rss-bridge 2026-02-12T00:02:00+00:00

SE Radio 707: Subhajit Paul on ERP Automation and AI

In this episode, Subhajit Paul joins SE Radio host Kanchan Shringi to discuss how enterprise resource planning (ERP) systems work in practice and where machine learning and generative AI are beginning to fit into real-world ERP environments. Subhajit grounds the conversation in ERP fundamentals, explaining core business flows such as order-to-cash, procure-to-pay, and plan-to-produce, and why ERP systems are central to running large enterprises. He then walks through the realities of ERP implementation, sharing examples of both successful and failed projects and highlighting common challenges around testing, process coverage, integrations, and change management. The discussion also explores how AI is being applied in ERP today, including practical ML use cases such as inventory optimization and anomaly detection, as well as emerging generative AI and agent-based approaches. Brought to you by IEEE Computer Society and IEEE Software magazine.

---

In this episode, Subhajit Paul joins SE Radio host Kanchan Shringi to discuss how enterprise resource planning (ERP) systems work in practice and where machine learning and generative AI are beginning to fit into real-world ERP environments.

Subhajit grounds the conversation in ERP fundamentals, explaining core business flows such as order-to-cash, procure-to-pay, and plan-to-produce, and why ERP systems are central to running large enterprises. He then walks through the realities of ERP implementation, sharing examples of both successful and failed projects and highlighting common challenges around testing, process coverage, integrations, and change management.

The discussion also explores how AI is being applied in ERP today, including practical ML use cases such as inventory optimization and anomaly detection, as well as emerging generative AI and agent-based approaches.

Brought to you by IEEE Computer Society and IEEE Software magazine.

---

Show Notes

#### Related Episodes

- SE Radio 698: Srujana Merugu on How to Build an LLM App

- SE Radio 641: Catherine Nelson on Machine Learning in Data Science

- SE Radio 305: Charlie Berger on Predictive Applications

- SE Radio 689: Amey Desai on the Model Context Protocol

#### Other References

- “Redefining AI in ERP Project Implementation Methodology and Future ERP AI Consultant Role in the Electronics Manufacturing Sector” by Subhajit Paul

- IEEE Article – Integration of Digital Twin and AI-enabled ERP Systems: A Conceptual Framework for Intelligent Electronics Manufacturing Services by Subhajit Paul

- SAP Sapphire session – “Mobile material placement in warehouses with AIML – SCM1661” SAP Sapphire Orlando

- ASUG Florida chapter session – “Unifying Pre-Receiving Processes in SAP: A BTP-Driven Approach to Supply Chain Continuity”

- IEEE Computer Society Professional Development Webinar – “AI in ERP – Smarter Mobile Material Placement in Warehouses“

---

#### Transcript

Transcript brought to you by IEEE software magazine.

This transcript was automatically generated. To suggest improvements in the text, please contact [email protected] and include the Episode number and URL.

Kanchan Shringi 00:00:18 Hello everyone. Welcome to Software Engineering Radio. Today we are talking about ERP or Enterprise Resource Planning systems and how they actually work in the real world. We’ll get into how ERP systems are implemented and where they tend to break. We’ll also talk about where Machine Learning and GenAI are starting to show up inside ERP and what people mean when they talk about agent existence. Basically, systems that can decide what to do next instead of just following a fixed workflow. And our guest Subhajit Paul. Subhajit has over two decades of experience leading ERP implementations at large global enterprises, particularly in electronics manufacturing and supply chain. Welcome to the show, Subhajit, great to have you here. Is there anything you’d like to add to your bio before we get started?

Subhajit Paul 00:01:10 Hi everybody. Thank you, Kanchan for having me. It’s a pleasure to join in this podcast. You have covered my introduction very well, so let’s get started.

Kanchan Shringi 00:01:20 Before we start, I would like to point out our listeners that SE Radio has several episodes that go into the background of AI and ML and LLM based apps. I will point to a few of them. A recent one is Episode 698 Srujana Merugu on How to Build an LLM App, also Episode 641, Catherine Nelson on Machine Learning in Data Science and the last one I’ll point to is Episode 305, Charlie Berger on Predictive Applications. With that, let’s start with an introductory question. For someone who’s never worked with ERP, Subhajit, how would you explain what it is and why it’s central to running a business?

Subhajit Paul 00:02:03 Enterprise Resource Planning system. That’s the full form of ERP, so we can see like in the full form al the enterprise is there. This is a total software solution for enterprise to run the business. ERP will integrate from finance to procurement to production, to inventory to sales, to logistics to shipping and each and every business system. Without ERP, all these processes will be running in silos. That will be creating inefficiency, that will be creating a lot of lagging and there will not be visibility of the business processes at any level. That’s the main reason. If you see all the Fortune 500 companies are running in one kind of ERPs or some of the companies will be running in multiple ERPs, I can tell the ERP vendors who are really leaders in ERP industry, like SAP, Oracle, Microsoft Dynamics, and now from the example, I can tell me kind of idea like who have never worked with ERP or who doesn’t know what is ERP?

Subhajit Paul 00:03:09 Let’s say without a total software solution like ERP, sales department got a sales order, so it’ll be in an Excel and then sales department will be sending those sales order requirement to production department in a mail or Excel. And then from that production department they will be requesting to the warehouse for having the components for manufacturing the finish groups. So now we can see here there are a lot of manual processes, there are a lot of manual interaction with different departments. There is a lot of data lagging and inefficiencies. It can happen, but with ERP it’ll never happen because it’ll be a total same software platform, and you can say ERP is the nervous system of an organization to run the business.

Kanchan Shringi 00:03:51 Thanks Subhajit. With that background, can you walk us through a flow how ERP shows up in day-to-day operations? You mentioned it joins multiple departments together; maybe could you talk about how a customer order turns into procurement, production, shipping and finally cash?

Subhajit Paul 00:04:13 Yeah, if we are talking about the basic ERP flows, I would like to mention three basic ERP flows. Order to cash, plan to produce, and procure to pay. Now with order to cash, ERP manages those customer orders like a sales order through delivery to customer invoicing. And then for plan to produce process, it’ll be a demand from sales order. There will be production order to manufacture the finished goods and then we procure to pay process. ERP will create purchase order to external vendors for the components which are required to manufacture the product. I can give one example from electronics manufacturing industry where the products is like laptop, mobile cloud computing, infrastructure like racks, servers, then industrial drones and you name any kind of electronics component. It’ll be a product for electronics manufacturing industry. So now let’s take one example for one familiar product like a laptop.

Subhajit Paul 00:05:15 So we all are ordering laptop. Whenever we are ordering laptop, we are basically selecting a different specification, right? Like we are selecting what will be the microprocessor speed, what will be the chassis, what will be the RAM. So from ERP point of view, these are all components, and these components are in a buildup material. That’s the way we are tracking all the components for manufacturing the laptop. Now let’s say from manufacturing the laptop, if sales department got a sales order for manufacturing let’s say a hundred numbers of laptops, for hundred numbers of laptops there will be components, all will be needed for procuring if those components will not be in the company stock. With procurement, like an external purchase order to vendors, all the components will be coming to company stock, then only all these manufacturing orders will be fulfilled with the components to manufacture the hundred numbers of laptops.

Subhajit Paul 00:06:16 After having a hundred numbers of laptops that will be delivered to customer. Now if I’m just mapping all these processes in these three basic ERP processes, so order to cash is the sales order, and delivery of a hundred numbers of laptops plan to produce, will be the manufacturing order for which will be internal to the company where these hundred numbers of laptops will be manufactured. And then procure to pay cycle will be the external purchase order that went to the vendors for supplying all the components. In this way, you can have all this business processes mapped in the three basic ERP process.

Kanchan Shringi 00:06:56 Thank you, that’s very helpful. Now before we dive in further, can you just take a step back and describe how ERP has evolved over time? In the beginning there were traditional on-prem systems, then there was the SaaS evolution to SaaS and how that set the stage for ML, GenAI and now agent-based systems.

Subhajit Paul 00:07:17 Yeah, so ERP is an old software concept, it’s nothing new, right? I would say ERP has begun maybe in 1990s and then I started working on ERP in on-premises software. It was around 2005 and then it really evolves from there. Around 2010 it went to SaaS ERP delivery model where all the infrastructures, like all the servers, application software data and managed by the ERP vendor. Company doesn’t have to take any kind of initiative for installation maintenance of the ERP software and that SaaS ERP will be available over the internet to the user. Now from SaaS ERP, like I would say two, three years back we started with AI capability. That time we are talking about ML model that is the machine learning algorithms where we can have all the patterns with the big ERP data from their ML model will be having predictive analytics for any kind of business function like anomaly detection, like warehouse efficiency improvement. And now we are talking about GenAI agents. If we are having a Chatbot, we are always familiar with the Chatbots. Now if we are thinking about the Chatbot that will be directing us, how to operate the business process, that’s where the GenAI agents have come. Like you can just write in a ChatGPT like thing and then that GenAI agent will direct you for the correct insight from the ERP. This is the way it really evaluated.

Kanchan Shringi 00:08:59 Do you have a real example of what happens when an ERP system goes down or went down? What broke first in the business, and did it escalate? Just walk us through what the chaos might have insured.

Subhajit Paul 00:09:12 Yeah, it’ll be a complete chaos obviously because as I said earlier, ERP system is the nervous system of a business. If ERP system goes down, then production order will be halted, procurement will not be happening because the buyer will not be sending the purchase order to vendors. Shipment will be delayed; order will not be fulfilled. So there will be lot of revenue loss. So, to have this kind of downtime — because ERP is a software, right? — so, for any kind of software this update maintenance is required. So, to avoid this kind of disruption, there are planned downtime that is required always for having the update in the ERP for latest kind of software update. And that time there will be a different strategy that organization needs to take. Like there will be a lot of an offline entries like the manual entries that will be done and after that it’ll be inferred in the ERP. In that way those kind of downtime can be managed.

Kanchan Shringi 00:10:13 Got it. Now that we’ve covered the basics of ERP, let’s talk about implementation. What does an ERP implementation look like from start to go live?

Subhajit Paul 00:10:25 Yeah, ERP implementation is a structured process. I would say it’ll be six to seven stages. It’ll be starting with the project preparation where the business problems will be finalized, and the budget will be finalized. Project manager and the project team will be assigned and then there will be business blueprint phase where that business problem will be discussed with business team to understand what will be the changes in the ERP system. And then as is send two B documents will be prepared. And then in the next phase it’ll be system design and realization phase where project team will be changing in the ERP system. It can be an enhancement for the existing functionality, or it can be a new transaction that is needed for users, or it can be the integration with the third party systems. Like it can be MES system that needs to be integrated with the ERP, that is the manufacturing execution system, which is very common in electronics manufacturing industry because all the production operations in electronics manufacturing needs to be tracked at very granular level.

Subhajit Paul 00:11:37 MES system will be integrated for that. We need to develop different framework and then we have to also do different framework for business-to-business communication with the different customers and vendors like we have to send maybe inventory data to the customers, and they want to have the visibility of company stock, how the raw materials are coming. According to that they’ll be placing the orders. Having said this will be the phase where ERP team will be really doing the changes in the system and then there will be the next stage where all these changes will be tested. It can be testing within ERP system; it can be testing with the MES system. For MES system there will be different team. It can be the testing with the vendor or customer team because vendor and customer also needs to test their own functionality with this B2B communication.

Subhajit Paul 00:12:37 After all these testing successful, there will be final preparation and go live phase where there will be user training because user needs to know what will be the new functionalities. Already some of the users like super users, they will be testing those functionalities and after they are testing certification only, all these new changes will be going to the production. So for all other users there should be training to operate the new functionalities. And then for new site implementation there will be data migration, like there may be some of the new inventory that needs to be updated in the new site and there will be different new bins that needs to be created. Those are my kind of example for data migration. After all this preparation, training data migration strategy program, then there will be go live. In this phase, it can be technical go live and after that it can be business global live.

Subhajit Paul 00:13:35 Sometimes what will happen, the bigger changes which can affect the whole ERP system globally, those changes will be sent at the system downtime. The business go live will not be happening immediately. That business go live will be happening after the technical go live. After going live, obviously we need to support the system if there will be any kind of issues. Usually, it’ll be saying as hyper care support, like we have to resolve that issue very fast. There will be very strict SLA for that. After this hyper care support, it’ll be handed over to support team and that will be the end of that project. This is a pretty good flow of the ERP implementation methodology.

Kanchan Shringi 00:14:19 You did call out integrations, you give an example of that, you talked about data migration and then you also said in some cases there are enhancements needed. Do you have an example of an enhancement?

Subhajit Paul 00:14:31 Yeah, sure. Let’s say we need to have one interface that needs to build so that we can send the inventory data to our customer. And we do not have this B2B communication with that customer. We have to do the changes in ERP system and we have to do that testing with the file, whatever we are sending, whether that file is getting received by the customer or not. This is a kind of enhancement that we have to do in our system as well as we have to do integration testing between customer system. This is an example like if we are building some kind of new interface.

Kanchan Shringi 00:15:14 Could you maybe share an example of an implementation that went well and talk about what made it succeed?

Subhajit Paul 00:15:21 I would like to share one of the biggest implementations that I was a part of and that went very well. It was SAP HANA implementation three to four years back. As I said, SAP is one of the biggest ERP vendors and their latest technology is 4HANA. So, HANA is in memory platform. The transactions of the ERP will be very fast. That’s the main benefit by going to SAPS/4HANA. Earlier we are having on-premises system as I said earlier, it’s a kind of evaluation. We went to Cloud and two SAPS 4/HANA Cloud. So that time three to four years ago, there is no program available from SAP like today it is available SAP rise where you can go to private cloud. That time it was not there. We have to basically go through our own methodology, own framework, own testing strategy, own role strategy role means like the authorization, what authorization will be having for the user.

Subhajit Paul 00:16:27 Like warehouse manager will be having different authorization, warehouse operator who is only doing putter or picking there will be different authorization and then all the transaction are kind of changing with the new SAP 4HANA system. So, we have to go for rigorous testing and a rigorous training because there are a lot of customization in our system. To adopt all this new process, we really need to test each and every process that nothing will fail in the production. And we went for the big bang go live where all the 100+ sites are affected worldwide. Like there are sites in US, Mexico, Brazil, Asia and in Europe. You can understand there are multilingual users in different time zone in referend business process so it was not easy to go with direct implementation. We went with a little bit phased out implementation, but the final implementation of SAP/4HANA, it was a kind of a big bang. With this implementation there is a good amount of savings, like a $1.8 million yearly cost savings and there is a lot of implementation experiences that has been shared in different SAP conferences. I would say it was a big strategy implementation that I would like to share.

Kanchan Shringi 00:17:56 What was it specifically during the implementation that caused the success?

Subhajit Paul 00:18:01 I would say it’s a proper planning, team collaboration and team coordination because there are multiple teams. There is a business team, there is a technical team, there is partner teams, there is a different vendors team. It’s a proper planning that is key of success and there is a lot of webinars that we have done. We have gone for the different languages also for doing the user training and obviously there is a technical upgradation that has been heavily supported by ERP vendor, SAP. All these things all clicked together and that’s the main success we got with very little downtime and disrupting the business.

Kanchan Shringi 00:18:46 Proper planning was a key to the success. Now could you share an example of an implementation that went sideways and why and what were the early warning signs?

Subhajit Paul 00:18:56 Yeah, I can give a recent example from our FAST SAP BTP project in inventory management where pre receiving process was defined and implemented in more than 40 sites globally, but initially it was a failure. What was the reason? What is SAP BTP? Let me tell fast. SAP BTP is the latest innovative platform from SAP where you can develop your own solution that will not be affecting ERP and all the solution will be connecting to ERP with different APIs. So that is the advantage. You are not doing a lot of customization to the ERP system. Now in this case we have developed this pre receiving platform where user will not be able to enter the material, whatever is coming from the vendor because those materials will not be having the correct number or maybe the purchase order is not having the capacity for doing the good receipt or maybe some wrong descriptions or some wrong values are there.

Subhajit Paul 00:20:02 Whatever is the reason? User is not able to do the good receipt in the ERP. That’s the business problem. After developing this platform, what will happen, buyer can go into this platform and see what was wrong with that particular good receipt and then back can correct that. With these functionalities we went live with all the processes and testing everything, but after going live user came back and told that they are not able to do the physical inventory process. What is that physical inventory process like? You will be going to the warehouse physically and you’ll be counting that and that is needed for their daily checking of the inventory. Then again, we have to stop. From project management point of view, it was a strategic project it was a very big failure. There was escalation. We have to again start from the beginning, so there will be more budget obviously needed because the project team will work with the new functionalities. And then after doing all these functionalities implemented like the physical inventory and then the other functionalities, well there they will be required different printout and different reports.

Subhajit Paul 00:21:12 Then it went live and now it is in 40 sites, it is running successfully. It’s a lesson learning and doing the correct business process testing and then it’s a success.

Kanchan Shringi 00:21:24 Those are good examples. Just thinking through them, the two examples that you walk through in the failed case, nothing really broke technically, just that a real operational process wasn’t modeled or tested and surfaced only after go live. In the successful rollout, there was proper planning, testing, training and good coordination. With that contrast in mind, let’s move to AI. What’s really new with AI? Is it adding brand new functionality or helping with some of these issues by surfacing missing processes earlier or by reducing cost and effort?

Subhajit Paul 00:22:03 Yeah, it’s a very good question. So, ERP is a static transaction, right? All our static rules, static workflows, everything is kind of static. Now with AI means its artificial intelligence, right? We are getting predictive analytics from ML model. It is basically increasing the efficiency how the ERP process will work.

Kanchan Shringi 00:22:27 Can you maybe just spend a minute on explaining practically what’s the difference between ML operative models and generative or LLMs?

Subhajit Paul 00:22:36 Yeah, sure. So, ML is Machine Learning algorithm, that means we are getting predictive analytics kind of functionality from there. It is not giving any kind of generative functionality like in with whatever we are getting from GenAI or LLM somewhere and this kind of GenAI and LLM functionalities are creating new content like creating new text, creating new documents, creating new images. These is the basic differences between these two.

Kanchan Shringi 00:23:06 Let’s start with ML, which is, as you explained, the predictive side of AI. What are me of the high impact ML different driven capabilities that you are seeing in ERP today?

Subhajit Paul 00:23:17 Yeah, I can give an example from my earlier my implementation with third party ML into SAP, like that’s the ERP that we have implemented. We have integrated that third party ML. This is for having the bin efficiency where we are talking about more than 20,000 bins in electronics warehouse. Picking costs is more effort for the due to large amount of buildup materials and it’s a big warehouse so our house operator needs to travel a lot.

Kanchan Shringi 00:23:53 Just stepping back, you’re talking about the plan to produce flow and how that is enhanced with ML? Yeah. Okay, continue please.

Subhajit Paul 00:24:01 In plan to produce what is happening that we have already discussed, there will be a production order and for that we will be required the component. Now we are talking about the components, how to get it from warehouse and I also talked about if any component will be needed. So that needs to be coming from the warehouse. Now here, to get it from the warehouse, the warehouse storage bins which will be near to the production floor. And if the warehouse operator is going to those bins and taking all the components, then this picking will be very, very fast and the production order will be fulfilled very fast. But now let’s think over like a warehouse operator needs to travel a lot, and they need to travel to the end of the warehouse and then they’ll be coming from there. The travel time will be long. So to mitigate that risk and to make that picking efficient for the faster manufacturing order, this ML model is required here to have that kind of predictive analytics which will predict what will be done, the production schedule for the future and then how the material is getting stored according to the layout of the warehouse with the bin feature near to the production flow.

Kanchan Shringi 00:25:16 How does this data actually get supplied to the model? Isn’t the model already trained?

Subhajit Paul 00:25:22 In this case the model was not trained with ERP data because we are talking about third party ML model, which are very general model which are not ERP specific pre-train model. I would say I can give some of the examples like Cains cluster running, linear regression XGBoost. Those kinds of models are basically tested. For this we have to really train those ML models with ERP data we have to send all this ERP data, like inventory data, then production schedules data then that demand production demand data because all these parameters are needed and also the physical bin sites and the package site because otherwise it’ll not be possible to calculate how much space is needed in the particular storage bin. Those data has been sent to different data lakes and from their ML model who got this data and getting trained and then come back with the bin solution and we got it from that in the ERP.

Kanchan Shringi 00:26:26 You mentioned training the models, don’t vendor supply ML models out of the box? Where are other use cases where you would extend or augment these models versus training them?

Subhajit Paul 00:26:37 Yeah, it’s a very good question because this was a question earlier in my mind also. With embedded AI functionality or the AI capability which are given by the ERP vendors, ERP vendor is already supplying the pre-trained ML model. Now in my earlier example where we need to really train the ML model for custom AI or third-party ML model like k-means clustering or linear regression or XGBoost, those ML models which are not having any context of ERP data that needs to be really, really trained. Now for embedded AI capabilities or where ERP vendor will be supplying the model. For an example, I can give very latest innovation from SAP, which is called SAP RPT1, it’s a pre-trained transformer model. This model is basically trained with SAP ERP data with programming context also. You do not have to train a lot and it can be also validated with different testing. That means whether it’s a vendor-supplied model or whether it’s a third-party ML model or the custom AI, you have to really test to see whether you will be getting the desired output or not.

Kanchan Shringi 00:27:57 So, the inventory management example about placement of inventory in the right bins in the right place. Are there other examples?

Subhajit Paul 00:28:06 I can give one example that I have as in the capability like let’s say anomaly detection and here I will be talking about the models which is already supplied by the vendor. In this case it’ll be SAP — SAP Joule — I will be referring, Joule is the conversational AI. That means it’s a Chatbot and it’s AI chatbot like a ChatGPT, but Joule will be sitting within SAP ERP. That means it’ll not be available outside like we are doing in ChatGPT and Joule will be having capability with all the ERP contexts. There will be SAP knowledge graph with that. Joule will be having all these capabilities for SAP context. In ChatGPT, if we are going and typing, we’ll not get that kind of ERP context or the inside the organization ERP data. But here if you are comparing that you will understand SAP Joule will be having that kind of capability.

Subhajit Paul 00:29:09 now for one example, anomaly detection. That is actually coming with SAP Joule with anomaly detection if any equipment that will be failed that will be predicted with that particular ML model like k-means clustering, which is basically running in the background and getting the data from IoT devices from the manufacturing line. With the historical data, that kind of ML model will be basically setting up one baseline if there will be any kind of difference data that they will be getting that, that time, that particular ML model will be basically notifying well so this particular part that can fail maybe in 10 days. There will be actionable insights, and SAP Joule will be notifying the users if there is no stock you have to procure. Do you want to create one purchase order? This kind of capability we can see in future in the ERP system.

Kanchan Shringi 00:30:18 You mentioned knowledge graph, can you provide me insight into that?

Subhajit Paul 00:30:23 SAP knowledge graph will be having all this SAP insights means SAP context, like what is there in the ERP business processes inside the ERP business data? Like what event will be happening after what? This is a kind of a knowledge graph that will be used by Joule.

Kanchan Shringi 00:30:43 And you mentioned in this case you’re able to directly use the vendor supplied ML model. There is no training or fine tuning required?

Subhajit Paul 00:30:51 In this case. It is not because this is one capability that is getting provided by SAP Joule. Like if SAP Joule is there in the ERP means in SAP, then all the underlying ML models that will be active. And if you are activating that functionality, basically you have to activate all these functionalities. If you’re activating that functionality for annually detection, then it’ll be actually already presented. You do not have to train to get the desired output.

Kanchan Shringi 00:31:21 But isn’t there a difference in the data that SAP would have trained on and what is valid anomaly at the customer side?

Subhajit Paul 00:31:30 Yeah, definitely because, like I have told about setting up the baseline, right? The setting up the baseline is actually a framing of the enterprise related data. Already that particular ML models are having the knowledge for SAP data, but the enterprise related data will be there to train the model and that will be basically happening in the background that they will be sitting on baseline with the historical data.

Kanchan Shringi 00:31:59 Where do you think most enterprises are today in terms of being ready for ML with ERP?

Subhajit Paul 00:32:05 Yeah, enterprises are, basically passed with the experimental phase. Now everybody is looking for implementing AI functionality and they’re looking for more use cases, more real-life example. With those use cases and real-life examples, those will be the parameter to decide whether leadership will be investing in AI or not.

Kanchan Shringi 00:32:28 Let’s shift to GenAI now as compared to ML, how is GenAI changing the way in which people interact with the ERP? The examples you gave for ML were more around enhancing already existing processes, is that right? And then if it GenAI, is that the same or that’s changing? Basically, the model of interacting with ERP.

Subhajit Paul 00:32:53 GenAI will definitely change the interaction with the ERP because, the thing is with ML we are getting predictive analytics, right? It is only giving the insight. Now with GenAI, it’ll be adding the conversation layer with the natural language processing and like my earlier example, SAP Joule. With SAP Joule, it’s on not only predicting, but it is al actually guiding the user like a colleague. There will be in Oracle also there will be Oracle AI agents which can summarize the documents, orchestrate the workflows, and then even create different business documents like purchase order, maintenance order or production order. These are the functionalities that we’ll be getting from GenAI where it’ll be working as a colleague, it’ll be helping the user with actionable insights.

Kanchan Shringi 00:33:54 As you explained, the conversational interface is really on top of the existing ERP workflows and context underneath these workflows could still be largely deterministic. For example, how procure to pay or plan to produce execute would not be fundamentally changed. But what has changed was the examples is that machine learning is making those deterministic flows better, like using ML to optimize the warehouse placement bins. So that picking is faster or anomaly detection to surface equipment or supply risk earlier before the workflow even hits an exception. However, as I understand with GenAI, besides the conversational interface, parts of the workflow can start to feel more agentic where the system can choose which step or tool to invoke next based on the ERP context in this case. If that’s true, how do you decide when a workflow should stay strictly deterministic versus when it starts to make sense for it to become more agentic?

Subhajit Paul 00:34:57 Deterministic orchestration is nothing new. It was there traditionally ERP also. With time it involved, like I can give an example, like PO release strategy means, purchase orders are created let’s say for a thousand dollars. Now buyer is not having the approval limit of a thousand dollars. Maybe buyer is having approval limit of $500. If any purchase order is created with thousand dollars, it’ll be going to purchase manager. Now if it’ll be created with maybe $5,000, then it’ll go to senior purchase manager. With this hierarchy and approval limit, it’s called really strategy in SAP. This is a kind of example for the workflow where that deterministic orchestration is happening. Now let’s talk about the AI agents. With AI agents there will be the task that really require the adaptability that will be requiring the problem solving. It’s not just the programming or the static workflows that will be just working with like the pure release strategy.

Subhajit Paul 00:36:09 For an example, Joule can identify high risk suppliers. I’m giving example from procurement side only, like procure to pay. Joule can have that kind of ability by analyzing the supplier performance, then compliance scores and the potential supply chain disruptions because Joules can analyze the supplier delivery data. From there Joule can see whether supplier has supplied all the orders in that correct timeline or not. That is very important because if supplier is not supplying the orders regularly, then that supplier will not be feasible to supply big orders in the future. With that it can suggest alternative suppliers, it can update directly the ERP documents for that supplier. That will be kind of a capability with Joule or AI agents.

Kanchan Shringi 00:37:06 What is the right granularity for these agents?

Subhajit Paul 00:37:09 Let’s talk about the same example. I talked about creating the purchase order earlier, right? Like release documentation or release strategy. Now in this scenario, if you’re talking about procure to pay cycle, creating the purchase order, that will be a granularity for AI agent, but it’s not the full procure to pay cycle like the supplier collaboration or the supplier determination or the good receipt because for good receipt there will be different agent. That will be the granularity level because it’ll be bounded by the business object and business object, it’ll be purchase order, sales order or production order. In procure to pay scenario, creating the purchase order will be one agent.

Kanchan Shringi 00:37:55 As you mentioned, the vendors are delivering a lot of these agents out of the box. You gave a few examples on what these agents can do. What are you seeing in practice? Are you able to adopt these agenting capabilities as delivered by the vendor or is there need for extension?

Subhajit Paul 00:38:14 Yeah, to answer this question, I need to really go to the different technical approaches or technical architectures like embedded AI MCP eight-way and I will be basically explaining all those.

Kanchan Shringi 00:38:27 Before you get into the technical implementations, can you give an example of where you feel it’ll be, take an example of the agent that you had and explain why do you think it might be important to enhance it or extend it?

Subhajit Paul 00:38:40 For an example, I’ll talked about the agent where we are doing anomaly detection. With that anomaly detection, what is happening? We are only getting the actionable insights from Joule that I need to create the purchase order. Now let’s think about one scenario where there will be an agent or there will be much more capabilities required, like whether it’ll be a purchase order or whether it’ll be a plan-to-plan transfer order. If that actionable insights are only giving purchase order functionality, then we need to write some kind of custom codes or do some kind of adjustment to have that kind of functionality where we’ll be getting that component from different plant. If that is not learned by the actionable insights of Joule, then it’ll not be there. I’m just, trying to give you an example where it can happen. That will be having the agents, which is already our ML models, which are already supplied by supplier, but sometimes it’ll not be able to fulfill the whole business requirements.

Kanchan Shringi 00:39:59 Thanks about that. Now we can get into the technical aspects that you started to talk about and you mentioned MCP or Model Control Protocol and A2A or Agent to Agent. Before you get into the details, maybe give a quick intro on what is MCP and A2A. I’d also like to point our listeners to Episode 689 Amey Desai on Model Context Protocol.

Subhajit Paul 00:40:23 The approaches and technical architectures of this MCP and agent to agent. I would say these are kind of paradigm shaped and its latest innovation in, ERP space because, all the big ERP vendors, like SAP, Oracle and Microsoft Dynamics, they have included this kind of protocols directly into ERP. I can give some examples then I can start like for embedded AI, we talked about embedded AI, right? With embedded AI we have talked about SAP Joule and then for other vendors just to have some kind of idea like, for Oracle it’ll be Oracle Fusion AI agents. Then for Microsoft Dynamics it’ll be Microsoft Dynamics Copilot, we all are Oracle Copilot, but with Microsoft Dynamics Copilot, it’s a kind of embedded AI agent that you can use AI also. Now coming back to MCP and eight-way, MCP stands for Model Context Protocol.

Subhajit Paul 00:41:29 This is defined as an open-source standard how AI application will communicate with external systems through a unified interface. Like how it is happening without MCP, right? If any AI agent, which is a foreign or AI agent or custom AI agent, it can communicate with API or it can communicate with different method where we’ll get all the communication from the third-party ML model and we’ll store in ERP, right, that we talked about. Now with MCP, this will be really easier because MCP will be the medium for communicating with any kind of custom AI with ERP. Having said that, MCP will be act as a server and custom AI will be doing query to MCP and then MCP will be having all this ERP context from ERP module and get back to the AI agent. That’s the way it’ll happen. Let’s say for an example, if any kind of SAP agent wants to reconcile accounts or supply invoices, in this kind of scenario what will happen, it’ll query to MCP server and then MCP will be querying to the invoice management model and from there it’ll be getting all this context back to the AI agent.

Kanchan Shringi 00:43:00 In this case, the vendor is actually exposing their business logic through the MCP protocol?

Subhajit Paul 00:43:06 Yes, that’s correct. It’ll be kind of a unified platform earlier. We have to create a separate, separate API for let’s say separate, separate agent or for separate, separate integration of different tools. Here we do not have to do that because if MCP will be there, MCP will take care. The traditional API method, it’ll be better than the traditional API method I would say.

Kanchan Shringi 00:43:31 Are you seeing a lot of functionality and business logic exposed as MCP service?

Subhajit Paul 00:43:36 Well, this is a really emerging technology, and we have seen the news or capabilities that is basically included in different ERP system. Like recently in SAP ticket, SAP announced about this capability. We are yet to see how it’ll be happening in future. We would like to see more use cases and real-life examples to understand how it’ll be happening.

Kanchan Shringi 00:44:03 Talk about A2A now maybe with a concrete example.

Subhajit Paul 00:44:06 Yeah, again, A2A is also an emerging technology that is agent to agent communication. It can be custom AI agent, or it can be the agent which is provided by the ERP vendor. Now with MCP, both agents are integrated with ERP right? Now if these two kinds of agents, what is provided by ERP vendor or the customer agent, if they’ll be talking with each other, it’ll be the agent-to-agent communication. For an example, if a supply chain planning agent interacts with an inventory agent and a procurement agent for proactively managing the stock levels so that, there will be new purchase order that needs to be created to place to the vendor. So that the stock will not be having any kind of shortages. In this kind of scenario, what will happen, this supply chain planning agent will be talking to inventory agent and procurement agent with A2A protocol.

Kanchan Shringi 00:45:13 Do you envision using agents across multiple vendors?

Subhajit Paul 00:45:16 Yes. that’s the capability basically. Like with MCP, it is possible like if there will be different, different vendors are creating because it doesn’t say that it has to be any ERP vendor specific agent. It can be anything like it can be custom agent also. That can be easily integrated with MCP too ERP.

Kanchan Shringi 00:45:36 You touched upon this a little bit, but maybe could you summarize on what determines whether the vendors GenAI works out of the box versus needing customization?

Subhajit Paul 00:45:45 With out of the box, like we are getting the use cases. All these use cases are very specific to the ERP software, right? It’s not specific to a company, it’s not specific to our industry. We need to have that kind of use cases which is specific to the similar company or similar industry. Because all the use cases are very generic, right? Now one size will not fit for all because every enterprise will be having that different business processes, different business requirement. For any kind of standard ERP process, when we are doing a lot of customization to fulfill the business requirement and to have that kind of capability, whatever is needed for a particular company. Let’s say for electronics manufacturing, whatever capabilities are needed that may not require for one auto manufacturer. These are the different capabilities will be required for the agents which are basically provided by the ERP vendors. This is where we need customization if the business requirement is not getting fulfilled.

Kanchan Shringi 00:47:02 With that, let’s get into our final section or discussion about how ERP implementation is changing now that AI is part of the implementation. To start with what is the process to evaluate upcoming AI enhancements? Either what vendors are supplying or the new models or new technologies?

Subhajit Paul 00:47:21 With ERP implementation methodology, we talked about six to seven phases. Now with AI integration it’ll be required similar kind of stages, but with the AI steps obviously because we need to have more AI specific activities now. The fast stage I would say it’ll be AI evaluation and strategy. Where with the business problem discussion we need to evaluate the strategy for AI. We have to talk about the data strategy of the organization. I’m not talking about any kind of, one project. I’m talking about the organization wise policy and I’m talking about the organization wise data strategy because with AI there will be a lot more dependency with the data security. After this there will be business process analysis and data integration and where there will be discussion from the earlier data strategy, how it’ll be connected with the AI.

Subhajit Paul 00:48:24 With embedded AI functionality it’ll be a little bit easier because the data directly will be connected with the embedded AI. You do not have to do lot more investment there. But with the custom AI functionality, we have to do lot more investment or strategy for integrating the data because data is the key as I said earlier. And then we need to think about the AI model selection training. With embedded AI functionality, if it’ll fulfill your requirement, then we do not have to think about, we can easily utilize the already pretend models from ERP vendors, but if it is not fulfilling the business requirement. In that case we have to really think about how to do the customization on top of the vendor provided models or if we have to take some different custom models and MCP, we can really integrate with ERP.

Subhajit Paul 00:49:27 Some of the models that I can tell, like it was a decision, Creed, Random Forest came in, Clustered Linear Regression, those are the models actually we looked earlier and these are very popular models. And then after model selection there will be the integration of AI model. If it is embedded, again, we do not have to think about that, but we need to do that testing. That is really important. Embedded AI model may not be giving you the desired output because with AI implementation you will not get the desired output on the first day. It will take time because the model will learn and with pretend model it’ll be a little bit easier. But with the custom AI model, it’ll be little bit more time for getting the desired output. Then in the next phase we have to do that AI adoption preparation. This is really important because organization change management is very critical in this scenario.

Subhajit Paul 00:50:29 We need to go for the expandability training, we need to train the user so that user will not be having that kind of feeling that AI will be replacing their jobs. AI will be helping them in the operations. And then after all this testing, training and all this policy in place, organization needs to sign different documents also because AI security policy and AI policy will be at organization level, as I said earlier. Then after that AI deployment will be happening. With AI deployment, as I said, will not get the result on the first day so we need to really monitor the changes, whatever is happening with the AI model deployment, and we need to see whether we are getting the desired output or not. If we are not getting, then we have to evaluate where is the fault, like whether we need to use more custom things to have that on top of the embedded AI or our custom AI, whatever we are using, whether we have to change that model or we have to fine tune that model. Basically, these are the phases that I can think over with AI implementation.

Kanchan Shringi 00:51:41 Stepping back a little bit, what’s your methodology to even research on what are the latest and greatest technologies or models? You mentioned several models; how do you keep up?

Subhajit Paul 00:51:53 Whatever models I have mentioned, those are actually available in GitHub. As an ERP consultant there will not be that kind of capability to select the models. Obviously, we need to work with the AI architects and that data scientist, but the training, the business process, this is the area where ERP consultants will work very closely with AI architects and data scientists because AI architects or data scientists will not be having the full-fledged idea about the business processes. Those models will be trained with the ERP business data, right? We have to really prepare the data as an ERP consultant and the correct data. That is where ERP consultant roles will be very critical. After selecting and testing, only validation will be with the business data, right? In that case, if one model will not perform with the desired output, then we have to see different other model or we have to fine tune the model. Everything is depending on the testing result.

Kanchan Shringi 00:53:05 Thanks for covering all the different new phases, but a lot of your focus, whether it was in explaining that testing and monitoring is very critical, especially to make sure that you’re actually getting the benefits of AI. You didn’t really talk about extending with MCP and eight-way, which we covered earlier in the GenAI context. Why is that?

Subhajit Paul 00:53:25 Basically we do not have any use case to tell, frankly speaking because this is really a new capability that has been shown by SAP. I’m talking about SAP because I have integrated all these things with SAP. Not only SAP if you are talking about me other big vendors like Oracle and Microsoft Dynamics. A lot of this MCP and eight-way are integrated in this year. We are yet to see the use cases. We are yet to see the pilot implementation or at least small kind of implementation with ERP partner. It doesn’t have to be that a big customer or something and somebody can do prototype and show, well this is happening like this. We are yet to see all these things. To have that kind of validation.

Kanchan Shringi 00:54:12 What are the roles or skills that now become essential for an implementer to have?

Subhajit Paul 00:54:19 I would say for AI integration and whatever we did, like we are having our ERP team and then we are having our AI central of excellence where there are AI architects and the data scientists. Now whenever we are talking about two different capabilities like ERP and AI together and all the ERP vendors are giving all the AI capabilities as an embedded kind of a model. We need to more focus on the data science as an ERP consultant and we need to know the basic AI models, how it’ll work. Like for an example SAP-RPT-1. If you are having any kind of CSV file, you will be uploading to SAP-RPT-1 and then you will be trying to ask some question there. This is a latest innovation from SAP and that has been released. I am trying to say that they will be trained where it’ll be easier for an ERP consultant having the basic knowledge of AI with their next level evaluation from ERP vendor so that ERP consultant can evaluate the models easily.

Kanchan Shringi 00:55:32 What’s your advice for teams starting to integrate AI with ERP? What is the most key technical investment or organizational investment they should make early in the journey?

Subhajit Paul 00:55:42 Yeah, from my experience, obviously it’ll be data. I told earlier also data is the key and then that AI mindset. So there will be AI awareness training, all over the organization and there will be AI security policy, and these are all organizational mandate. This is not particularly for one (?) project, there will be a lot more systems. Whenever we are talking about AI, it’ll expose a lot of business data. To mitigate the risk, we need to have AI security policy all over the organization. And then there will be a lot of training on AI awareness, not only ERP and after that, whenever ERP AI will be implemented, like will be integrating AI within ERP with a different business function, that particular business users, they have to be trained wisely so that they can understand the value of artificial in intelligence.

Kanchan Shringi 00:56:37 Is there any topic that we haven’t covered that you’d like to drill into?

Subhajit Paul 00:56:42 Yeah, I’d like to mention that in recent days I can see a lot more fraction on interest from organization’s perspective. Like in my organization, there are a lot of workshops happening for having AI in place into ERP, all the AI capabilities with different ERP partners, different ERP vendors. One point I would like to mention in each and every workshop we are looking for the use cases. We are looking for the pilot implementation at least, by the ERP partner. Those are basically missing, like there are lot more AI capabilities which are explained on the PowerPoint, but that may not be useful for a company which will be investing a lot of money on AI innovation. If there will be a lot more use cases, lot more real-life example, there may be small kind of pilot implementation. ERP vendor will be basically collaborating with customer and doing this kind of pilot implementation or small use cases. And then that will be also a good point for selling the product for ERP vendors.

Kanchan Shringi 00:57:50 Thanks for that validation that there’s a lot more happening with AI. Now, how can listeners contact you or learn about your work?

Subhajit Paul 00:57:58 Please check out my IE papers AI Redefined Implementation Methodology for ERP and then Digital Twin Integration with AI enabled ERP because digital twin is an important part for any kind of manufacturing scenario. Where digital twin has that predictive thing for having new product or a new product line. And in this context, we can get all this digital twin capability that actionable insights directly into AI and all that capabilities or insights can be used in the ERP. Apart from that, I will be having IEEE webinars, and I am also an active member of American SAP user group where I will be having presentations in EC chapters meetings or in a SAP software.

Kanchan Shringi 00:58:54 I’ll put some of those links in the show notes. Thank you so much for coming on today’s episode.

Subhajit Paul 00:58:58 Thank you very much. Thanks for having me.

[End of Audio]

---

[Original source](https://se-radio.net/2026/02/se-radio-707-subhajit-paul-on-erp-automation-and-ai/)