Infrastructure teams are being pushed to put AI agents to work. But before you give an agent permission to provision services, change configurations, or make other infrastructure decisions, thereʼs a more fundamental question to address: Can it trust your data?

Data that works for human-driven automation may not work for autonomous agents. Engineers can spot missing information, investigate inconsistencies, and work around systems they know are unreliable. Agents need much clearer rules. Here are five signs your infrastructure data isnʼt ready for them yet.

Five signs your infrastructure data isnʼt ready for AI

1. Schema changes break your automation. One global entertainment company had Ansible playbooks pulling from its source of truth. Whenever the database schema changed, its automation templates broke. Agents need explicit, predictable data structures they can rely on.

2. Youʼre not working with your source of truth. Instead of working with ServiceNow directly, the network automation team at a global financial technology company built a custom UI, Git workflows, and Python scripts around it. The workaround made them faster but the resulting toolchain became so complex that only senior developers fully understood it. When your team works around your source of truth instead of with it, every agent you add to the workflow inherits that complexity.

3. Your source of truth doesnʼt contain the whole truth. One large telco had infrastructure data spread across its source of truth, network inventory, and commercial inventory because its schema couldnʼt represent everything the team needed. An agent can only reason from the information available to it. Missing infrastructure and relationships create blind spots and can lead to dangerous hallucinations

4. You canʼt model intended state before deployment. Another team relied on spreadsheets for Day Zero configuration across 200 global PoPs. That may be workable with enough human coordination but an agent needs a structured record of what should exist before it can safely help build it.

5. Your infrastructure data lacks business context. Ownership, contracts, service dependencies, and the reasons behind infrastructure decisions often live outside the source of truth. Without that context, an agent may understand what a device is without understanding what depends on it or why it matters.

These problems have existed for years. AI raises the stakes because agents remove some of the human interpretation that has made imperfect data systems workable.

What AI-ready infrastructure data looks like

Giving agents access to more data isnʼt enough. They need infrastructure data that is structured for machines to understand and safe for them to act on.

Start with a flexible, explicit schema. Your data model needs to represent your actual infrastructure, including its quirks, without forcing everything into predefined objects. But flexibility canʼt mean ambiguity. Fields, types, relationships, and constraints should be explicit so an agent can understand what data exists and what it means, and create valid data when needed.

Connect infrastructure to its context. Devices, circuits, services, customers, locations, owners, and contracts are connected records. A knowledge graph makes those relationships explicit and queryable, giving an agent the context to understand dependencies.

Version your intended state. Agents need a safe way to propose changes. Treating infrastructure data like code lets you version, test, and review changes before they become part of the production state. With branching, an agent can work on proposed changes separately while humans remain in control of what gets approved.

Give agents the right data, not all the data. We learned this firsthand while experimenting with AI at OpsMill. When we gave an LLM the full Infrahub GraphQL schema, we overwhelmed its context window before it could do useful work.

The better approach is to give agents task-specific slices of infrastructure data. An agent working on a circuit doesnʼt necessarily need your entire infrastructure model. Give it the data and relationships required for the task, with paths to retrieve more when needed. Smaller contexts reduce ambiguity and make agent behavior more predictable.

Build the data foundation before you hand over the controls

Infrahub provides these building blocks in a data platform designed for infrastructure automation. Its flexible schema lets you model infrastructure and business context around your requirements, while its knowledge graph captures the relationships between them. Git-style version control and branching let agents propose changes while keeping testing and human review in the workflow.

GraphQL and Infrahubʼs MCP server can give agents access to the infrastructure data they need for a specific task rather than exposing the entire data model at once. As you move toward agentic infrastructure automation, giving AI access to your existing source of truth isnʼt enough. You need a data foundation designed to provide the structure, context, and controls agents need to act reliably.