be seen in the picture. This occurs, for example, when they want to remove a fee that is not covered by the company or indicate the amount reimbursed by the company instead of the total price. Then there are other users, who submit expenses claims with unusable supporting documentation, such as a screen from the gas station, blurred, handwritten, or multiple receipts, etc.” he adds.
Security for our customers
Our databases are secure with regular backups at multiple remote sites. All receipts are anonymized once they are in the database. Regarding GDPR and data protection, the receipt image itself disappears after three days.
Is Artificial Intelligence employed anywhere else in Notilus?
Yes, of course, especially in the Automotive Fleet section with Notilus YourWay!
“We have created a chatbot from scratch whose goal is to help the user take control of the car fleet solution. It can enter expenses, appointments, mileage meters, etc. Another area of interest is the automatic integration of supplier invoices, for which artificial intelligence can automatically identify a file and its provider. For example, if your leasing company sends you the contract for any given vehicle, it is automatically recognized and linked to the vehicle file in Notilus YourWay.”
Other examples include:
- Estimating the fuel consumption of vehicles;
- Readjusting the mileage stipulations, which allows the fleet manager to be closer to reality and to tailor his contracts;
- Detecting anomalies entered in the odometers.
There still are many other features that we have developed but that are not yet integrated into the products. Artificial intelligence in Notilus still has plenty of room for growth!
Our solution is dedicated to the moblity management. In addition to the expense reports management you can plan, book and track your business trips. It is also possible to manage your car fleet.