By: Nelson M. Nones CPIM, Founder, Chairman and President, Geoprise Technologies Corporation
Nearly a half-century ago, Cray Research, Inc. owned the high-end scientific computing market. It faced practically no direct commercial competition ever since launching its CRAY-1 supercomputer in 1975.
But the niche market for high-end scientific computers was small, so the company’s president, John Rollwagen, engaged my father, Walter Nones, to conduct a marketing strategy study to discover new growth opportunities.
Drawing upon in-depth interviews with senior business executives across the United States, my father concluded that the commercial business data processing market long dominated by International Business Machines Corporation (IBM) was ripe for exploitation, but required a new kind of high-speed input-output bridge allowing IBM mainframe computers to offload scientific number crunching to Cray supercomputers at the back end.
Accepting those findings, Rollwagen initiated a strategic shift to multi-processor architectures and parallel computing, culminating in the launch of the CRAY X-MP in 1982. Thereafter, Cray’s revenues and net earnings grew over fourfold, from 516 million to 2.2 billion USD of revenue and 69 to 289 million USD of net earnings in today’s dollars by 1991, and the number of employees grew apace from about 1,300 to nearly 5,400.
According to a 1985 New York Times article,
“all three of the nation's largest auto makers have installed supercomputers in their design centers, seeking to test-drive their prototypes without ever building them … Aerospace and chemical manufacturers, and now even movie companies - seeking to create three-dimensional special effects for science fiction films - are also buying the machines, which provide far-greater computing power than any mainframe can deliver. ‘It's clear the market is poised for explosive growth,’ said Jeffry Canin, an analyst at Hambrecht & Quist Inc., ‘because now it's not just the Defense Department and the universities that are grabbing them up.’”
The Model Context Protocol Bridge
The Cray Research story reminds me of the way many businesses are adopting advanced, state-of-the-art frontier large language models (LLMs) today. Just as the CRAY X-MP connected powerful new supercomputers to existing business mainframes, Model Context Protocol (MCP) harnesses the reasoning, planning, and decision-making capabilities of LLMs such as Anthropic Claude, Microsoft Copilot, Google Gemini, and OpenAI ChatGPT by reliably connecting autonomous AI agents to application programming interface (API) endpoints exposed by Web applications.
Simply put, MCP gives LLMs a way to discover, rationalize, and interact with Web applications including mission-critical enterprise resources planning (ERP) software. The business opportunity driving the adoption of LLMs is the automation of repetitive, tedious data entry and reporting tasks using AI agents.
First released by Anthropic in November 2024, MCP is today’s universal standard for connecting AI agents to APIs. Any API that is documented according to the OpenAPI Specification―formerly known as the Swagger Specification―can be transformed automatically into functional MCP servers that expose the API as a tool to LLM Applications such as Claude Desktop, as shown in the illustration below. Transformation is done using popular MCP generator services including Postman, Speakeasy Gram, liblab, Prefect FastMCP, and Stainless (now part of Anthropic); and self-hosted generators such as the open-source openapi-mcp-generator are also available.
Each tool visible and available to an LLM Application has a title, description, and input schema which AI agents can discover and interpret to create natural-language requests. The MCP server converts those requests into API calls which activate specific Web application functions that generate and return the results to the AI agent.
The Legacy ERP Conundrum
Generally, ERP applications are a class of business software dominated by well-known legacy incumbents including SAP ECC and older SAP R/3 releases, Oracle E-Business Suite (Oracle EBS), Oracle NetSuite, PeopleSoft, JD Edwards, Infor, and older Microsoft products like Dynamics Great Plains (GP) and Dynamics NAV.
Those mission-critical systems were expensive to buy, even more expensive to customize and implement, and are proving exceedingly difficult to replace due to their broad enterprise footprint. Increasingly, they are obstructing the adoption of AI agents for the purpose of automating repetitive, tedious data entry and reporting tasks, because they expose few―if any―APIs that comply with the OpenAPI Specification (OAS).
In fact, not one single legacy ERP application on the market exposes its entire set of transactions (functions, screens, and reports) as OAS-compatible APIs today, and many of them require proprietary adapters or API gateways to generate OAS-compatible documentation for specific transactions like Accounts Payable or Order Management.
According to Seth Ravin, CEO of Rimini Street, a global provider of end-to-end enterprise software support,
“A Censuswide survey of 4,295 CFOs, CISOs, CIOs, and CEOs found that 70% do not see traditional ERP as the future of enterprise software. That is not a fringe opinion from a handful of disgruntled IT directors — it is a majority position among the people who sign off on enterprise technology budgets. The alternative they are increasingly pointing to is headless ERP: an architecture that separates the system that runs your business from the interface your employees actually use.”
Headless ERP Architecture And Its Pitfalls
Ravin continues,
“Traditional ERP systems are monolithic: the database, the business logic, and the user interface are bundled together and shipped as one product … Headless ERP inverts this. It separates the backend, the database, business logic, and core operational functions like finance, inventory, and HR, from the ‘head,’ the user interface layer that employees, customers, and partners actually interact with. In a headless architecture, the ERP engine exposes its functions through APIs rather than through a fixed set of screens.”
A much newer class of “API-first” or “metadata-driven” ERP applications is emerging that expose all their transactions as OAS-compatible APIs requiring no proprietary middleware or custom-coded wrappers, but the leading headless ERP challengers have two major weaknesses.
First, they are truly headless in that they provide no user interface whatsoever, which makes them impossible to use when businesses must entrust the execution of specific transactions (such as the approval and initiation of electronic payments) to people because AI agents are too risky.
Second, the leading headless ERP challengers have a much narrower ERP footprint than legacy ERP incumbents. As just one example, none of them offer any finance capabilities such as General Ledger, Accounts Payable, Accounts Receivable, or Fixed Assets.
GM-X Auto – A Revolutionary Breakthrough
Today, Geoprise Technologies became the first and only ERP software house in the world to offer an ERP application—GM-X Auto—that exposes every single one of its 3,871 functions, screens, and reports to MCP servers through 456 OAS-compatible APIs requiring no proprietary middleware or custom-coded wrappers.
Unlike the offerings of headless ERP challengers, GM-X Auto also provides a browser user interface for each of those 3,871 transactions, so they are equally accessible and useable by people as they are by AI agents. Originally introduced in 2018, the GM-X Auto user interface is fully responsive, allowing it to run equally well on smartphone, tablet, and desktop browsers.
GM-X Auto, therefore, can function in either headless or traditional (“headful”) modes. The motto for our unique capability is,
“ERP for AI when you can; for people when you must.”
When tasks must be entrusted to people instead of AI, the GM-X Auto ERP application works exactly the same as GM-X itself has done for many years.
When tasks can be delegated to AI agents, on the other hand, MCP servers expose them as tools that execute those transactions in exactly the same way that human users would when using a smartphone, tablet, or desktop browser.
Also, unlike the software capabilities of headless ERP challengers, 21 different GM-X Auto subsystems cover an exceptionally broad enterprise footprint encompassing finance (General Ledger, Accounts Receivable, Accounts Payable, Fixed Assets), customer relationship management (CRM), supplier relationship management (SRM), supply chain management (SCM, including manufacturing and logistics), strategic human resources (HR), and business process management (BPM).
How GM-X Auto Mitigates Autonomous AI Agent Risks
Without human oversight, it’s no secret that AI agents introduce major security and governance risks because of their ability to plan, decide, and act autonomously. Those risks can multiply exponentially when AI agents hold broad, long-lasting API keys or tokens, potentially allowing attackers to move laterally through an ERP system.
The GM-X Auto ERP application substantially mitigates those risks by employing exactly the same role-based access controls for AI agents that you can configure the application to enforce for human users.
As a non-negotiable best practice, Geoprise strongly recommends assigning a permanent and unique GM-X User Id to each running AI agent instance, just as you would assign a unique User Id to each human user. This practice will hold every AI agent accountable for its actions through the built-in GM-X Audit subsystem which automatically maintains a complete history of who (or which agent) changed what, what they changed, and when they changed it.
It also allows you to assign roles you configure for AI agents that grant access to the narrowest possible scope of data and transactions needed to perform the tasks delegated to AI agents, thereby enabling you to erect strong guardrails that prevent autonomous AI agents from spawning cascading streams of errors, unforeseen liabilities and financial costs, or leaking vital secrets, should they escape and run amok.
GM-X Auto also saves you money. Assigning a permanent and unique User Id to each AI agent instance can become prohibitively expensive under user-based ERP licenses or Software-as-a-Service (SaaS) subscriptions, because each one requires a separate paid-up account no matter when or how often the instance runs. But this is never a problem with the GM-X Auto ERP application, which is licensed and paid for by application server. Each GM-X Auto server supports unlimited users, whether AI agents or people.