Microsoft Foundry September 2026 Actions Every CIO Should Take

In this blog post Microsoft Foundry September 2026 Actions Every CIO Should Take we will explain which September announcements matter, what risks they introduce, and where technology leaders should act first.

The problem for most CIOs is no longer a lack of AI options. It is the growing number of models, agents, tools, pricing choices and preview features competing for attention. Without clear controls, an impressive pilot can quickly become an expensive and difficult-to-govern production system.

What Microsoft Foundry does in plain English

Microsoft Foundry is the layer organisations can use to build, deploy and manage business AI applications. It brings AI models, automated agents, business data connections, security controls and performance monitoring into one managed Azure environment.

An AI agent is essentially a software worker that can interpret a request, use approved tools and complete several steps. For example, an agent could review support tickets, check customer records, suggest a response and assign the ticket to the right team.

The value of Foundry is not simply access to models such as OpenAI’s GPT family or Anthropic’s Claude. It is the ability to manage different models and agents using common identity, networking, policy and monitoring controls. We covered that foundation in more detail in our guide to using Microsoft Foundry as a unified control plane for AI governance.

1. Stop choosing one model for every task

September expanded the production-ready model choices in Microsoft Foundry. The GPT-6 family now includes Astra for demanding work, Sol for a balance of capability and cost, and Luna for focused, high-volume tasks. Claude Opus 5.5 also became available for long-running coding and knowledge work.

The CIO-level takeaway is simple: the most powerful model is not automatically the best business choice. Using a premium model to classify routine emails or summarise standard documents can produce an unnecessarily large bill without improving the outcome.

Ask your team to benchmark models against real business tasks. Measure accuracy, completion rate, response time and total cost per completed task—not just the price of each AI request.

This extends the model evaluation approach discussed in our article on comparing Claude across Microsoft Foundry and Anthropic. The decision should cover operating controls, data handling and support arrangements as well as model quality.

2. Test Model Router before model selection becomes manual work

Microsoft’s Model Router can automatically direct a request to an appropriate model. In plain English, it can send straightforward work to a cheaper model and reserve more capable models for difficult requests.

September added preview controls for keeping related conversation steps on the same model where possible. It also added more information showing which model was selected, whether another model was attempted, and why a fallback occurred.

That visibility matters. Automatic routing can reduce costs, but CIOs should not approve it as a black box. Teams need to confirm that routing decisions do not reduce answer quality, move workloads into an unsuitable data location or create unpredictable monthly spending.

Start with one high-volume, low-risk process such as internal document summarisation. Compare fixed-model and routed results for four weeks before expanding the approach.

3. Treat Routines as operational automation, not another chatbot

Routines became generally available in September. Generally available means Microsoft considers the capability suitable and supported for production use, unlike a preview feature that may still change.

Routines allow an agent to run on a schedule or respond to an approved event. An agent could prepare a daily operations summary, review a backlog every Friday, respond to a Microsoft Teams event or perform a recurring compliance check.

This removes some of the infrastructure normally required to schedule work, manage identities, retry failed jobs and record execution history. The business outcome is faster automation with less custom software to maintain.

However, scheduled AI actions need stricter controls than a chatbot that waits for a user. Define what each routine can read, what it can change, how much it can spend and which actions require human approval.

A useful rule is to let an agent prepare, classify and recommend before allowing it to approve, delete, pay or publish. Human approval should remain mandatory for high-impact decisions.

4. Use Tool Search to reduce cost and permission sprawl

Agents become useful when they can access tools such as a customer database, document library, ticketing system or finance application. The problem is that giving an agent too many tools increases complexity, cost and security exposure.

Tool Search, now generally available, allows an agent to find the relevant approved tool when it needs it rather than loading every available tool into every request. This can reduce the amount of information processed and make large agent environments easier to manage.

Microsoft also made agent-to-agent communication generally available. This allows specialised agents to pass work to each other—for example, a service agent could ask a billing agent to check an invoice without receiving broad access to the entire finance system.

Before adopting either capability, create a tool register. Record the owner, purpose, permissions, data accessed and approval level for every tool an agent can use.

This is where existing security practices remain critical. Microsoft Entra manages identities and access, Microsoft Defender detects threats across the Microsoft environment, and Wiz identifies cloud risks and dangerous configuration paths. Foundry does not replace these controls, and AI governance does not replace Essential 8, the Australian Government’s baseline framework for reducing common cyber risks.

5. Make monitoring part of the business case

Foundry Insights entered public preview in September. It uses traces—a step-by-step record of what an agent did—to help teams investigate poor performance and identify possible improvements.

This addresses a common executive concern: “How do we know the AI is actually working?” A successful response is not enough. You need to know whether the agent completed the business task, used the correct information, followed policy and cost less than the manual process.

Because Insights is still in preview, use it for evaluation rather than making it the only control protecting a critical workflow. Establish a monthly dashboard covering task completion, human corrections, failures, response time, cost per task and any inappropriate data access.

6. Revisit cost and Australian data requirements

September also brought Foundry pricing changes and a new Asia-Pacific Data Zone option. This may give Australian organisations more deployment flexibility, but “Asia-Pacific” does not necessarily mean data remains only in Australia.

Before moving a regulated or sensitive workload, confirm where data can be processed, which model provider receives it, how prompts and outputs are retained, and whether the arrangement meets your privacy, contractual and industry obligations.

Flex processing also offers a lower-cost option for work that does not require an immediate response. It may suit overnight document processing, large evaluation jobs or scheduled summaries, but it is not appropriate for customer interactions where response time matters.

These choices should feed into a wider cloud cost review. Our earlier article on why cloud cost is back on every CIO’s agenda explains why provider estimates are only the beginning.

A practical mid-market scenario

Consider a 200-person professional services business receiving hundreds of client documents each week. Its first AI pilot uses one premium model for every document and relies on staff to trigger each review manually.

A better Foundry design could use a Routine to start processing, Tool Search to access only the approved document tools, a lower-cost model for standard classification and a stronger model for unusual cases. Staff would approve low-confidence or sensitive results before information reached a client system.

The likely business outcomes are lower processing costs, faster turnaround and a clearer audit trail. More importantly, the company can expand the workflow without giving one agent unrestricted access to every system.

What CIOs should do in the next 30 days

1. List every AI model, agent and business tool currently being tested or used.
2. Select one workload and compare at least two models using cost per successful task.
3. Identify one low-risk scheduled process suitable for a controlled Routine pilot.
4. Create approval rules for any agent capable of changing data or triggering actions.
5. Confirm data locations, retention settings and provider terms for sensitive workloads.
6. Review upcoming model retirement dates and document a migration owner for each production deployment.

The priority is controlled progress

September’s Microsoft Foundry updates make agents more capable, automated and easier to operate. They also increase the need for disciplined model selection, access controls, cost monitoring and human oversight.

CloudPro Inc brings more than 20 years of enterprise IT experience across Azure, Microsoft 365, OpenAI, Claude, Defender and Wiz. As a Melbourne-based Microsoft Partner and Wiz Security Integrator, we help organisations turn new AI capabilities into practical, governed services rather than disconnected experiments.

If you are not sure whether your current Foundry plan is ready for production—or whether it will cost more than expected—we are happy to review the architecture and identify the first actions worth taking, with no strings attached.