I spend a lot of time with organizations that are under real pressure to prove AI can deliver more than a polished demo. That pressure is even greater in highly regulated environments, where every new capability has to clear a higher bar for security, governance, auditability and operational control.
That’s why this project stands out to me. A large defense-sector organization came to us with a very practical challenge: catalog management. Its Oracle E-Business Suite environment remained the system of record, but cataloguers were carrying a heavy manual burden. They had to interpret manufacturer data sheets, safety sheets, specifications and product images, then translate that information into structured item master data across a complex, multilingual environment.
The goal was not to replace the ERP, but to make the work around it faster, more consistent and easier to govern. In other words, it was a perfect use case for an enterprise AI implementation: focused, constrained and close to the workflow where value is created. As the broader AI conversation moves from ambition to execution, the question for enterprise IT leaders will no longer be whether to pursue AI, but where to apply it first, how to govern it and how to prove value without creating unnecessary risk.
Enabling AI-powered datasheet auto-fill with human review
Catalog management may not be the flashiest AI use case, but it is the kind of work that quietly determines how well procurement, maintenance and supply chain teams operate. When item descriptions are inconsistent, attributes are incomplete or duplicate items are created, the impact shows up everywhere: slower requisitioning, more friction for buyers, avoidable rework and less trustworthy master data.
In this environment, the complexity was significant. The catalog used a five-level category hierarchy with roughly 1,900 valid category combinations. For each item, a cataloguer had to determine the correct category, extract the relevant attributes and create descriptions that met the organization’s standards in multiple languages.
The AI capability was designed to assist that process without taking control of it. A cataloguer uploads a manufacturer document or product image, and the system proposes values such as item description, manufacturer part number, unit of measure, item type, currency and category-specific descriptive attributes. The AI solution does not post directly to Oracle EBS. It pre-fills a wizard for the cataloguer to review, correct and submit for approval.
This is AI-assisted cataloging, not autonomous decision-making. The vision model runs on infrastructure that the organization controls within its own network perimeter, so that catalog data, supplier data and product documentation do not leave the building. For security-sensitive industries, that architecture can be the difference between an AI project that is approvable and one that isn’t.
Building guardrails before scaling AI
The most important work in enterprise AI is often not the model itself. It is the guardrails around the model. In this case, the system was designed to make useful suggestions while keeping the organization’s rules, validations and approval processes firmly in control.
The model receives only the permitted attribute keys for the selected category and, where applicable, the permitted list-of-values options. It is instructed to use only what appears in the source material, with omissions preferred over invention. Server-side validation, language validation, schema filtering and audit logging provide additional controls.
Identity and authorization also stay where they belong: in the ERP. Users sign in through the organization’s identity provider; every action is resolved to a real ERP user, and the ERP determines what that user is allowed to access. If the system can’t verify that a user is authorized, it blocks the action by default. The AI layer never acts as a privileged service account, which means the audit trail can answer the question regulators will ask: Who did this? Were they allowed to?
There are productivity gains here, too:
- Intelligent category suggestion helps cataloguers navigate a deep hierarchy
- Part-number matching across internal and external catalogs helps prevent duplicate item creation before it happens
- Item descriptions are generated deterministically from approved category and attribute values, helping enforce naming standards by design rather than depending on policy compliance alone
This is the practical balance that I see highly regulated organizations seeking: faster, smarter workflows without giving up security, auditability or control.
3 lessons learned about enterprise AI implementation
For enterprise IT leaders, this project offers more than an interesting AI use case. The takeaways from this implementation translate into three practical rules for enterprise AI execution:
1. Innovate around the edges
The ERP stayed the untouched system of record. Every new capability was built on top of it, via supported public interfaces, without changing vendor-delivered objects. That allowed the organization to create a modern, multilingual experience and an AI assistant without taking on the cost, risk or disruption of a migration.
If your AI strategy starts with “first we replace the ERP,” you may have already lost the business case. Most organizations need a faster, lower-risk path to value, and that usually means layering innovation around the core systems they already run.
2. Be deterministic where you must, probabilistic where it pays
The AI proposes; it never commits. It reads a manufacturer document and pre-fills a wizard, which is exactly where probabilistic AI earns its keep: interpretation. But the data that actually reaches the item master must be governed, validated and approved.
That is why item descriptions are generated deterministically from approved values in all required languages. Naming standards are enforced by construction rather than by hoping people follow policy. The lesson is straightforward: know which parts of your workflow must be exact, and do not hand those decisions to a probabilistic system.
3. Design human oversight from the start
Nothing the AI produces reaches the system of record without a person reviewing, correcting and approving it. Oversight is not a compliance afterthought bolted on before go-live. It is the workflow itself.
That is what makes the capability approvable in a security-sensitive environment and also makes it trustworthy enough for people to use. In regulated industries, trust comes from visible control, not from promising the model will behave.
Practical enterprise AI starts with the right problem
This is deliberately bounded AI assistance. The model interprets; people and deterministic controls govern what reaches the ERP. There are no agents here and no self-directed workflows. In a defense environment, bounded, human-reviewed AI is not the cautious choice. It is the correct first capability. Prove the pattern where the blast radius is small, then extend it.
Enterprise AI delivers value when it is grounded in real workflows, constrained by strong governance and embedded where people already make decisions. This defense-sector catalog project is a useful proof point, showing that even complex, security-conscious organizations can move quickly when AI is applied to a well-defined business problem.
At Rimini Street, we use the Rimini Smart Path™ methodology to help organizations pursue modernization: keep the governed ERP core stable and supported, optimize the experience around it, then introduce Agentic AI capabilities where they create measurable value without forcing a disruptive replatforming project.
For IT leaders, the Rimini Smart Path™ turns modernization into a practical, outcome-driven roadmap. This flexible, lower-risk approach helps organizations address the rapid advancement of AI, constrained IT budgets and rising demand for innovation to move faster without surrendering control.
- Watch our webinar series in partnership with MIT Technology Review for more insights on building a strong foundation for your AI strategy.
- Learn more about how we help clients take a practical path to Agentic AI with our Rimini Street Agentic AI ERP solutions.
Key takeaways
Enterprise AI can address high-friction workflows without replacing the ERP systems organizations already rely on. Use AI to interpret unstructured information, while deterministic rules enforce validation and naming standards. Build human review, authorization and auditability into the workflow so AI suggestions cannot become ERP records without approval.
