The Agentic AI Elephant In the Room
- Charles Rathmann
- Apr 18
- 5 min read
Updated: 3 days ago
AI-Enabled ERP Requires Modern APIs, Granular Web Services and Consistency
Success with agentic AI (artificial intelligence) depends on more than a well-trained AI model.
It needs enterprise software, usually enterprise resource planning (ERP), organized on a granular enough level to present information and allow required actions. It must be consistent across a broad spectrum of functionality. And it requires ERP that speaks the lingo. And that lingo is JSON.
Let’s define some terms, and then figure out how each of them plays into the agentic AI equation.
What Is Agentic AI?
Researchers from MIT and Harvard in a 2025 paper defined AI agents as autonomous software system utilities that sense, reason, and execute in a digital environment against goals set forward by humans. They can use tools, complete transactions and interact with people and other systems.
On its Hype Cycle, Gartner recently positioned agentic AI at the Peak of Inflated Expectations. And one reason for this is not just the reasoning abilities of large language models (LLMs). Once these LLMs are integrated with ERP software, where the tools and resources and transactions agentic AI needs are located, it still needs to understand the underlying business functionality and data set.
Once it understands the data, it can act as a guide-on-the side tool, offering recommendations. To be able to act within the application, it must gain access to the levers that trigger business actions.
RESTful APIs, JSON and Agentic AI
This is where JSON plays a role. JSON (JavaScript Object Notation) is a lightweight, text-based, language-independent data-interchange format used to store and exchange data, primarily for web APIs and configuration files.
But not every ERP system does JSON well. ERP and enterprise software built on RESTful APIs generate JSON messages natively.
As outlined recently here on The Explainer Blog:
“Representational State Transfer (REST) is an architectural style for designing networked applications, primarily over HTTP, using a stateless, client-server model. It enables efficient, standardized communication by treating data as resources, manipulated through standard methods like GET, POST, PUT, and DELETE, making it popular for web services. REST is a more modern method than Simple Object Access Protocol (SOAP), which communicates to other applications using .XML messages. RESTful APIs on the other hand can support multiple protocols including JSON, XML, HTML, YAML, and CSV.”
JSON is human-readable, easy to parse, and based on key/value pairs and ordered lists, making it a popular alternative to XML. JSON enables agentic AI to do things—if the LLM is the brain, JSON messages back and forth between systems and within the ERP software are the hands.
Asynchronous methods like JSON enable AI agents to retrieve real-time data, perform create, read, update, or delete (CRUD) operations, and engage with public and private web services.
JSON enables agents to work in real-time by receiving data updates and executing tasks, interacting with web applications on behalf of a user, and automating multi-step tasks. It does this better than older-school APIs built on simplified object access protocol (SOAP).
TAKEAWAY: In Agentic AI, LLMs are the brains, but JSON through APIs are the hands.
Why is this? XML files are long and complex, and must adhere to strict rules. JSON files are smaller, can extend to unexpected, new fields and data types. JSON is lingau franca of programming languages like Python and is the foundation for emerging standards for AI including:
What’s more, trying to run an LLM with XML files would be slower and consume more tokens. And while RESTful APIs can output a broad spectrum of file types, including JSON, SOAP will only output XML.
Apart from RESTful APIs, ERP must ensure the API is not an afterthought, Scotch-taped onto the outside of a legacy application. This afterthought approach results in a limited number of points at which an AI application can access data to take agentic action in an application.
Granularity Matters to Agentic AI
A modern ERP should be built natively to use RESTful APIs inside the application, as well as as integration points, yields hundreds of projections, which select specific fields available to an API, and thousands of endoints, the locations where data can be found as a web service for integration.
IFS Cloud, for instance, inherits the functional granularity of IFS Applications, which since IFS Applications 7.5 was already built around thousands of components, all available as web services. Marketing of the applications leaned heavily on the ability to deliver the “extended enterprise”—essentially opening the ERP stack to external software and systems.
By comparison, Sage X3 relies heavily on SOAP (Sage is working to rapidly build out its REST capabilities), while flagship Sage Intacct offers a more granular design built on RESTful APIs.
TAKEAWAY: Agentic AI requires ERP built on RESTful APIs. If possible, these APIs should be used on a granular level within the application, as well as for integration points.
Consistency Matters to Agentic AI
The pertinence of the Sage X3/IFS Cloud comparison does not end with SOAP vs REST.
JSON may be the lingua franca of agentic AI, but if data models are not consistent or processes not buttoned down in the application, AI will only be able to do so much with it.
A pure-play enterprise suite of modules is designed to consistent standards to work together support AI well. Enterprise suites comprised of multiple software products united through merger and acquisition activity, meanwhile, will throw up roadblocks to AI.
Prior to being acquired by Sage and rebranded as Sage X3, the Adonix ERP product grew multiple times by acquisition. Adonix acquired Prodstar, Italian ERP vendor Gruppo Formula SPA, and North American vendor GSI Transcomm (formerly ADP-GSI).
IFS has undertaken acquisitions as well, most notably Metrix and 360 Scheduling, which comprise the company’s highly-touted field service management capability.
AI may not care where the software came from, but it does care if it treats data and processes the same throughout an application suite. In some cases, the quality of the processes facilitated by the software will present barriers as well—modern ERP with strong business rules will be friendlier to agentic AI to those that enable rapidly-changing, ad hoc processes that may not respect segregation of duties, GAAP or enterprise-wide data standards.
TAKEAWAY: Like a gothic horror story, agentic AI can go horribly wrong due to the sins of the past. Were multiple acquisitions of software functionality well-rationalized in the product? Were all products acquired built to modern standards? If not, these ghosts will come back to haunt your agentic AI efforts.