
Ask technology leaders at large pharmaceutical or medical-device companies whether they are exploring generative AI, and many will describe active pilots. Ask how widely those systems can safely use internal, regulated documents, and the discussion becomes more cautious. The obstacle is not only model capability. It is whether the systems can operate within the controls that govern sensitive business information.
In life sciences and medical-device manufacturing, access to information can depend on a person’s role, site, product line, study and the jurisdiction that governs the data. Documents may be subject to access controls, data-residency requirements, validation procedures and audit obligations. An AI assistant that returns information to someone who was not authorized to see it can create a serious governance problem, even if the answer itself is factually correct.
Vishnu Vardhan Reddy Yeruva, a solution architect at MyGo Consulting Inc., has worked on enterprise technology projects involving pharmaceutical, medical-device and specialty-manufacturing operations in the United States. His work focuses on integrating enterprise systems and designing AI-enabled workflows for environments where access control, operational reliability and auditability matter.
“In a regulated company, an AI system has not succeeded simply because it gives the right answer. It must also respect what the user is permitted to access and make the result reviewable.”
The gap between AI adoption and governance
Retrieval-augmented generation, commonly known as RAG, allows an AI model to consult source documents before responding. In a simple demonstration, the model searches a collection of documents and produces a fluent answer. In a regulated enterprise, however, the system must also preserve the source systems’ permissions and provide a reliable account of how information was retrieved and used.
If authorization checks are applied only after retrieval, restricted information may enter the processing pipeline before it is filtered. If controls are applied too broadly, employees may receive incomplete results and return to manual work. A governance-aware design therefore needs to consider permissions as part of the retrieval process, not merely as a final presentation-layer check.
Yeruva’s enterprise architecture work has included systems designed to apply source-system permissions during retrieval and make results easier for compliance teams to review. He is identified as a named inventor on a pending U.S. patent application related to enterprise retrieval optimization. The application’s status, scope and any public details should be confirmed against the official filing before publication.
A supply-chain example: reconciling data across systems
Some of the most consequential enterprise problems are less visible than an AI chatbot. In one large pharmaceutical and medical-device manufacturing engagement, inventory information was distributed across SAP enterprise resource planning software, third-party logistics providers and trading partners exchanging electronic data interchange (EDI) transactions. Records could differ because of timing, partial matches or inconsistencies between systems, leaving teams to reconcile reports manually.
Yeruva designed and built an automated reconciliation layer on SAP Business Technology Platform. The system brings together EDI transactions and internal inventory movements, normalizes records, matches transactions and routes exceptions for human review. The intended operational benefit is to move away from periodic, manual reconciliation toward more continuous visibility into discrepancies.
The beneficiary reports that his project work has reduced operational processing time by approximately 50 percent and improved system performance by around 30 percent. Those figures should be treated as reported outcomes until supported by documented baselines, measurement periods and project records. Any final publication should describe precisely which system or engagement each metric covers.
The distinction matters in regulated supply chains. A discrepancy may be more than an accounting issue: depending on the product and process, it can affect traceability, inventory planning and the ability to investigate a problem. Automated reconciliation does not replace human accountability, but it can help teams identify exceptions earlier and focus attention where review is needed.
Why the issue extends beyond one employer
The broader challenge is how to deploy AI in industries where mistakes can affect sensitive information, manufacturing operations and product traceability. Pharmaceutical companies, medical-device manufacturers and advanced manufacturers operate under complex requirements and depend on interconnected systems. AI tools that cannot respect permissions or explain the provenance of their outputs may be difficult to approve for important workflows.
For that reason, the potential value of governance-aware AI is not limited to a single installation. Reusable methods for permission-aware retrieval, source traceability and auditable workflows could be relevant to other organizations facing similar technical constraints. Whether those benefits materialize at national scale depends on the systems’ performance, validation, adoption and the extent to which approaches can be reused across organizations.
“The goal should be for permission checks and traceable sources to be built into enterprise AI from the beginning, rather than treated as optional safeguards.”
From enterprise delivery to further research
Yeruva began his software-engineering career in India in 2019 and completed a master’s degree in software engineering at San José State University in 2023. He has undertaken work in enterprise application development, analytics and applied AI, and holds SAP Business Technology Platform certifications. Two research papers were reported as accepted at IEEE-affiliated international conferences in 2026; their titles, venues and acceptance status should be verified before publication.
His planned research addresses a related trust problem: whether data extracted by an AI system from a document can be traced to the precise passage that supports each extracted field. He is preparing an independent study on field-level verification for document-to-structured-data extraction, using public, permissively licensed datasets and open-source tools. The work is intended to examine whether systems can provide evidence for individual outputs rather than merely produce plausible-looking values.
That question is particularly relevant where records must be reviewed or validated. A system may return a correct value without making its basis clear; in a controlled environment, that can make the result difficult to verify. Research into provenance and field-level grounding could help establish more transparent evaluation methods, although the practical impact will depend on the study’s findings and subsequent adoption.
A practical direction for enterprise AI
The next stage of enterprise AI will require more than better model responses. Organizations will need architectures that preserve existing permissions, provide traceable sources, expose uncertainty and fit within operational review processes. For regulated industries, these safeguards can be part of the foundation that makes responsible adoption possible.
Yeruva’s work sits at the intersection of enterprise platforms, applied AI and regulated operations. His proposed research into source-grounded extraction is a continuation of that focus: making AI outputs easier to verify and govern. The broader test is whether these methods can be validated, reused and adopted beyond the projects in which they were first developed.
Internal verification checklist – remove before media submission
These notes are for preparation only. They are not part of the proposed article and should not be sent to an editor.
- Confirm every biographical fact, job title, employment date, degree and certification directly with Vishnu and reliable records.
- Verify the quoted statements word for word and obtain Vishnu’s approval for any edited quotations. Do not present reconstructed quotations as verbatim if they were not actually said.
- Substantiate the reported 50% processing-time reduction and 30% performance improvement with project documentation, baseline definitions and measurement periods. Remove or qualify metrics that cannot be supported.
- Confirm the patent application number, filing type, filing date and named-inventor status. Do not disclose confidential or non-public patent information.
- Verify the two conference-paper acceptances and name the venues only if documented.
- Obtain written clearance before identifying clients or disclosing confidential project details. Keep client descriptions generic unless authorized.
- This is a proposed pitch, not a published article. Do not represent it as published or as independent editorial coverage unless an outlet independently accepts and publishes it.