When a B2B application gains an AI feature — summarising support tickets, drafting quotes, answering questions over a document archive — a new data flow appears. Records that used to travel only between your database and your users now travel to a model, often hosted by a third party, sometimes outside the EU. Every prompt is a potential transfer of personal data, trade secrets, or both. The GDPR did not stop applying when AI arrived; it simply gained a new, very busy pipe.
Start with roles and contracts. In most B2B integrations your customer is the controller, you are the processor and the model provider is your sub-processor. That chain has to exist on paper: a data processing agreement under Article 28, an up-to-date list of sub-processors, and clear answers to the three questions every customer DPO will ask — where is the data processed, how long is it kept, and is it used to train the provider's models? Enterprise tiers from the major AI providers typically offer EU data residency, no training on customer data and zero or short retention, but these are options to contract for and switch on, not defaults to assume. If data leaves the EU, the transfer needs a legal basis: the EU–US Data Privacy Framework, or standard contractual clauses backed by a transfer impact assessment.
Then design for minimisation. A model rarely needs the full customer record to do its job. Mask or pseudonymise names, email addresses, tax codes and IBANs before the prompt leaves your perimeter, and re-identify the answer on the way back. In retrieval-augmented generation (RAG), the vector index must enforce the same permissions as the source system: if a user cannot open a contract in the document management system, the assistant must not be able to quote it. Permission-aware retrieval is the most overlooked control in enterprise AI projects — and the one that leaks the most when it is missing.
AI also brings a new attack surface. Prompt injection — instructions hidden in an email, a PDF or a web page that the model reads — can push an assistant to disclose data or call tools it should not. AI agents that can act, sending emails, updating records or calling APIs, widen the blast radius. The countermeasures are classic security engineering applied to a new component: least-privilege credentials for every tool the model can use, human confirmation for irreversible actions, output filtering, rate limits, and complete logging of prompts and tool calls, stored under the same retention and access rules as the data they contain.
When AI influences decisions about people — screening candidates, scoring creditworthiness, evaluating employees — the bar rises. A data protection impact assessment under Article 35 becomes the rule rather than the exception, Article 22 limits decisions based solely on automated processing, and the AI Act may classify the system as high-risk.
On the AI Act front, the calendar has just moved. The Digital Omnibus on AI — Regulation (EU) 2026/1744, in force since 27 July 2026 — postponed the obligations for stand-alone high-risk systems from August 2026 to 2 December 2027, and those for AI embedded in regulated products to 2 August 2028. It is a delay, not a repeal. Obligations for general-purpose AI models have applied since August 2025, and the Article 50 transparency rules — telling users they are interacting with an AI, marking synthetic content — apply from 2 August 2026, with time until 2 December 2026 for marking generated content in systems already on the market. In Italy, Law 132/2025 has added national AI rules since October 2025. A separate omnibus also proposes changes to the GDPR itself, including a new Article 88c on legitimate interest for developing and operating AI: worth tracking, not worth betting a project on.
The practical lesson is that compliance is cheaper as architecture than as paperwork. An AI gateway between your applications and the models — handling masking, routing to EU endpoints, permissions, logging and policies in one place — turns GDPR requirements into reusable infrastructure instead of per-feature exceptions. At Svapna we integrate AI into enterprise systems — ERP, PLM, ecommerce, middleware — and we treat data protection as a design constraint from the first sprint, with application hardening and monitoring built in. If you are planning AI features for your B2B platform, start from the data flow.