Skip to main content
Industries

How AUMOVIO centralized AI access with Databricks Unity Gateway

by Louis Rottmeier, Maximilian Fest, Michael Boesl, Michael Weber, Martin Gindler and Kelly Albano

As AI adoption grew, AUMOVIO teams developed different ways to access large language models. Today, Databricks Unity Gateway provides a shared entry point for our AI workloads. This blog discusses how we brought these approaches together.

The challenge: Managing AI access across multiple providers

Teams at AUMOVIO developed different ways to access LLMs, including provider integrations and gateways built for specific use cases. These approaches all addressed real needs, but as AI adoption grew, their variety added complexity to operations, cost attribution and consistent governance.

The main opportunities to improve were:

  • Shared visibility. Bring model usage into one place for consistent reporting.
  • Cost transparency. Attribute spend to a user, team or use case.
  • Simplified operations. Manage provider updates, incidents and support through a shared service.
  • Consistent governance. Apply model approvals, logging and guardrails through a common approach.

Together, these approaches highlighted the value of a shared operating model for managing our AI workloads with a single AI gateway solution

The decision: One centralized AI gateway

AUMOVIO's cross-functional AI taskforce brought together stakeholders from security, data protection, business areas and IT. The taskforce decided on one central AI Gateway as the shared route for LLM access and governance, establishing a common direction for model access across the company.

From there, our job was to make the central path useful to teams and their applications.

Why we chose Databricks Unity Gateway

Building an AI gateway is straightforward, but operating one as an enterprise service is a permanent commitment: provider APIs change, dependencies need patching, security requirements evolve, and users need support. Once applications depend on it, a gateway needs a roadmap, an on-call rotation and a support process. Weighing that against buying yet another AI gateway, five arguments decided why Databricks Unity Gateway was the solution for us.

1. Databricks was already our data platform. 1ADP, the One AUMOVIO Data Platform, runs on Databricks. Putting model access next to our governed enterprise data meant no new vendor, no new network path to review and no isolated stack to operate.

2. Corporate identity integration. Databricks is integrated with our corporate Microsoft Entra ID tenant, and other AUMOVIO applications authenticate against that same tenant. An application presents its Entra ID token and gets access to Unity Gateway through token exchange, with no separate account, no API key to distribute and no bespoke integration. Onboarding a new AI-enabled application takes a permission grant, and access to every model ends automatically when the corporate identity does.

3. Governance and usage tracking. Model approvals, access policies, usage tracking and logging sit in one place instead of being reimplemented per application. Payload logging through inference tables is off by default and only enabled with an explicit, approved justification, which is the posture our data protection agreements require.

4. We stayed free to choose providers. This mattered most. We publish a curated catalog of 1adp-* model aliases, and Unity Gateway routes each alias to whichever provider serves it: Databricks-hosted foundation models, Google Vertex AI, Azure OpenAI, AWS Bedrock or others. Users code against a stable, vendor-neutral name, and we can re-point that name to a different provider or a newer model without a single consumer changing a line of code. The aliases give teams a consistent route to approved models.

5. The setup is infrastructure as code. Terraform expresses everything Unity Gateway does for us: the endpoints behind each 1adp-* alias, the provider each model routes to, the fallback order when a provider degrades, rate limits and the guardrails applied per endpoint. Adding a model, re-pointing an alias or tightening a safety control is a pull request, reviewed, versioned and reproducible across environments. When production misbehaves, our version control explains the configuration that caused it. Because we provision the rest of 1ADP the same way, the Gateway is one more module for the engineers who already run the platform.

How Unity Gateway helps us with AI cost management and model access  

Centralizing model access was the first step. We then built three capabilities on top of Unity Gateway to support how our teams manage, access, and use AI:

Cost attribution and chargeback. Unity Gateway reports on how many tokens were consumed, and every user assigns their consumption to a personal cost center or to one of the projects they work in. Pipelines join that assignment with Databricks system tables and charge usage to the right cost center. We publish the result as dashboards: which models are used, by whom, how consumption develops and what it costs. Budget owners can see their spend and identify changes in consumption.

A curated model catalog. Standardized access does not make models interchangeable. The catalog is the single source of truth for which models are approved, what they map to and which consumption tier they run in. Teams still test their application against the model they choose, but they no longer build infrastructure to reach it.

An interface for employees. The 1ADP AI App runs in production on the same gateway and gives employees a governed way to use AI without writing code. Developers get an OpenAI-compatible endpoint and can point compatible agentic coding tools at the central models. Same infrastructure, two audiences: we standardize model access and leave the choice of development interface open.

The results so far: Centralized AI governance, Cost visibility, and developer access

Unity Gateway provides a shared route for LLM access. The 1adp-* aliases connect applications to models, guardrails apply centrally, and usage is attributed to cost centers in near real time. 

The next phase moves up the AI platform stack: sharing agents inside 1ADP projects and connecting knowledge sources, web search and other tools through managed and bring-your-own MCP (Model Context Protocol) connections. Model access stays centralized; what changes is the range of tools an agent can reach. That raises questions about what an agent may retrieve, which tools it may call and whose permissions it acts under. Having one framework to answer them in is why we consolidated in the first place.

Lessons learned: Building vs. buying an enterprise AI Gateway

If you are bringing different model access approaches together, involve security, data protection, business areas and IT in the operating model. Then make the shared service useful to the teams expected to use it.

When you compare building to buying, price the decade rather than the prototype. Ask who patches the provider integration, who answers the support ticket, who handles the incident, and who does it again next year.

Then decide what is genuinely yours to build. The gateway was not. Cost attribution, the model catalog and the employee-facing AI App were, because they encode how AUMOVIO works and no vendor was going to hand them to us.

We remain accountable for our architecture, our security decisions and the quality of our applications. Unity Gateway provides the shared model access service, while teams can continue to choose the applications they build.

Frequently Asked Questions

Why did AUMOVIO choose Databricks Unity Gateway instead of building its own AI gateway?

AUMOVIO chose Databricks Unity Gateway to centralize AI model access without the ongoing maintenance, security updates, and operational overhead of building its own gateway. Because Databricks already powered its enterprise data platform, AUMOVIO could reuse existing infrastructure, Microsoft Entra ID integration, and governance controls. This allowed its teams to focus on building company-specific capabilities such as cost attribution, a curated model catalog, and an employee-facing AI application.

How can enterprises centrally govern access to AI models across multiple providers?

Databricks Unity Gateway provides a single, governed entry point for accessing AI models across providers such as Azure OpenAI, Google Vertex AI, AWS Bedrock, and Databricks-hosted models. AUMOVIO uses centralized model approvals, access policies, logging, and guardrails to apply consistent governance across applications. Its curated model aliases also let administrators change the underlying model or provider without requiring application code changes.

How does Databricks Unity Gateway help enterprises track and allocate AI costs?

Unity Gateway provides token usage data that enterprises can use to track AI consumption across models and users. AUMOVIO combines this data from Databricks system tables with project and cost center assignments to attribute AI spending to the appropriate business owners. Dashboards provide near-real-time visibility into model usage, consumption trends, and costs, supporting internal chargeback and budget oversight.

How can enterprises give developers access to multiple AI models without managing separate API keys?

Databricks Unity Gateway provides centralized model access through an OpenAI-compatible endpoint, reducing the need for developers to manage separate provider integrations. At AUMOVIO, applications authenticate through Microsoft Entra ID token exchange, eliminating the need to distribute individual API keys. Developers can use approved models through compatible coding agents and applications while IT manages permissions and provider connections centrally.

What are the benefits of using a centralized AI gateway for enterprise AI applications and agents?

A centralized AI gateway simplifies access to multiple models while bringing governance, cost visibility, and operational management into one shared service. AUMOVIO uses Unity Gateway to support both employee-facing AI applications and developer tools through the same infrastructure. Its next phase will extend this approach to shared AI agents and MCP connections, with a focus on governing tool access and permissions.

Related tags:

Unity Gateway and Manufacturing

Get the latest posts in your inbox

Subscribe to our blog and get the latest posts delivered to your inbox.