automation in finance: 3 use cases.

There is a keen desire to work more efficiently in many finance departments. A good digitalisation and automation process takes repetitive tasks off the hands of your finance colleagues and gives them the opportunity to contribute strategically to the company. Moreover, it also increases the quality of the work that’s been done. 

The cases below give you some realistic examples. They show that it is okay to take your time on your work and not to expect spectacular results overnight. Every step you take – no matter how small it may seem – saves you time. And that time can then be invested in the next, new stage of your automation project.

• automatic invoice processing
• predicting turnover with machine learning
• facilitate the credit controller’s work

automatic invoice processing.

A company in the transport sector recently came knocking at our door to ask if we could make their invoicing process smoother. That’s a question we get asked a lot these days. 

From monitoring and registration to approval and archiving, there is a lot involved in processing invoices. Our customer mentioned that their current process takes a long time and is prone to errors. And such mistakes can be costly for any company. On the one hand, you can take that literally: late payments can result in high fines. On the other hand, you may jeopardise the good name of your company.

deep-dive into the process.

We sat down with a number of stakeholders and looked into the current invoice processing procedure together. The transport company still receives a lot of paper invoices. All data is manually retyped into the accounting system, which is, of course, time-consuming and prone to errors. 

Invoice processing is usually an ideal candidate for automation: it is a repetitive and standardised process, with structured data. In the next step, we went deeper – via task and process mining – into their specific procedure. 

Wim, Senior Solutions Architect at Ausy, was involved in that step: “In task and process mining, we find out which specific steps a human colleague takes, so that we know which actions the software bot will have to perform. Sometimes this reveals very important concerns that – if we didn’t know about them – would cause the bot to stumble on a certain step during the first run.” 

“In this case, the additional invoicing cost of two euros for paper invoices was such a concern. It means that the final amount invoiced differs from the previously approved quote. That’s something the bot needs to know: to process such invoices correctly, it needs to look up a specific booking code.”

We also took the following concerns into account when developing the software bot: 

• the layout of each invoice is different
• the length of invoices varies: sometimes it’s one page, sometimes it’s several pages
• invoices arrive from all over the world, so in different languages and currencies

first step: train the bot.

Having the paper invoices scanned in and then having a software bot run them helps control errors and speeds up the work. But a bot cannot be brought into the team overnight. It has to be trained first.

Here’s a clarifying example: we as humans can immediately see where the total amount to be paid is on any invoice. A bot has to learn something like that. It must be able to recognise that the largest amount with a euro symbol in front of it is the amount to be paid. The model is therefore presented with invoice after invoice in order to learn to recognise what the sum is, what the product or service reference codes are, what quantities of a given product are invoiced, which amount includes VAT and which doesn’t, and so on.

So the bot really learns to understand the content, not just the layout of an invoice. That is a sustainable way of working. It also means that the bot doesn’t have to start learning all over again if the supplier changes its logo or the design of its invoices.

from attended to unattended bot.

Once it’s been trained, the software bot can start working. It’s always with human intervention first anyway – which are referred to as attended bots. We show the bot an invoice on which it must interpret the information correctly. The outcome of that interpretation is submitted to a human being for approval. Once the bot is 99.99% correct, it can continue on its own – as an unattended bot.   

The software bot works with a combination of RPA and AI, and can:

• retrieve scans or mails containing invoices based on specific criteria
• automatically add invoices to your accounting software
• compare invoices with the corresponding quotes
• read, recognise, enter, and process invoice data
• initiate an approval flow based on the name on the invoice
• book invoices automatically and link them to the correct cost centre
• make approved payments automatically

Human intervention is only required when the software bots detect anomalies.

satisfied customers, satisfied employees.

If you digitalise your invoice processing and then automate it (using RPA and AI), you will soon reap the rewards. For the transport company, this project should therefore ultimately lead to: 

• faster and more accurate invoice processing: erroneous or late payments are a thing of the past  
• a digital and uniform way of working that provides a better overview: everyone knows the status of an invoice at all times, and paper invoices no longer get lost  
• a digital archive in which all processed invoices are stored: everyone can easily retrieve old invoices – also useful during an audit

Customer satisfaction increases, while employees are presented with more challenging work and feel more in control. Who wouldn’t want that?!

predicting turnover with machine learning.

We don’t need to tell you that a lot of relevant insights can be gained from the abundance of data within your company. Figuring out which data to utilise when and where is an intensive and error-prone process. That’s why more and more predictions and visualisations today are made using a system linked to machine learning. 

Such a smart solution extracts the necessary information itself from different data sets and then combines them. You can derive various predictions from all the different analyses that can be performed on that data. That is, it allows you to make certain business decisions based on actual data, not just past data.

Let’s look at a specific example: forecasting turnover.

Revenue forecasting and determining budgets is something that can be optimised in our own company. To give you an idea: today, the finance team asks for a lot of input from the business, and everything is manually recorded in various Excel files.

There is close cooperation: our finance experts check if the data passed on are correct and see if the expectations are in line with what is possible. Based on the results of this analysis, they challenge their colleagues in the business. Sometimes, their assumptions are too optimistic, but mostly, they’re just too conservative. 

In the latter case, finance must be able to encourage the business to take more risks. But that’s sometimes easier said than done, which is why our finance team is asking for a system that will allow them to challenge the business’s forecasts more quickly.

feeding the machine with data.

We’re now experimenting with a solution based on machine learning. This solution is being rolled out internationally in several Randstad and Ausy countries. The machine returns a certain result and, based on that, expectations can be adjusted accordingly.

A number of steps preceded the start of this experiment in our country last year.

First of all, we collected stable data to feed to the machine. The choice of data depends on the KPIs you want to measure. Take our case: the number of people we employ has an impact on our turnover. The more people employed, the higher the turnover. We want to know what growth in the workforce is possible in certain departments. The question in this case is then: can we predict what the growth in the workforce will be on the basis of other parameters – and can that factor then help determine what our annual turnover will be?

Gross domestic product, for example, is a stable parameter in this case. There is several years’ worth of data on this, which can be correlated with growth in the workforce.

what data mimics reality?

Such a system doesn’t work completely correctly overnight. That’s why it is important to carry out sufficient tests, which is the phase we’re in at the moment. 

What do these tests entail? Well, we have all the turnover data from last year. So we could feed the machine with data from the first half of last year and expect a forecast for the second half of that year. We add a new parameter each time and try to get as close as possible to the effective results of the second half of last year. Once that’s been achieved, you know that the data you are providing is the right output. You can then provide the same data to make predictions for this year or the years to come.

There are usually some adjustments to be made after the first run. We noticed, for example, that the model indicated a so-called corona dip every year because the entire economy was brought to a standstill at the beginning of the pandemic. Today, however, we know that there is little chance of such a dip recurring annually. So, to make sure there are no impurities in our data, we adjust things like this manually. 

future plans.

Our next step? That’s to compare the machine’s result with the business’s assumptions, and then challenge them. 

Kenny, Ausy’s Team Lead Business Controlling, has been involved in the process from the start. He has been enthusiastic from the beginning, and he’s now also convinced of the benefits this solution will bring in the long run. “We can eliminate the errors that human assumptions bring. This higher degree of accuracy will allow our controlling department to focus on tasks that can add more value. Our controllers will be less busy with number crunching and able to work more analytically.” 

“The predictions the machine makes provide us with information that allows us to make informed, proactive decisions,” Kenny continues. “Potential problems are also spotted sooner. We can then work on that more quickly, together with the business.”

facilitate the credit controller’s work.

Another example from our own Ausy finance department: iController has made the work of our credit controllers a lot more pleasant and easier for the past two years. But how?

If you work in finance, then you know that customers don’t always pay the invoices you send them on time. This may be because of gaps in their own processing, because they disagree with what has been invoiced, or because there are administrative issues with the invoice (because a PO is required, for example).

For credit controllers, the follow-up of all those arrears is an intensive process: they have to call and email constantly and need a correct overview of the unpaid invoices at all times.

iController automates part of the credit controller’s work. This is how it works here, by and large:

• the basic information for all overdue invoices is in Unit4
• the invoices themselves are in Changepoint
• iController retrieves all information from those two systems twice a day
• iController contains an overview of which amount is outstanding with which customer

The different procedures we use to deal with late payments are set up in iController. These procedures – from standard emails to official notices from the lawyer – are then carried out automatically. If an invoice is one week overdue, for example, a first reminder is automatically sent via email. This is a standard mail that you can create yourself and add to iController. You can also manage your mailbox from there. 

The manual determination of who should be reminded is now fully automated. We now have a structured process, in which many actions no longer have to be performed manually one-on-one with regard to late payers. 

Such an automated way of working also allows our credit controllers to easily draw up all kinds of interesting reports, partly thanks to the history that is present in iController. Which customers always pay on time? And which ones are always late? What is your DSO ratio? At the end of the month, we will also clearly know which files are still pending and may or may not need to be escalated.

“iController has helped our department tremendously,” says Koen, Finance Director at Ausy. “The dunning process has become much more efficient since it was automated, which means that arrears are settled better than before. The work of our credit controllers has not only become easier, but also more interesting. The tool takes over the administrative and repetitive part, freeing up more time to perform analyses and thus provide added business value.”

conclusion.

You probably noticed what all these cases have in common: they want to give finance experts more time to work on the story behind the figures. You don’t have to take 10 steps all at once in order to benefit from the advantages of digitalisation and automation. Even if you start with your low-hanging fruit, you’ll soon see an impact on the efficiency of your processes and the reliability of your data. 

If you want to know more about these cases or are curious about other automation applications in finance departments, please don’t hesitate to send us a message. We have plenty of stories and tips that we would be happy to share with you!