When integrating generative AI and ERP, focus on use cases

ERP invests in AI

Integrating AI into ERP has been happening for years, but with the rise of generative AI, ERP vendors are stating their cases more forcefully now.

SAP recently introduced Joule, a generative AI assistant that will be integrated across the company's cloud product portfolio starting with its HCM platform SuccessFactors later this year. SAP will be using third-party large language models (LLMs) to power Joule.

Previously, Microsoft Dynamics 365 ERP added generative AI assistant capabilities with Copilot across several applications including finance and supply chain management. In January, Microsoft said it was investing $10 billion in OpenAI, maker of ChatGPT.

These new AI assistants will be added to a lineup that existed before the rise of generative AI. Infor's Coleman was rolled out in 2018, while Epicor introduced its Epicor Virtual Agent (EVA) in 2019.

Although the goals are essentially similar, there are differences between the underpinnings of the new generative AI tools and the older AI assistants. SAP's Joule and Microsoft's Copilot use a variety of third-party LLMs and data from SAP and Microsoft business applications. Infor's Coleman and Epicor's EVA use natural language processing user interfaces and connect with different data sources via APIs.

Strategy, protocols and data

Although the generative AI integrations for ERP tools are plentiful, industry experts caution customers to have firm goals and strategies in place before they deploy.

One factor to keep in mind is that the integration of generative AI and ERP is still in an early evolutionary stage, and no one should draw conclusions on where things are headed, said Brian Sommer, founder and president of TechVentive.

These protocols must be there to protect the customer who's going to use this stuff and the customer's data from all kinds of unintended consequences. Without this, it's not even worth engaging with the vendor.

He suggested customers ask vendors to provide a clearly articulated strategy and set of protocols that specify the types of data that will train the LLMs and how customer data will be protected before implementing.

"These protocols must be there to protect the customer who's going to use this stuff and the customer's data from all kinds of unintended consequences," Sommer said. "Without this, it's not even worth engaging with the vendor."

ERP vendors are using generative AI in chatbots and assistants to provide a better customer service experience, but it will be embedded more deeply into ERP technology, he said.

"We'll see AI used to actually generate code," Sommer said. "For example, you can now feed it and let it train on all of SAP ABAP code, and then give it some pseudo-code specifications for a new module, and it could generate new ABAP code."

Customers also need to consider the data that's being used to train the LLMs, said Joshua Greenbaum, principal at Enterprise Applications Consulting.

SAP has made a case that its data is positioned to underpin generative AI-based business decisions because its technology underpins a good chunk of global business transactions, Greenbaum said. This means SAP has deep repositories on business processes, successes and failures, which could be useful in populating ERP LLMs, depending on its quality, governance and accessibility, he said.

Because LLMs will be used to drive important decision-making, high data quality is critical for generative AI, Greenbaum added.

"Enterprise data quality has been the biggest problem in enterprise migration since the dawn of time," he said.

Sommer echoed the point about data fidelity, saying, "AI stops providing value when it runs out of meaningful data. The questions for vendors are, how do they know which customers are going to run out of AI capability because of a data limitation and [what is] the telltale sign for knowing when that will happen?"

Vendors don't have adequate responses for these questions yet, Sommer said. Some vendors say they can aggregate data from different customers into one LLM; others claim they can develop one AI that works with the company's internal data for a private AI system and another that is trained with outside data stores.

"That sounds good, but the problem is, to get anything statistically significant on the private side, it has to be a big company, and it will be driven off of things that have huge transaction volumes," he said.

Prediction is still powerful

Customers might be willing to pay ERP vendors premiums for generative AI, but likely only if they can connect it to specific industry use cases, said Vinnie Mirchandani, analyst and founder of Deal Architect, an enterprise industry-focused blog.

Vendors will need to gather data from high-value applications, which will require employing domain experts and finding domain-specific data from sources that are not always accessible in the cloud, Mirchandani said.

This kind of data will not come cheap. He pointed to Oracle's purchase of Cerner for $28 billion as an example.

But if vendors make these kinds of investments in industry-specific data, they can offer AI functionality that provides more value than generative AI right now, Mirchandani said.

"If you can preclude unplanned shutdowns of expensive assets with preventive maintenance AI or you can reduce production and logistics footprint, waste and scrap through better demand forecasting AI, that may be exactly what your customers need," he said in a blog post.

The best business model for AI in ERP is for vendors to find out what their customers are willing to pay for as premium use cases, then figure out how to acquire the infrastructure, domain knowledge and data to train the AI systems for those use cases, Mirchandani said.

Put AI to internal use

In some cases, the intent of how generative AI is used in ERP needs to be broadened, said Jon Reed, co-founder of Diginomica, an enterprise industry analysis firm. AI could be used to solve internal pain points other than the ones that vendors are putting forward right now.

For example, SAP customers should be able to use generative AI to ease their migrations to S/4HANA Cloud, he said. Customers should also be able to apply AI toward optimizing their licensing and data agreements with vendors.

"In theory, if these AI tools are so great, [customers] should never be out of compliance with any digital data implications of third-party issues or any product licensing clashes," Reed said. "If this technology is so great, shouldn't it be used for that?"

Jim O'Donnell is a senior news writer who covers ERP and other enterprise applications for TechTarget Editorial.