If you use Microsoft 365, you’ve probably already seen a message about Copilot doing more than answering questions. From October 2026, Microsoft is rolling out changes that let Copilot Chat carry out multi-step work directly: drafting emails, managing meetings, organising files in OneDrive and SharePoint, and creating documents, without you doing each step yourself.
That’s a meaningful shift, and it’s arriving inside Microsoft 365, a platform most mid-size Australian businesses already run on. It matters now because the decision to switch this on, or leave it on by default, isn’t really a technology decision. It’s a question of what you’re prepared to let software do inside your business before a human checks it, and it’s the practical version of a question more leaders are now asking: is my business ready for AI agents?
An AI agent readiness checklist for business covers four things: what the agent can already access, who approves its actions, what staff are told before it goes live, and how you’ll know if something went wrong. Get those four settled before enabling agent features, not after.
What Changed in Microsoft 365, and Why Now?
Microsoft’s own Message Center notice MC1479277 (published 25 September 2026, updated 12 October 2026) confirms that Copilot Chat is gaining the ability to complete long-running, multi-step work across Word, PowerPoint, Excel, Teams, OneDrive, and SharePoint, including creating files and managing meetings. Microsoft states that sensitive actions require user approval, and a related notice, MC1493223 (published 9 October 2026), describes “Skills” in Copilot Chat: task-specific instructions that staff can create and share, and that admins can deploy or block, rolling out from October to mid-November 2026.
The multi-step Chat features are rolling out first to organisations in Microsoft’s Frontier early-access program, and they need a Microsoft 365 Copilot licence. Some of the larger agent capabilities also use usage-based billing that an admin has to switch on. So whether and when this reaches your staff depends on licensing and admin settings, which is exactly why it’s worth deciding deliberately rather than finding out after the fact.
What Is an AI Agent, and How Is It Different From a Chatbot?
An AI agent is software that completes a task on your behalf, rather than just telling you what to do. A chatbot answers a question and leaves the action to you. An agent takes the action itself, often chaining several steps together: reading a request, checking a calendar, drafting a reply, and sending it, all in one pass.
The practical difference is where the human check happens. With a chatbot, you review the answer before you act on it. With an agent, the action has often already happened by the time you see it, unless an approval step was built in deliberately. That’s the entire basis for the checklist in this article: approval has to be designed in, because it’s no longer automatic.
What Can Go Wrong When an Agent Acts on Its Own?
Most AI agent risks for business come down to three failure modes that show up consistently once agent-style tools are switched on without preparation.
Over-broad access. An agent inherits whatever permissions the person using it already has. If someone has access to a folder, mailbox, or system they technically shouldn’t, the agent can now act on that access too, at machine speed, across far more items than a person would manually touch in a day.
No clear owner of the approval step. Microsoft’s “sensitive actions require approval” is a real control, but your business still decides what counts as sensitive, and who the approver is. Left unconfigured, a default setting decides this for you, and it may not match how your business actually wants decisions made.
Staff who don’t know what the tool is doing. An employee who doesn’t understand that Copilot can now draft and send on their behalf, not just suggest, is more likely to let an action through without a second look. Confusion about capability is itself a risk factor, separate from whatever the agent actually does.
None of these are reasons to avoid agent-style tools. They’re reasons to check your own setup before the default configuration makes the decision for you.
What Should You Check Before Turning On Agent Features?
Work through these in order. Each one is quick to check and expensive to skip.
- List what’s connected. Identify every mailbox, shared drive, and system the relevant Microsoft 365 accounts can currently reach. An agent can only act within that same access, so this list is also your risk map.
- Decide what counts as a sensitive action. Sending an external email, sharing a file outside the organisation, and deleting anything are reasonable starting points. Write the list down rather than relying on Microsoft’s default definition.
- Name the approver for each category. “Sensitive actions require approval” only works if a specific person, not a vague “the user,” is responsible for giving it.
- Check who has broader access than their role needs. This is the moment to tidy up permissions that have drifted over time, before an agent can act on all of them at once.
- Decide which roles get agent features first. A staged rollout to one team, with a defined review period, surfaces problems while they’re still small.
This is the same groundwork an AI Readiness Audit covers for any AI initiative: knowing where you stand across people, process, data, and technology before you spend time or trust on the tool itself. Agent features just make the data and technology questions more urgent, because the tool can now act on what it finds.
Who Should Decide Whether to Turn This On?
This is a question of AI agent governance, not a line item in an IT rollout plan. It isn’t a decision for IT alone, and it isn’t a decision for individual staff to make by default either.
IT and your Microsoft 365 administrator own the technical configuration: what Copilot can access, what counts as a sensitive action in the tenant settings, and who the approver is. Operations leadership owns the practical question of which workflows an agent should touch first, and which ones are too high-stakes to automate yet. Whoever owns risk and compliance in your organisation, often the CFO or COO in a mid-size business without a dedicated risk function, should sign off on the policy before it goes live company-wide.
Leaving this entirely to IT treats it as a technical rollout. It isn’t. The technical settings just implement a decision about risk tolerance that the business, not the IT team, should actually be making.
How Do You Train Staff to Use AI Agents Safely?
A policy that sits in a shared drive and gets mentioned once in an email doesn’t change behaviour. Staff need to know three concrete things before agent features reach their accounts: what the tool can now do without being asked twice, what they’re expected to check before approving an action, and who to tell if something looks wrong.
This is where training earns its place alongside the technical setup. Akoora’s Staff Training & AI Adoption work exists specifically because tools bought without training on how to use them well tend to sit half-used, or used in ways nobody intended. The same pattern that causes low adoption, people not understanding what a tool actually does, is exactly what causes agent-related mistakes. Fixing the training gap addresses both.
How Do You Get Started Without Guessing?
Start with the access list and the sensitive-action list from the checklist above. Both can be done internally in an afternoon for a single team, and they tell you immediately how exposed your current setup is.
If that review surfaces more questions than answers, or if nobody in the business is confident they know what Copilot can currently access across your tenant, that’s the signal to bring in a structured review rather than guessing your way through it. It’s the same gap that leads to most AI projects failing: moving straight to the tool without first mapping what it touches.
Business process automation projects run into a related version of this problem. As covered in our guide to business process automation, automating a process you haven’t mapped properly just automates the mistakes in it, faster. Letting an agent loose on unmapped access is the same risk, applied to permissions instead of workflows.
Frequently Asked Questions
What is an AI agent?
An AI agent is software that can carry out a multi-step task on your behalf, not just answer a question. Where a chatbot gives you an answer to act on yourself, an agent drafts the email, updates the file, or books the meeting directly, often across several systems in one go.
Is Microsoft Copilot’s new agent feature safe to turn on?
It can be, but safety depends on your setup, not the feature itself. Microsoft requires user approval for sensitive actions, but who counts as the approving user, what counts as sensitive, and what Copilot can already access in your tenant are all things your business configures, not Microsoft.
Do we need a written policy before staff use AI agents at work?
Yes. Without a written policy, staff will each decide for themselves what counts as an acceptable action for an agent to take, which means your actual risk exposure depends on whoever is least cautious. A short, practical policy closes that gap.
How is AI agent risk different from regular AI chatbot risk?
A chatbot’s worst case is a wrong or misleading answer that a human still has to act on. An AI agent’s worst case is the wrong action already taken, in your inbox, your files, or your calendar, before anyone reviews it. The review step moves from before the mistake to after it.
How long does it take to get ready for AI agents at a mid-size business?
A focused review of what’s connected, who can approve what, and which roles need training typically takes two to three weeks for a mid-size organisation. That timeline matches a standard AI Readiness Audit because the questions are the same ones the audit already asks.
The Takeaway
AI agents inside tools like Microsoft 365 Copilot change where the human check happens, from before an action to after it, unless you deliberately design the approval step back in. Checking what’s connected, deciding what counts as sensitive, naming an approver, and training staff before rollout turns a default configuration decision into one your business actually made on purpose.
Not sure what Copilot can currently access across your business, or who’s actually approving what? An AI Readiness Audit answers exactly that, before you spend a dollar on rolling AI agents out further. Book a conversation →
AI assisted with the research, structure, and editing of this article. The views and recommendations are Akoora’s own.