Enterprises are increasingly adopting large language models (LLMs) to augment and automate many tasks. But even the best LLMs hallucinate 20-30% of the time out of the box. So, vendors are exploring various approaches to improve accuracy, including fine-tuning, retrieval augmented generation (RAG), and other techniques.
Stampli recently announced a significant upgrade to its accounts payable (AP) automation platform, which claims 97% accuracy for matching purchase orders. It can also further improve with ongoing feedback. The new PO matching service is available as an add-on service for Stampli customers who use Oracle NetSuite, Sage Intacct and SAP, with support for additional financial systems coming in weeks.
The company has been applying various AI models to improve AP processes for over a decade, and the gen AI enhancements build on its existing approach. They call this combination of traditional AI, generative AI, and business process modeling cognitive AI. Eyal Feldman, CEO and Co-Founder at Stampli, explains this branding:
What we're doing is building an agent, and in essence, try to explain it in a simplified way. You want to look at generative AI like a brain, someone that is intelligent that can perform and reason. However, they don't understand the profession of AP because they were never trained on AP. They were trained on general data, on things that were available out there. But we have very specific knowledge of AP because we run AP for, you know, many, many organizations for billions of dollars of transactions.
It's also important to look at this new generation of AI like a new employee without experience in your business, but you can grow to trust over time. Results are surfaced within Stampli's existing service that provides an audit trail and streamlines communications between finance teams and the rest of the business. AI-generated results are personified as being from "Billy the Bot," allowing finance teams to approve suggestions or amend them, which helps improve accuracy over time.
Feldman says they thought about how to reverse engineer the user's thought process. This helped them break it down into many micro-components with checks and balances that could allow it to perform the task at hand to the level of an experienced human. In finance, it is critical to provide the certainty that there are no mistakes, so Stampli developed many different ways of double-checking results.
For example, POs and invoices can run into dozens or hundreds of line items, and discrepancies are routine. Discrepancies can include inconsistent descriptions, quantities and prices; mismatched unit types; missing deliveries or line items split across multiple deliveries; and taxes, freight charges, credits, discounts, rebates and many other variables. Each discrepancy requires careful investigation to resolve.
Under the hood, Stampli works with all the existing LLMs as the foundation. This allows them to choose a given model based on cost, latency, or suitability for various tasks. Multiple checks and balances are added on top.
Feldman argues that it is important for customers to start by asking how AI capabilities can provide real business value. This can reduce the risk of getting caught up in AI washing, which can confuse practical discussions. He suggests a better approach is to treat it like a new employee and see how it performs in practice:
You want to think about it a bit like when taking on an employee. When you interview the employee, they can say they can do amazing things, but then the employee is going to come and you're going to see, are they performing, or they're not performing? Are they doing the job I expect them to do, or they're not doing the job I expect them to do? And at the end of the day, it's the same thing.
Some Stampli users have already reduced the time for various AP processes by 90% with its existing capabilities. Further improvements using the new cognitive AI capabilities begs the question of whether it will result in redundancies or provide opportunities for more meaningful activities. It will probably be a bit of both.
June 2024 US Bureau of Labor Statistics data counted 308,000 finance and insurance job openings, but only 132,000 reported hires. Feldman said one cause might be that finance can be a stressful job. Teams might spend long, grueling hours looking for information, figuring out how to categorize it, and then double-checking results. Sometimes, they have to chase down managers for approvals who might be too busy to take the requests seriously.
Stampli's original vision was to make finance a little more pleasurable, which can increase retention. For example, when a finance person requests an approval, the service captures the communication chain and notifies managers when they are holding up a process.
So, how might more meaningful work improve the bottom line? One example is activity-based costing which can improve visibility into the total cost of a particular product or project. This often involves in-depth discussions with various teams to determine better labels and categories for different expenses and revenues. In other cases, greater efficiency could mean giving finance teams more time to plan new projects. Feldman explains:
If you look at the typical controller in a company, they will come with this notebook with so many items that they need to do. And the executive will come and say, 'I want you to do this project, which I think is really important for us, and I think it's going to really be very effective.' And they say, 'I want to do it, but I cannot get there until I'm not getting these things done.' And what we're doing, we're taking these things, and we nail them down for them, so they have the time to do these other important things.
Every company will strike a different approach between improving the quality and capability of finance and reducing headcount. Feldman says:
The reality is that the work is going to change dramatically, and it's already changing dramatically. Where is every company going to take that is the company's choice. You know, some companies will decide that they want to do more with less. Some companies will decide that they actually can do more with what they've got.
One important distinction for companies is whether accounts payable is just about paying the bills to keep the lights on or a strategic enabler for better business decisions. Feldman argues there is more value in the latter:
Accounts payable is there to make sure that at the end of the day, management has a very good report that reflects the effectiveness of the businesses they're managing. Accounts payable is not about making payments, but it's really a lot about how I capture this transaction and allocate that cost. And the real essence is spending time asking someone in marketing about that expense. How should I allocate it between product line A and product line B now?
But at the end of the day, it's up to the organization how they want to look at it. I think the smart organizations will basically leverage their accounting teams to be much more helpful to the business and actually to help the business have that mirror of how well we're doing or what it is that we need to do differently based on real data, based on real information of what's happening in the reality of the business.
The prevalence of AI washing increasingly means buyers have an increasingly confusing array of flash gen AI products characterized by monickers like LLMs, copilots, bots, agents, and autonomous and cognitive AIs. As Feldman points out, it's probably important to consider how easy it is to bring on a new tool, how much it costs, and how much it saves. Stampli is surfacing the new capability as a bolt-on for its existing service, which should make it easy to bring on, test, and confirm how well it works in practice.
