This is more than an infrastructure agreement
Microsoft and Databricks announced on 23 July that they are extending their strategic partnership into the 2030s. Databricks will run more of its own core operations on Azure Databricks and use Microsoft's Cobalt infrastructure. Microsoft, in turn, will integrate Databricks capabilities such as Genie and Unity AI Gateway more deeply across its products.
The infrastructure commitments are substantial, but the language around them is more revealing. Both companies emphasise AI that understands customers, products, operations, metrics and business processes. Their promise is not simply access to a stronger model. It is AI grounded in trusted enterprise knowledge, governed consistently and delivered inside the tools where people already work.
That points at the real bottleneck for enterprise AI: not a shortage of models, but a shortage of usable business context.
Business context is not the same as giving an agent more data
Most organisations already have plenty of data. The problem is that meaning is scattered across databases, documents, dashboards, SharePoint sites, inboxes and the heads of experienced employees. Two systems may use the same customer name but disagree about status. A KPI can have different definitions in finance and operations. A procedure may be documented, while the exceptions that determine the actual work live in email threads.
An agent cannot solve that by reading more rows. It needs to know which source is authoritative, how terms are defined, which information is current, who may see it and what action is permitted. It also needs the methods around the data: business rules, approval thresholds, exception paths and the evidence required before a decision can move forward.
That is why business context should be treated as part of the operating architecture. Without it, an agent can produce fluent answers while acting on the wrong definition, an outdated document or incomplete permissions.
Putting AI inside Teams does not automatically put it inside the process
The expanded partnership connects Databricks with services across the Microsoft ecosystem, including Entra, Purview, OneLake, Power BI, Microsoft 365, Teams and Copilot. That makes it easier to bring governed data and AI closer to daily work. But proximity to the user is not the same as integration with the workflow.
A useful operations agent must do more than appear in Teams. It must recognise the case being discussed, retrieve the correct records, apply the organisation's rules, preserve access controls, prepare or execute the next system action and escalate exceptions to the right person. It must also leave an audit trail that explains what information and rules shaped the result.
The channel is only the front door. The value sits behind it: a controlled path from data to method, tool, decision and action.
The context layer becomes the foundation for an AI-native operation
For companies moving beyond isolated assistants, the architecture can be viewed as an operating layer with six connected parts: data, methods, tools, governance, channels and agents. Data provides the facts. Methods capture how work should be done. Tools connect the agent to operational systems. Governance sets permissions and boundaries. Channels bring the capability to employees and customers. Agents coordinate the work across those layers.
The Microsoft-Databricks announcement reinforces this direction. Genie and governed enterprise data can help make business knowledge accessible; identity, security and governance services can constrain its use; Microsoft channels can place it near the user. The hard implementation work is connecting those parts around a concrete process without losing meaning or control between systems.
This is also why platform choice alone does not create an AI-native company. The platform supplies components. The organisation still has to identify its authoritative context, encode its operating methods and decide where humans remain accountable.
A practical way to start
Choose one process with a visible bottleneck and a clear owner. Map the sources the process relies on, the definitions people use, the decisions made, the systems touched, the permissions required and the exceptions that trigger human review. Then test whether an agent can prepare one useful next step with evidence, rather than asking it to automate the entire process at once.
Measure operational outcomes: turnaround time, error rate, consistency, manual handovers and the number of cases that can be completed without rework. That creates a better business case than measuring prompts, licences or token consumption.
The lesson from Microsoft and Databricks is straightforward. Enterprise AI becomes useful when it understands how the business actually works. Models matter, but business context is what turns their capability into reliable operational value.