AP Automation and AI

Why Business Networks and Artificial Intelligence are a match made in heaven 

The past eighteen months have seen an explosion in interest and innovation around artificial intelligence (AI). Businesses worldwide are racing to catch up, and according to Gartner, 90% of CFOs plan to level up spending on AI-driven tools in 2024. 

We caught up with Rolf Njor Jensen, SVP of Engineering at Tradeshift, to understand how his team is deploying AI and machine learning within Tradeshift’s e-invoicing and AP automation solution and its impact on accuracy, efficiency, trust and more effective resource allocation among customers. 

We also asked him about their transformative potential across the broader finance and procurement function and why workers should welcome these technologies, rather than fearing being replaced by increasingly sophisticated algorithms.

Watch the video highlights from this broad-ranging and insightful conversation below, and scroll down for a more in-depth Q&A with Rolf. 

You’ve been at Tradeshift pretty much since the beginning. Where did AI figure in those early conversations? 

Back then AI was still at a relatively immature phase in its development, but we saw the potential from the beginning. At the time, much of the focus was on robotic process automation, which helps make existing processes go faster. 

That’s great if you’ve got a process that already works well, but if the process itself is inadequate or inefficient, all you’re doing is replicating something that shouldn’t be there in the first place. We looked at the potential of AI to fundamentally reimagine business  processes and achieve a much better outcome overall.

What were some of the earliest applications of AI on the Tradeshift platform?

When I started,  most invoicing was paper-based. And if it wasn’t paper-based, then it was typically PDF-based. As a buyer, you need that PDF converted into structured data before posting it into your accounting system. 

A lot of what Tradeshift does for our customers is put structure into data. One of the first applications we developed on Tradeshift was CloudScan, which takes a PDF and applies a machine-learning model to convert that document into a structured invoice. A version of that same core technology still exists on Tradeshift today.

What makes Tradeshift such a fertile environment for AI applications? 

Every machine learning model is trained on data; so the old adage of ‘garbage in, garbage out’ applies! At Tradeshift we’re focused on taking data and giving it structure so it can have far greater utility across the business . 

The quality filter for that data is our network. We have buyers and suppliers collaborating and validating that data in real-time. It’s also on an individual level when the buyer or suppliers processes and validate that data through their accounting processes. What you end up with is a data set that’s both structured and of a super high quality.

Can you give us an example of a challenge you are currently helping to address through AI?

Identity is a significant issue for most organisations dealing with thousands of suppliers. Tradeshift is a business network with over a million active businesses on our platform. 

New companies are always joining, and it’s not uncommon for several departments within a single organisation to need to connect to Tradeshift. As a buyer, that could cause confusion. How do you determine which connections are unique to one company and which are multiple instances from the same company?

We’ve deployed ML (machine learning) models in our onboarding process to address this issue. These models quickly determine the likelihood that a new company coming onto the network is the same as another. This notion of identity verification and trust is a critical property of any business network.

Rightly or wrongly, some businesses might be wary of ceding control too early on to automation and AI-powered systems. How does Tradeshift’s solution help alleviate those concerns?

Some jobs are done better by a machine, and invoice coding is a great example. It’s heavily manual and very error-prone. We can introduce huge efficiencies by effectively removing the manual element of this process. But we do it in two modes: assisted coding, where we have an ML model suggesting coding options to a human operator, and automated mode, where you simply let the algorithm do its thing. 

Of course, this raises the question of when is the right time to switch from assisted to automated coding. That’s why we developed the ADA Automation Dashboard, which provides a visual representation of how your manual coding operations are performing compared to the automated model. Users can play around with different thresholds around the cost of manual labour, error tolerance and much more to see when it makes sense to switch to automated coding over time. We’ll always need human operatives, and this is a great example of how technology can make their working lives so much easier.

We’re in the midst of a giant leap forward in AI capabilities. Where might we see Tradeshift leveraging these technologies on its platform? 

Another really exciting use case for AI is in our B2B marketplace solution for procurement. We see a significant opportunity around SKU enhancement, using generative AI to help sellers get their products in front of buyers more effectively. 

Returning to accounts payable, we see a massive opportunity to deploy AI to alleviate some of the challenges businesses face in linking invoices to a specific transaction. We’re often talking about thousands of documents that tie into a particular transaction, including multiple POs, transport orders, goods receipts, delivery notes, etc. 

Doing the right linking is the precursor to touchless processing and automation in the accounts payable process. This is often very complex, with potentially thousands of different invoices and documents linking to one transaction. Disintermediation models like Continuous Transaction Controls add another layer of complexity. We’ve already done a lot to automate in this area and see massive potential in technologies like generative AI to streamline the process further. 

There’s a fair amount of uncertainty around AI, particularly regarding safety and security. Is that something we’re conscious of as we explore the potential of this technology?

This links back to what we discussed regarding data and data quality. We have some instances where we train ML models based on our platform’s entire data cohort to provide generic suggestions. We may then do more specific training on individual customer data sets. 

We’re careful to avoid any potential cross-contamination of customer data sets, which we can do quite easily due to our clear segmentation across different customer data. 

What does the future look like for finance teams as AI becomes more ubiquitous? To what extent are fears that these technologies might replace humans justified?

Relationships are at the heart of every buyer-supplier connection on the Tradeshift platform. Machines are great when it comes to doing many jobs that get in the way, but when it comes to nurturing and growing those relationships, human beings win hands down. 

Machines can do a lot of the transactional work currently done by humans, but you still need governance and oversight from a human operator, and you still need people to decide how to use those tools in the most efficient way.