The most powerful language models run in the cloud. The most sensitive corporate data — customer names, contract terms, internal business information — must not easily reach external services. This seems like a contradiction. It isn't, if the architecture is right.
This post explains how companies can operate AI agents with cloud models without disclosing personal or confidential data.
The Problem: Where Does the Agent Process the Data?
When an AI agent processes a customer inquiry, it processes data: names, addresses, order numbers, contract information, internal price lists, and potentially product issues that are not yet publicly known. Transmitting all of this unfiltered to a cloud language model carries significant data protection risks — and in many companies, it is simply not permitted by their own compliance department.
„The objection that data is too sensitive for AI is not a technical judgment — it is an indication of missing architecture."
The solution lies not in foregoing powerful cloud models, but in separation: What is processed where?
The Solution: Three Processing Stages
A hybrid architecture divides processing into three clearly defined stages.
Stage 1 — Local Pre-processing: A locally operated model analyzes the incoming request and extracts the factually relevant information. Everything personal — name, address, customer number — is replaced by abstract identifiers. This processing takes place entirely within the corporate network. No personal data leaves the company's own infrastructure.
Stage 2 — Cloud Decision: The prepared package with the abstract identifiers is passed to the cloud model. The model makes the process decision based on factual information — customer segment, product category, type of request. It sees no real people, no real addresses.
Stage 3 — Re-enrichment: A local service within the corporate network replaces the abstract identifiers with the actual data and compiles the final response. The personalized response to the customer is generated entirely internally.
Practice: The Warranty Case of Markus Meier
Markus Meier (Meier IT-Services, Gold Customer) has reported three defective Smart Hubs. His email contains his name, the company address, the order number, and internal contract data.
In Stage 1, the local model extracts the factual core information: a Gold customer, a defective product of category 9942-X, a warranty claim. Everything personal is internally replaced by a neutral customer identifier. What goes to the cloud model contains no real name and no real address.
The cloud model in Stage 2 makes the decision: express replacement according to the Gold contract, approval from the regional manager required due to a reported supply shortage for this item.
In Stage 3, the local service fills the response template with the actual data. Markus Meier receives a personalized, correct answer — even though his name never reached the cloud model.
Classification: Who Makes This Decision in the Company
The question of which data may go to the cloud and which remains local is not a technical footnote. It is a fundamental business and legal decision that must be made before an agent's first production deployment.
In a corporate context, this is the task of an AI Governance and Compliance Officer: classifying data categories, defining processing boundaries, and ensuring conformity with the GDPR and internal guidelines.
Conclusion
The objection "our data is too sensitive for AI" is justified when it refers to uncontrolled cloud transmission. It loses its basis when the architecture separates local pre-processing and cloud decision logic from the outset.
Blog 06 poses the next question: What happens if the cloud model changes — without anyone having touched anything?
The prompt is just the beginning. Architecture decides which data the cloud never sees.