AI Readiness for Procurement Teams

Good afternoon, everyone. Welcome to episode seven of our AI Readiness series. And this episode, we will be talking about what it means to be ready for AI in procurement. And I am Balaj Horbath. I'm the founder of Visual Labs, and I've been working all my professional life with ERP and CRM solutions. And for a good few years, I actually worked as a procurement functional consultant for international ERP projects. So this topic is very close to my heart, and today, I will be addressing a problem about AI based OCR scanning and invoice processing, which I know is is is a prime place for automation for for many clients. So let's without much ado, let's dive into this because we do have two different actual demos that I'd like to show you. So before we jump into the procurement part, there are a couple of concepts that I'd like to clarify and a couple of beliefs that need to be addressed. So many people think that AI is easy. You can just turn it on and get value out of it. But in reality, AI isn't easy unless your organization is ready for it. So to get those foundational pieces, your organization needs to be prepared and ready for AI. To do that, every every department, every business function needs to take different steps, and that's why we have created this webinar to show you what it means to be ready for each function. In the previous weeks, we have covered marketing, sales, customer service, field service, and now we're moving into more of the back office ERP functions. Today, we'll be talking about procurement. Next week is going to be finance. As a matter of fact, not next week, we're going to be taking a bit of a hiatus, but we'll continue with finance and HR and IT in the remainder of these sessions. So organizational readiness is a key point. And another misbelief is very many people think that AI needs to be perfect. In reality, you just need to get going. You just need to get there so that you can you can get started and start learning the benefits. Another one is Copilot is transformation. But in fairness, Copilot is not AI transformation. Copilot just gives you efficiency, gives your employees a more efficient way of working. But what they really need is AI agents, which is what will bring the proper value, and it's not Copilot. So this is why we've created the this webinar series so that you can you can get closer to these topics. And about Visual Labs, we are a Microsoft Dynamics and Power Platform consulting company. We help our clients make the best of their ERP CRM systems, and we implement them, support them, and more recently, them into the AI enabled world. As you can see, we've got all the badges. Our colleagues have all the certificates that is needed to do their job. This is our wall of pride. And and without further ado, let me jump into why this is such a hot topic, why this is so important. So for very, very many people, we see that AI is, is the buzzword and it still is, it's not slowing down as people were expecting. So most likely this is not the bubble that people have been waiting for to blow. It is still going. And Microsoft has made a study that companies who embrace AI readiness and agentic readiness will build agents two point five times faster than the companies who don't. And that knowledge will compound and you will get benefits from it. And the earlier you start, the more benefit you will gain compared to the ones who started later. So this is why we reckon that it's it's such an important piece to get going with with AI as soon as you can. And just a very few terminologies that I want to clarify. We won't go through these in detail. In the first session of the webinar series, I have covered these and you can watch them on the Wistia link. So FrontierFirm is a term coined by Microsoft, which is an organization that is human led, but AI operated. And this is the hybrid future that we are increasingly seeing where human capabilities are augmented with AI and AI agents are increasingly taking on end to end flows. Another important terminology that is being used or misused is Copilot, which is AI assistance designed to improve human workers efficiency, and coincidentally, is also Microsoft's brand for AI. It is essentially what they call the UI for AI. And in this world, we are talking increasingly about agents, which are AI assistants designed to automate and execute business processes working with or working for a human. So a lot of the people who have attended previous webinars have seen all of these. There's just one more slide that I'd like to show you here, which is the spectrum of agents. We have very basic agents that are just doing some of the bare bones basics, basic chat agents, knowledge retrieval agents. There are agents that can undertake certain tasks, and then there are all these autonomous agents that can undertake an end to end process. So as I mentioned today, we'll be talking about procurement, but where and how do we get started with procurement? And that is the question of the day. As with customer service and field service, I've come up with a maturity model that I've seen throughout the years work well for different organizations. And a lot of the organizations in the procurement space when it comes to automation and and and essentially accounts payable processes. There is no formalized procurement process. They just people just go go out to the market. They buy stuff. They if they need approval. If it's a big thing, they ought to get it via email or just phone or or verbally. But for for many cases, it's just you buy and you explain, you bring the invoice, and it gets, it gets expensed or paid. So, this is the the first level, just manual invoice processing. The second one is purchase order is getting registered at the point of commitment and you can take that and and essentially match the invoice against it. You might or might not have a product receipt or a confirmation of the of delivery in between. And next up is when you actually have a proper requisition process, requisition flow with your after that purchase order is issued to the supplier, and then a receipt is posted, whether it's services or goods, and then the invoice is registered, and some of that is automated. And then many special large organizations have come to the stage where they've got end to end automation with minimal human touch, essentially. Once a requisition gets approved, the purchase order gets out, to the supplier automatically. The supplier does their services, delivers their goods. Someone confirms that delivery. And after that, the supplier submits the invoice. If the quantity, the amount, the VAT is all correct on it, it will just get automatically processed and and picked up by the payment cycle. And that is what we're doing with end to end AP automation. And none of that is already, you know, pre generative AI has already been done by AI, mainly through OCR, optical recognition. So let's talk a bit about generative AI advances and how OCR has become really easy and simple to do previously. But there still are several OCR and API automation provider companies out there who are making a good living out of this. But what I would like to show you today is how this new new way, new new generative AI based world is making it easy. So let's look at the different steps and what companies can do to to graduate into the next step. So if your company is still doing manual invoice only procurement, then start enforcing preo creation. But if you're stuck at that stage, you can still make benefit of AI by registering invoices with AI, which is what I will show you today by two ways. And and the AI benefit is obviously reducing the time to register those invoices and and reducing human errors. The next stop is if you've got purchase orders already, you can once you have invoiced enforced PO creation, your AI opportunity is that you can match invoices against the purchase orders using AI. You you can retrieve available information from your system. And then the real benefit here is obviously the further automation and and getting automated invoice exception matching. So what happens if things don't match up? You can introduce workflows there. So to move on to the next stage, you need to introduce budgets, purchase order approvals to take out that human element, and then introduce a proper purchase requisition process. And once you do that, you can really start relying on AI driven approval processes, approval flows, messages, within certain criteria. AI may take out to approval, instead of a human person, which is what we're seeing increasingly. You have a human in the loop process, but the human in the loop can be reduced if there are certain straightforward pieces. And in order for you to evolve to the next step, which is end to end API automation, as I as I described earlier, you want to introduce a three way match with a typo there and proper matching workflows. And that will essentially get you to the nirvana of no touch procurement and payment, and you would have a PO to payment without essentially human involvement, maybe just proof of delivery. So once you do that, you have unlocked sort of the AI opportunities, and then you can rely on AI for your spend reporting and further analysis, which is not something we will cover today. So what I'd like to show you today is is essentially the piece that we we feel people are struggling with most, and that is data entry. Once you can do data entry, start registering those invoices, then it will really make it more straightforward for your users. So there are two approaches here. I won't go into too much technical detail. If you've been following our series, you have seen Copilot Studio, which is Microsoft's AI agent builder tool. And I will show you how you can well, show you what it's like when Copilot Studio is connected to an ERP, and you will just upload an invoice and ask it to create the the actual invoice entry within the system. But what we are seeing often is people will just want to, you know, email an invoice or a PDF to the accountant and ask them to register it for them. So what we have done is created a WhatsApp based agent where you can just attach the invoice paper, ask it to register, and it will automatically register that document in Business Central for you so that you don't need to, send it to an email, pick it up automatically. It will read out the details, and it will tell you what it what it ran out and do the registration. So this is something that we are still fine tuning, but we see this as a recurring pattern of how you can interact with with people who are always on the go and are struggling with with documentation, with carrying paper based receipts, so on and so forth. If they can just text essentially the ERP system via WhatsApp, that's a massive win. So let me let me jump into into our good old Copilot studio here, and I will reshare my screen with the Copilot agent. So what you'll see here is something that our team has built over one of our hackathons that we held, and we are making this into a production ready setup. Just one second until I find the screen. So here we go. Here's our ERP purchase invoice OCR agent that you can see, and what this has is several tools with Business Central. Again, not going into the technical details. It's got instructions. It's got an Excel based knowledge here. And what we'll do is just look at the tools and add an invoice up here, and I will ask it to please process this invoice. So once I click that, I can now see that I didn't restart a session, but I can see that this is from the previous session. But I can see that it's actually calling in the business central MCP demo that you can see here, and it's actually going through you can see it's thinking. So it's going through its prompt, its base prompt, what's in the base prompt, and it's actually reading out the invoice header. So this is a German invoice from LingQ. So it read out the vendor, the vendor address, the VAT number, the invoice date. It read out the customer information. Oh, it disappeared. So it got the details, and now it's actually finding actions within its toolkit to start registering, and it found the invoice registration toolkit. And the reason why it's called the reasoning model, because it's you can actually track it, but you wouldn't be seeing it. It just just does this in the background. So if you're using WhatsApp or Office or Teams, you can just pick this over and it will go through these exact steps. So it's picking out, it's using the vendor information and the purchasing voice header was created. There we go. So we already have a purchase invoice number. So this is just it's thinking. So when it's done thinking, it will give me a result, and then we'll go right into Business Central, and we can see what was created. So let me just show you here that we have this vendor card created, and our vendor doesn't have any invoices created against it. Right? Just so you see that this is actually being created real time. But now that I clicked it, that invoice was created as we spoke. So the invoice is created, but it's still not doesn't have the lines, I don't think. Right? So no lines have been added to this invoice yet because that's what it's doing now. It's actually creating the purchase invoice lines right now we can see real time what's happening in Business Central. So still no invoices. Let me actually do these side by side so we can see the thinking, and we can also see what's happening. Uh-oh. That's a rookie rookie demo mistake. It stopped the process because the test session disappeared. What I'd like to do here is start clean. I will actually there you go. It actually did create that purchase invoice, but it didn't complete it. It didn't add VAT value. It didn't add the costs to it. So this is where the process got aborted because of my root mistake. So this wouldn't happen in real life. Right? What would happen here is let's run this test again. What would happen here is it would just complete it in the background. There we go. So let me add this invoice, and I will ask it to please process invoice. So this will take a little while. Just bear with us, and I should not be moving that. Alright? So new test session. Made the same rookie demo mistake twice. Please process this invoice. And we can see the way it's thinking. Alright. There. So it's going through its base prompt, and you can see the information that it's extracting right out of here. You've got the vendor, the customer number, the VAT number, the bill to details, the invoice line details, and then the total. So once it read the details as per the prompt, I don't actually need it's interesting that this hasn't come up during our testing. So I can allow it, but I shouldn't shouldn't be needing SharePoint here. Interesting. Let me see what what this could have been. Those triggers, that's ready and available. Let's run another test. Start a new test session. This is the thing with agentic AI, and this is why we are actually doing two different topics here, so that you can that you can see that it's not deterministic. Again, going through his base prompt, reading out the details from the invoice, recognizing the header, the customer, the customer details, the invoice lines, reading out the VAT amount, the totals, and it will look at the vendor setup. So the second step is it will need to check if that vendor exists, and it does. So it's actually going to tool call a tool for search, and it recognized that that vendor already existed. If it hadn't existed, it would have set up and created the vendor. Now it's got the vendor created and it's going to start mapping the GL account based on the invoice text line. It's got the monthly membership. And this is where it starts getting really clever and and, in my opinion, slightly scary, that it's actually going to go through the it's actually going through the the chart of accounts and going through and recognizing what are the subscription fees. It found a software license main account, and it found a rental main account. So it's actually going through and finding where it should be posting it to, and we'll start essentially mapping the posting against it. So it will use the vendor ID and the invoice details, and now we shall have that invoice soon created. There you go. I should have previous run still ran, but let's look at the latest one. So this is our invoice run that was just created now, and it doesn't have a line. So what it's doing now, it's adding that invoice line to it. So I just opened that up again. Still doing the mapping against the member rental fees. Good. Creating the purchase invoice line for this account. So if I come back here, that should have it created. There you go. So that comes in, and that's mapped against the GL account of of rental fees brought in the right values. And you can see that it pulled in exactly what we needed. So it's telling me invoice processing is complete, successfully processed it. This is our invoice number, one one seven. That's correct. That was a third and successful run. And we've got the number in euros, and it quoted it against a text code, extracted, found, mapped, created, added the invoice lines correctly. So we can now validate this. Now that can disappear. We can validate this and post the invoice if we wish to do so. So that's that's the Copilot studio based direct invoice registration. And as I mentioned, we are building another option here on a different environment where a user is actually using WhatsApp to communicate with Business Central, and they're posting an invoice. So let me share my WhatsApp screen here. This you can see that we've connected to WhatsApp via Twilio. Let me make that slightly bigger. And that was my previous trial run, and I would just do that right away from from here so you can see it's actually happening real real time. And this is using a different protocol in the background, which seems to be more reliable than the MCP one. You see at the m the MCP server took a couple of runs. So if I just say, please process this invoice, it will get triggered and go. So this is doing that thinking in the background without, you know, giving me that that context. So that invoice was already processed with the same amount. Okay. So it's actually doing a duplicate check. So it's clever enough that it is checking that. Is there a specific action? Would you like me to create a vendor record or purchase order based on this information? So, yes, please register it again. So it's actually smarter than we've anticipated. This duplicate check process wasn't built into the process. So again, the beauty of generative AI. Want to create a vendor record? Do you want me to create a purchase? Let's create a vendor record. Or alright. So you can see that it's actually reading information directly out of the out of the environment, and it just created that vendor record s zed zero zero four Berlin DE vendor record created successfully, added the contract, contact, the address, and everything. So let's actually check that out. If I reshare my screen, the correct one this time around. So this is where you can see the previous runs of how we have processed invoices into the incoming document register. So Business Central has this functionality where you can just start attaching and start registering documents so that you can process them further. But let's now look at our purchasing vendors. Okay. So it jumped into a different environment, and that is, again, the demo gods not being with me here. So I want to track oh, it is. They're with me. Yeah. So this agent is actually linked to multiple environments, and I had that previously when it was jumped into another one. So it created s zed zero zero four as expected based on the WhatsApp message, and that was just created as well. So here's you see it picked up the address, the phone number, everything that we needed. So that's a clever way of building an integration on top of Business Central. This is actually not relying on the MCP. You could see that how fast it was compared to the MCP. It wasn't reasoning as much. It was just going through because we pretrained them on the on the custom APIs. So that's a pretty clever way of working on this. And I'd like to go back to my presentation. That was the demo that I wanted to show you today. And here's the here's the critical bit. So what we just saw was a manual invoice processing within directly on top of Business Central. So that was no custom code based integration. That was all AI interacting with the Business Central endpoints by an MCP or our own custom custom way of of doing it. And once you have a proper business central cloud based environment, all this becomes available to you. You could see that the technology is evolving so fast that we now have multiple ways of doing it, And we are all trialing out what works best, what is more reliable. But at this speed, everything is becoming quicker and better and more reliable by the week, essentially. So what companies should be focusing is having these baseline stepping stones in place so that when AI comes at them, they are ready for it. They have the right building blocks and foundations so that they can build on top of it. So there are a couple of things that we hear and see when we talk with our clients, why they are not ready for AI or why don't want to do AI projects. Many people think that AI just easy just happens and people will take it forward. But in reality, AI is easy when an organization is ready for it. If you've got a proper Business Central or other ERP that has an MCP server, then these use cases are pretty straightforward. They are not as complicated as we think. Or some people think that everything needs to get sorted. Everything needs to get aligned. So in reality, you can just start using this. Look. It didn't have a full vendor list. So it created the vendor. If it doesn't have an item list, it will create the item. Right? It you can just get going and start reaping benefits quite early. You just need those foundations. You just need to be on the right system, the right ecosystem. Many Many people say that, Oh, we already have Copilot, so why wouldn't we need to do AI projects? Well, what you saw here was Copilot Studio or that WhatsApp based thing. But in reality, Copilot and Copilot Studio and other custom pieces only work if you have something to base them on. The reason I and the team could do all this relatively quickly was because we are building on the Microsoft ecosystem. If you're just using Copilot, it won't be true transformation. And very many people still think that AI is expensive and it hallucinates. Look, these runs weren't easy, so they cost about twenty cents per run. But probably if you take your colleagues time to see how long it takes for them to register, it's going to be more than twenty cents. Hallucination is no longer an issue. I've run these tests over and over again as part of the prep, and it read out the correct amounts every single time. We use different formats and it didn't have an issue. So we are now at a one percent misreading rate, whereas which is probably comparable to a human error rate. So and we are still keeping the human in the loop before an invoice actually gets posted. So hallucination is less of a problem. If you're giving it too much context, if you're asking too much from it, then, yeah, then you miss miss you're misusing AI, then hallucination will be a problem. Very many people say that AI needs the right data. It is right. And in order to start using AI, you do need the right data, but you don't need to get your whole data domain sorted. You just need the critical bits in piece in in place so that you can start building them. And another important aspect is that our people won't adopt it. So if you build this right, AI doesn't need adoption, or they will solve real proper business issues, business problems. So once you can do that, then you will be able to reap the benefits of AI without actually having to rely on people adoption, or you're solving their true problems, and people will actually want it because it makes their lives easier. So the real opportunity here is once you start doing these buildings and getting benefits of AI, you will start compounding that value and start getting proper business benefit. In twenty twenty six, the companies who are not building on AI will be left behind, and the companies who are just using AI for the sake of AI and not getting proper business benefit, those AI initiatives and AI center of excellence, so on and so forth, will be shut down for not getting the right traction. So there is going to be an increasing value pressure. This is why we have come up with our own methodology, value architecture design, where we start with how a company is building, is generating business value today, what is the underlying architecture that supports it, and based on that, we can come up with a set of agentic backlogs, and then we can just get on with execution. So this is what our one day AI readiness workshop offers for you. It's an in person event with your decision makers, right stakeholders, everyone in the room, and we understand your business in the morning, how you generate value, what your challenges are. And by the afternoon, you will lead leave with a set of agentic backlogs that you can start executing, or we will help you get execute. And that will give you the right foundations so that you become self sufficient. So if that's something that's of interest to you, just scan this QR code, and I will be in touch directly with you and see if we can help you. And that was it for today. We looked at procurement. We looked at an AI based OCR and AP invoice automation. We looked at how you can chat with your ERP system back and forth, create vendors, creating voices via either WhatsApp or different means. And technology is not a question here. The question is whether you get going. So if you want to get going with this, just do scan the QR code, and we'll be in touch with you. Thank you so much. Have a good one.