We build the Model Context Protocol layer that lets AI agents work inside your SAP, Salesforce and core banking systems, hosted in the EU.GDPR and EU AI Act designed in, not retrofitted.
European buyers are not slower to adopt agents. They answer to more people first. A DPO who needs a lawful basis and a record of processing. A works council that wants to know what is being monitored. And since 2 August 2026, an AI Act regime whose high-risk obligations are fully enforceable.
That changes what the integration layer has to be. It is where personal data is minimised, where transfers are contained, and where a human stays in the loop by design. Built as an afterthought, it becomes the thing that fails the audit.
Full high-risk obligations under the EU AI Act became enforceable. Logging, human oversight and transparency now carry consequences.
Regulation (EU) 2024/1689
Maximum AI Act penalty exposure, whichever of the fixed sum or share of worldwide turnover is higher.
EU AI Act penalty provisions
EU hosting moved from preferred to mandatory in 2026 procurement for regulated sectors and public buyers.
2026 EU MCP procurement guidance
Member states, each with its own supervisory authority. One architecture has to satisfy all of them.
Modernisation without migration, and without moving a single record outside the jurisdiction it belongs in.
The layer that makes your systems readable by agents is where minimisation, logging and oversight are enforced.
Your systems
MCP layer, EU-hosted
Minimised field sets, pseudonymisation at the boundary, Article 12 logging and human oversight on every consequential action.
Your agents
SAP S/4HANA and ECC, Dynamics 365, Salesforce, Sage, DATEV, core banking platforms and Snowflake.
Each tool returns the narrowest field set that answers the question. Personal data is withheld at the boundary, not filtered afterwards.
Purpose, data categories, recipients and retention logged as structured events that feed your Article 30 register.
Article 14 oversight is not a policy sentence. Sensitive actions stop at an approval step with full context attached.
Egress is explicit. Any non-EEA component is named, assessed and covered by SCCs before the build, never discovered in a log.
Built to MCP rather than one vendor, so Mistral, EU-hosted Bedrock or a sovereign self-hosted model all remain open to you.
Each one changes something concrete in the build. We tell you which article drove which decision.
Purpose limitation expressed as tool scopes. The default is that an agent cannot see a field, and every exception is documented.
Records generated from live telemetry, plus encryption, access control and resilience testing evidenced against Article 32.
A DPA on your terms, a named sub-processor list, EU-region processing by default, and SCCs with a transfer impact assessment.
We help classify the agent system, then implement the Article 12 logging and Article 14 oversight that classification triggers.
Register-of-information entries, documented exit strategies, incident reporting hooks and resilience testing built into delivery.
Supply-chain security, incident handling and management accountability reflected in how the layer is operated, not just built.
We are engineers, not your legal advisers. We build the architecture that makes a compliant position achievable and hand your DPO and counsel the evidence to assess it. The determination stays theirs.
Residency that depends on somebody remembering a setting is not residency. The topology is the control.
Deployed into your own AWS, Azure or GCP account in Frankfurt, Dublin, Paris or Stockholm, with your keys and identity provider.
For public sector and core banking: self-hosted on your infrastructure, or on OVHcloud, Scaleway or IONOS.
Mistral, EU-region Bedrock, Azure OpenAI EU data zones or Vertex EU, with zero retention verified in configuration.
Cases where the answer already exists in a system and a person is paid to go and retrieve it.
Transaction history, KYC records and prior alerts assembled per case, with every access logged for the regulator.
Policy, claim and correspondence systems read together to surface what needs a human today, with no new copy of personal data.
SAP, MES and supplier portals queried in one call to find the shortage that will stop a line next week.
Batch records, deviations and study documentation retrieved under strict scoping, with an audit trail that fits GxP expectations.
Citizen casework supported by agents running entirely within national infrastructure, with full transparency on what was consulted.
One layer serving several subsidiaries, each keeping its own controller relationship, scopes and retention rules.
The European sequence differs in one way that matters. The data protection work happens before the engineering.
Systems in scope, data categories, lawful basis per purpose, transfer exposure and AI Act classification. Output feeds your DPIA.
Tool schema, minimised field sets, role mapping, retention and oversight checkpoints, reviewed with your DPO and works council.
Built and validated inside EU infrastructure, with logging and evidence generation switched on from the first test.
Agent platform connected, then handover with runbooks, the Article 30 feed, the control mapping and DPIA material.
Opening production systems to an agent is a data protection decision first. Every boundary is agreed in writing before the first tool is exposed.
The contractual details that decide whether a programme starts this quarter or next year.
Daily overlap with CET and GMT hours for standups, incident response and reviews.
We sign your Article 28 agreement, disclose sub-processors up front and accept audit rights.
Mapping produces the data flow, necessity and proportionality material your DPO can lift directly.
Plain-language descriptions of what is monitored and what is not, for Betriebsrat consultation.
One architecture across subsidiaries, documented in English plus your operating language.
Code, infrastructure definitions and documentation transfer at the end of every phase.
Start narrow, prove it inside the boundary, then widen. Every phase ends with something running and something documented.
| Engagement | What it covers | Typical shape |
|---|---|---|
| Start HereMAPPING & DPIA SUPPORT | Fixed-scope assessment of systems, processing purposes, transfer exposure and AI Act classification, ending with a costed build plan. | Typically 1 to 2 weeks. Fixed price. |
| Targeted BuildSINGLE-SYSTEM LAYER | One system, one minimised set of tools, running in EU infrastructure before the scope or budget widens. | Typically 5 to 8 weeks end to end. |
| Full DeliveryGROUP-WIDE INTEGRATION | Several systems and entities under one governed layer, with per-entity scoping and a single evidence surface. | Phased, production milestone per phase. |
| OngoingMANAGED EVOLUTION | Continued engineering plus periodic control review as AI Act guidance, your systems and your workflows keep moving. | Monthly retainer, scaled to surface area. |
Not by default. The layer is deployed inside your EU tenancy or on European infrastructure, and the model endpoint is pinned to an EU region with zero retention verified in configuration. If a component you want has no EU option, we name it before the build, assess the transfer and cover it with SCCs. You never find out from a log file.
Where we handle personal data during delivery we act as your processor and sign your Article 28 agreement, including sub-processor disclosure and audit rights. In the steady state the layer runs in your own environment, so for most clients we are not processing personal data at all once it is live.
Most conversations start with a system map. Share the platforms you depend on, the data categories involved and what you want an agent to do. We will come back with a scoped plan, an EU-resident architecture and an honest read on what your DPO will ask.
Ready to Get Started?
Transforming businesses with data, AI/ML, and cloud solutions