AI agents inside your own infrastructure, not someone else's.
AI agents that work on your internal knowledge and systems, inside your own cloud region or data centre. Your data never leaves.
The model is ready. The question is where it runs.
Most companies have piloted a public AI assistant. Far fewer run agents on real customer, contract or engineering data. The blocker is rarely the model: it is where data goes, which law applies, and whether security, data protection and the works council sign off.
Open-weight models now handle most internal work and run on a handful of GPUs. The cloud tenant or data centre you already run can host everything an agent needs. Running agents inside your perimeter is now an engineering task, not a research project.
The hard part is everything around the model: connecting SAP and SharePoint, enforcing the permissions people already have, testing answers before staff see them, and keeping it running after the project team leaves. We do that work, then hand it to your teams.
- 96%of German companies would prefer AI from German providers, Bitkom 2026Bitkom, Digital-Gipfel: Unternehmen wollen KI aus Deutschland (November 2026)
- 87%say a provider's country of origin matters when choosing AI, Bitkom 2026Bitkom, Digital-Gipfel: Unternehmen wollen KI aus Deutschland (November 2026)
What we build
Model serving and routing
Open-weight models such as Mistral, Llama or Qwen on your GPUs or cloud region. Swap models without rebuilding the agents.
Retrieval over internal knowledge
Search across SharePoint, Confluence and file shares. Answers cite the source and respect each document's existing access rights.
Tool connections
Agents read and write in SAP, Salesforce or your own APIs, with set limits and human approval where you choose.
Identity and permissions
Sign-in through Entra ID or Okta. The agent acts with each user's own rights, never a shared super-user account.
Evaluations
Tests from your real cases run before every release. If quality falls below the agreed bar, the release stops.
Monitoring and operations
Prompts, tool calls, latency and cost traced in self-hosted tools, with runbooks and dashboards so your platform team runs it.
How it works
Every step runs on infrastructure you control, from first request to last log line.
A request arrives
An employee asks in Teams or a new ticket arrives. Sign-in passes the user's identity along.
Retrieve and plan
It searches only what this user may see, plans the steps and picks a model inside your environment.
Act through approved tools
It reads an order in SAP or opens a ticket. Actions above your threshold wait for a person.
Answer with sources
The user gets an answer or finished task, linked to its sources, so they can check it.
Trace and improve
Every step is traced in your environment. Evaluations and feedback show where quality drops, and your team adjusts.
How we roll it out
Discovery
3 weeksWe map your infrastructure, data and security needs, pick two or three use cases, and agree where agents run.
- Target architecture for your environment
- Prioritised use cases with success measures
- Security and data protection concept
Build
8 to 12 weeksWorking in your tenant, we build the first agents and test them on real cases before users see them.
- Agents live for a pilot group
- Evaluation suite and quality baseline
- Infrastructure as code in your repositories
Run and hand over
8 weeksWe run the agents with your platform team, widen access and tune them, then your team takes over.
- Runbooks and operating documentation
- Monitoring dashboards and alerting
- Trained platform and support team
In production elsewhere
Public examples from other organisations, not unus clients, showing in-house and sovereign AI already in production.
Public case studyFinanz Informatik (Sparkassen-Finanzgruppe) Savings bank assistant on open models, own data centres 170,000 workstations
- The situation
- Finanz Informatik runs the savings banks group's IT from its own German data centres. Advisers handle sensitive customer data daily. The group wanted generative AI without depending on global cloud providers or losing control of data protection and security.
- What they deployed
- Its AI platform runs open-source models from Mistral, OpenAI and others in its data centres, with no cloud connection. S-KIPilot answers from group knowledge inside Outlook. An adviser agent gathers account data, summarises recent contacts and drafts a meeting guide.
- The outcome
- According to the article, the S-KIPilot has been rolled out to 170,000 workstations in the Sparkassen, with more than three million professional prompts each month. The adviser preparation agent ran about 130,000 meeting preparations in March.
Open-weight models in your own data centre can carry an assistant and its agents at the scale of an entire banking group.
Source: Börsen-Zeitung, Keine Zukunftsperspektive ohne Vertrauen und Sicherheit (guest article by Andreas Schelling), May 2026Public case studyCanton of Aargau Canton runs generative AI in its own data centre Runs in the cantonal data centre
- The situation
- Aargau already used machine learning. Cantonal data protection law requires personal data to be processed under a public body's control, so it wanted generative AI on its existing infrastructure, with more than one model and no sensitive data given away.
- What they deployed
- The canton runs Swisscom's Swiss AI Platform entirely in the cantonal data centre, orchestrated with Open WebUI. Each department picks the model it needs, including foreign pretrained ones, as long as the canton orchestrates them. AargauGPT is the internal assistant.
- The outcome
- The canton says it keeps control over infrastructure and access logs, can switch models on its own, saves licence costs and stays independent of individual cloud providers. AargauGPT runs in IT and a few pilot areas so far, with a canton-wide rollout to be supported by training and practical guides.
Keeping orchestration, indexes and access logs in-house lets you change models freely while meeting strict data protection law.
Source: Swisscom B2B Magazin, Digitale Souveränität statt Abhängigkeit: Wie der Kanton Aargau die Swiss AI Platform einsetzt, July 2025Public case studyITZBund (Federal Information Technology Centre) Federal AI platform run in government data centres Cleared for VS-NfD classified data
- The situation
- Germany's federal administration handles personal and classified information under the strictest data protection, secrecy and security rules. Agencies wanted generative AI for everyday work, on their own knowledge, without tying themselves to one model vendor.
- What they deployed
- ITZBund built KIPITZ, a shared platform where each agency configures its own portal. Apps cover chat, document summaries, translation, redaction and transcription. Agency knowledge and specialist applications connect through retrieval and APIs. Mostly open-source models run in ITZBund's data centres.
- The outcome
- ITZBund reports that many federal agencies already use KIPITZ, with all portals operated centrally in secure federal data centres. The platform is planned to become part of the sovereign Deutschland-Stack, scheduled for the end of 2028.
A modular platform with swappable, mostly open-source models lets one central team serve many units without locking them into one vendor.
Source: ITZBund, Die KI-Plattform der Bundesverwaltung: KIPITZ, October 2026These are public deployments by other organisations, summarised from the sources linked. They are not unus clients.
Built for European rules
Data stays in your perimeter
Models, indexes and logs stay in your EU or Swiss tenant or data centre. Nothing reaches us or others unless you approve.
GDPR and Swiss nDSG
We document data flows and retention, support your data protection impact assessment, and enforce deletion and access rules in the index too.
EU AI Act duties
Agents talking to people disclose they are AI (Article 50). Annex III uses like recruitment face high-risk rules from 2 December 2027.
Security and human oversight
Least-privilege access, network isolation, full audit trails. Writes above your thresholds need approval; any agent can be switched off without redeploying.
Questions buyers ask
Do we need our own GPUs?
No. Many organisations start with models hosted in their existing Azure, AWS or Google Cloud tenant in a European region, or on a sovereign cloud. Own GPUs make sense when volumes are high, the data is very sensitive or you want full control over the model. In discovery we size both options against your use cases and costs.
Are open-weight models good enough for real work?
For most internal tasks, such as search over documents, drafting, summarising and classification, current open-weight models perform well once retrieval and evaluations are set up properly. For harder reasoning, specific tasks can be routed to a larger model hosted inside your tenant. Evaluations on your own cases decide which model handles what, not public benchmarks.
Can an agent see data a user is not allowed to see?
No. The agent acts with the signed-in user's rights. Retrieval filters documents by the permissions they already carry in SharePoint, Confluence or your document management system, and tool calls use scoped credentials. Before go-live we test this explicitly with users in different roles.
Who operates it after go-live?
Your team. We build with your platform and security people from the first week, everything is infrastructure as code in your repositories, and we hand over runbooks, dashboards and documentation. We can stay involved for new agents and model changes, but the setup does not depend on us.
What happens when a better model comes out?
Models sit behind a routing layer, so they can be swapped. When a new model is available, we run it against your evaluation set and compare quality, latency and cost. It goes live only if it passes, and the previous model stays available for rollback.
Does this work with SAP and our other core systems?
Yes. We connect agents to SAP, Salesforce, ServiceNow, Microsoft 365 and in-house systems through their APIs or MCP servers, with read and write scopes defined per tool. The effort depends on each system and its interfaces, which we assess during discovery.
Observability
We trace every prompt, answer and euro your AI spends. Quality, cost and compliance on one screen, kept in the EU.
Explore →Also from unusWorkflow automation
Agents process invoices, prepare the close, draft campaigns and keep your CRM current, inside your systems. Your team approves what matters.
Explore →