Medius Launches Two AI Agents, Medius Copilot and Medius Supplier Conversations, for Streamlining AP Workflows

For about eight years now, we’ve been using machine learning and neural networks to enhance invoice processing. Our machine learning technology uses pattern recognition to capture invoices, code them correctly, and route them for processing, all based on patterns specific to that company. We also use machine learning and neural networks in our Smart Flow functionality and in Medius Fraud and Risk Detection to detect indications of fraud. We’re leading the industry in our use of AI for these use cases.

What’s new is that we’re taking advantage of the recent advances in large language models (LLM), specifically in OpenAI’s GPT-4, which is the most advanced LLM on the market. OpenAI gives us the ability to interpret natural language, make decisions, and generate responses, all in natural language. We’re using LLMs now in a couple of ways to save AP managers time while improving the speed and accuracy of AP workflows.

For these new products, we’re running OpenAI models on the Microsoft Azure platform, deepening our longstanding partnership with Microsoft. Azure gives us reliability and security, so Medius customers can be confident that their data is always well-managed and stays secure, never leaving the Azure Data Centers.

Medius Copilot is an AI-powered assistant that answers common AP inquiries from within existing workflows without requiring the intervention of AP teams themselves. In a sense, Medius Copilot serves as an automated proxy for the AP team. Approvers can ask Medius Copilot questions about an invoice that shows up in their inbox. Copilot knows about the invoice from the context of data in the Medius platform and the customer’s ERP system. Using that context, Medius Copilot can use the reasoning power of LLMs to provide detailed answers. Approvers can instantly learn about the supplier, about previous invoices, read invoice comments, and get other relevant data. They get this information in moments through Medius Copilot, without having to involve the AP team at all.

This gives the approvers instant answers to their questions. And it saves AP teams time, since they’re not having to look up records and answer questions from approvers.

It also gives approvers and AP teams the confidence that best practices are being followed.

With our AI assistants, we’re trying to model the best chain of thought. For Medius Copilot, that means modeling the chain of thought that follows best practices. If you were following best practices, what steps would you follow as part of receiving an invoice, doing due diligence on invoices from that supplier, what the amounts of previous invoices from that supplier were, what comments were left about those invoices, and so on.

Today, an approver might open their inbox, find an invoice, and wonder why it was directed to them. Or, if they know why it was directed to them, they might wonder if it’s a legitimate invoice from this supplier. How are they going to look that information up? They might dig through their inbox, or they might send an email off to the AP team.

We want to eliminate that confusion and all the back-and-forth that it brings. And we want to eliminate the dread that approvers might feel whenever an invoice shows up.

With Medius Copilot, we want to put all the relevant information about an invoice at the approver’s fingertips. And we want to do that in a way that models best practices.

What is the model of the thought process in this case? They should check the previous invoices. It should check the comments on the previous invoices. It should check the approval history of the previous invoices. It should look for anomalies on the previous invoices.

This is what we consider the model thought process of an approver, and this is what we have instructed the LLM to do.

We are getting the context from our system — all the invoices flowing through Medius, including through SmartFlow and Medius Fraud and Risk Detection. So we take that context and combine it with this chain of thought of the perfect approver, so we give that to the LLM and ask it to execute this chain of thought, given this context, and say, please tell me, given this context and these steps in a chain of thought, whether I should approve this invoice. Or tell me if this invoice needs specific attention.

Medius Copilot features a natural language prompt, so approvers can ask all sorts of questions about an invoice. To help them get started, we offer several suggested questions. We never leave approvers staring at a prompt and wondering what to do.

It can immediately show previous invoices from that supplier, showing amounts, approval status, and comments. It can also analyze coding lines and dimensions and provide a recommendation for approving or not approving the invoice. Within minutes, an approver can make a decision about an invoice, confident that they’ve acted on accurate information and in accordance with best practices.

For all its responses, Medius Copilot provides links to relevant information, including our Success Portal for documentation. In this way, we’re providing explainability for our AI responses. Customers don’t have to wonder how Medius Copilot came up with something. We tell them directly.

Medius AI agents limit the questions that customers can ask to inquiries just about their own data. This ensures that one Medius customer doesn’t query an LLM to discover information about another Medius customer; the agent simply prohibits this kind of query. In fact, it won’t have the context to answer that type of question at all.

The only context Medius submits to Azure OpenAI comes from the context of one Medius customer and not only that, but also one certain user at that customer. For example, if you’re using Medius Copilot, and you start asking questions about invoices that you don’t have access to, the LLM will not give you answers. So, a manager in one division cannot ask Medius Copilot about the details of invoices in another division or even about invoices that were directed to another manager in their same division.

We will never inject any context into the LLM that an employee wouldn’t have normally. We have full control over the context. The LLM is only used to provide common sense and automated reasoning.

Because we have access to an LLM, we can do things like translate invoices on the fly from one language to another. That can be important in a global organization. For example, at Medius, our CTIO, Ahmed Fessi, lives in Paris and speaks French, English, Spanish, and Arabic. But he might get invoices written in Swedish. Medius Copilot can instantly translate the invoices into whichever language he prefers.

Medius Copilot can also read invoices using OCR and flag suspicious text, such as a line added that says, “Please pay this ASAP!!” Anomalous text like that might be an indication of fraud.

Medius Copilot also connects to the customer’s ERP system, so it can perform currency conversions automatically, using the company’s official currency conversion rates, not just some generic conversion rate discovered on the internet.

Finally, it’s worth pointing out that an LLM never gets tired, no matter how many invoices it analyzes. A human might naturally get tired after closely reviewing ten or twenty invoices. Medius Copilot never gets tired. It systematically follows the model chain of thought for every invoice it processes, whether that’s the first invoice of the day or the thirtieth.

Supplier Conversations is an AI agent — an AI application dedicated to performing specific tasks — that fields inbound email questions and requests from suppliers and answers them by retrieving and analyzing relevant data about invoices and generating natural language responses. Everyone wins: the AP team doesn’t have to spend time sorting through inbound emails, looking up financial records, and crafting responses. And suppliers get answers within minutes to the questions they’ve asked about invoices and status.

Medius gives customers full control over the level of automation that Medius Supplier Conversations provides. They can choose to review the email responses before they’re sent out, or they can choose to let Medius Supplier Conversations run autonomously, fielding questions and eliminating distractions from more important AP work.

Replying to suppliers is one of the most time-intensive manual processes that modern AP departments engage in. The inquiries are innocent enough, but they can arrive out of the blue and sometimes the tone is urgent. The supplier might be expecting an immediate reply about the status of their last invoice or if there was an issue with the payment. Meanwhile, the AP team is focused on making accurate payments and closing out the month.

AP teams can’t afford to ignore these emails. They don’t want to jeopardize their relationships with their suppliers in any way. Ideally, they’d like to answer these questions quickly and accurately, but that takes time.

Some AP departments dedicate one or more full-time employees (FTEs) to monitoring an inbox for supplier inquiries. But that is a costly approach. Workloads vary, so these FTEs might be very busy with email one week and mostly idle the next. More importantly, though, it’s now possible to answer all these questions automatically — if you have the right AI technology.

And we do. We have an LLM that can analyze the context of invoices and payments through connections to ERP systems and other financial systems.

So it’s simply a matter of having Medius Supplier Conversations do the work and freeing the AP team to work on more strategic projects.

Medius Supplier Conversations monitors an inbox for supplier inquiries. When an email comes in with a question from a supplier about an invoice, the AI agent calls OpenAI on Azure. But OpenAI itself doesn’t have information about that invoice – it lacks context about the customer’s history with this supplier – so it asks the agent to query invoice data on its behalf. For example, the agent might query the customer’s ERP system to find out about this particular invoice. The AI agent checks if the data retrieved from the context is sufficient for answering the question. If the context is not sufficient, the AI agent has the autonomy to query other context until it has all the information it needs. The agent never queries context to which the user should not have access. Once it has collected all the data it needs, it generates a response to the supplier, using language and phrasing recommended by the AP team.

They do. They can choose to review and approve each one. In this case, Medius Supplier Conversations is still saving the trouble of fielding the email and collecting all the information.

Even if a company decides to have Medius Supplier Conversations automatically answer all the routine emails it receives from suppliers, it can still raise a flag and alert the AP team if an email seems suspicious or if the email needs some kind of special handling. And it can copy the AP team on a blind cc: on every email it sends.

Medius is committed to protecting customer data. We don’t use customer data to train Azure OpenAI or any other model used by Supplier Conversations. Instead, we use customer data as a temporary context we submit to LLMs as prompts to return an answer. The LLM doesn’t learn from customer data. Instead, in the special pipeline created for a customer’s instance of an AI agent, it processes and returns an answer in a “stateless” manner.

This approach to building query pipelines is called retrieval augmented generation (RAG). RAG is useful not only because it protects the privacy of customer data; it also greatly reduces the risk that the LLM might hallucinate or “confabulate” an answer.

Our LLM architecture consists of OpenAI running Microsoft Azure. Supplier Conversations connects to Azure OpenAI to submit prompts through RAG pipelines with customer-specific data in order to return answers. The agents might do other things as well, such as look up invoice histories in a company’s ERP system, so that historical AP data can be passed along to the LLM as relevant context to be considered when generating a response.

The agents coordinate with ERP systems and other applications to collect or write data as needed. All this activity happens quickly behind the scenes when questions are submitted to either AI agent. And it all happens very quickly.

The Supplier Conversations AI agent limits the questions that suppliers can ask to only their own data, with access fully controlled by the AP team. This ensures that one supplier cannot query Azure OpenAI to discover information about another supplier working with the same or a different Medius customer; the agent simply prohibits this kind of query. In fact, it won’t have the context to answer that type of question at all.

The only context Medius submits to Azure OpenAI comes from the context of a supplier that sends an email with a question. If you’re using Supplier Conversations and start asking questions about invoices that you don’t have access to, the LLM will not give you answers.

We will never inject any context into the LLM that a supplier is not approved to see. We have full control over the context. The LLM is only used to provide common sense and automated reasoning.

That’s right. With all our AI products, we’re automating best practices. We’re never using AI to make shortcuts or making decisions on behalf of our customers. Instead, we’re using AI to consistently follow rules and the model chain of thought, including all the security practices that would be included in any model chain of thought.

Customers benefit because Medius Copilot and Medius Supplier Conversations are following model chain of thoughts quickly and automatically. These AI agents never get tired. They never make keyboard entry errors. They collect information, and if they need more information, they query to get it. And if something looks suspicious, they let people on the AP team know.

Using Medius Copilot and Medius Supplier Conversations, AP teams get to be more responsive while doing less work, improving their relationships with suppliers, approvers, and stakeholders across the organization.