Choosing a capable AI model is useful. Before giving it a task in your business, you also need to decide what it may access, what it may do and when a person takes over. A recent Microsoft announcement brings this question into focus.
What Microsoft announced
On 28 October 2026, Microsoft announced that its Copilot harness was being made available through Microsoft Foundry. This is the software framework around an AI model: it organises instructions, context, tools and task execution. Microsoft describes both an SDK and managed cloud execution. Read the official announcement.
The announcement is primarily aimed at teams building agents. It does not mean that a new assistant is automatically enabled in every Microsoft 365 subscription. Before starting a project, you still need to check feature availability, terms of use and suitability for your environment.
What does an agent actually do?
An agent can work through several steps using the tools it is allowed to access, rather than simply answering a question. Consider a fictional small-business example: gathering information from a working folder, preparing a summary and suggesting a draft reply. Sending that reply to the customer can remain subject to a colleague’s approval.
The distinction matters. Preparing a document and sending it do not have the same consequences. Neither do reading a folder and changing its contents. The right scope therefore depends on the work required, not just the name of the model being used.
Four decisions before you start
- Define a specific task. Start with a limited, repetitive activity whose output is easy to check. “Prepare a summary from these documents” is clearer than “manage our files”.
- Limit access. Provide only the data and tools needed for that task. A trial does not require opening up the entire mailbox or every shared folder. Have permissions checked by the person responsible for your IT.
- Set approval requirements. Identify actions that should still require approval: external messages, changes to records or commercial commitments. Also decide who steps in when information is missing or contradictory.
- Measure the actual outcome. Compare preparation time, review time and observed errors. A fast response is not necessarily a benefit if correcting it takes longer than doing the original task.
Test imperfect situations too
A useful pilot goes beyond a successful demonstration. Try incomplete documents, an ambiguous request or a temporarily unavailable tool. Check that the process stops cleanly and explains what is missing. Keep a manual way of working and appoint someone to oversee the process.
For an SME, an association or an independent professional, the aim is not to accumulate agents. It is to choose a use case that genuinely helps the team, with understandable rules and an appropriate level of autonomy. A limited pilot lets you make an evidence-based decision before expanding its use.
Start with your everyday work
Do you have a task in mind but are unsure whether AI is suitable? MYCIO can help you examine your use cases, tools and team training needs. Discover our AI support or contact us to discuss it.
Conceptual illustration generated with AI.