Until now, most AI at work has been something a person opens, asks a question of, and closes. The tool acts only when someone is in front of it, and it works through that person’s account.
That’s starting to change. Microsoft says its new Autopilot agent is “an always-on, cloud-based personal agent” with “its own account, mailbox, and files”. In the same update, published on 30 October 2026 and revised on 3 November, Microsoft says Autopilot “is rolling out in November to Frontier”, the programme through which Microsoft releases new Copilot features to some customers first. Most New Zealand businesses won’t see it this month. But the direction is clear, and the management questions it raises apply to any AI agent that can act without someone watching.
What changed, in plain English
Three things landed close together.
Agents with an identity of their own. An Autopilot agent isn’t borrowing your login. It has its own account and mailbox, can work with your local files and “its own remote computer”, and can work with teammates where that’s switched on. Microsoft says access controls and approval guardrails keep “consequential actions under human control”.
A way to see which agents are in use. Microsoft’s Copilot release notes for 4 November 2026 describe a new agents usage report in the Microsoft 365 admin center, under Copilot management. It covers agents used by both licensed and unlicensed staff, and shows who built each one: a user, your organisation, Microsoft or a Microsoft partner. Usage is broken down per user and per agent.
Agents managed like any other identity. Microsoft’s Agent 365 lets organisations give agents “a first-class identity” in its identity system and “manage agents in much the same way you would any identity in your tenant”. That includes sign-in rules, risk detection, an inventory of agents, and a named person responsible for each agent “throughout the agent lifecycle”. Microsoft’s own documentation says Agent 365 needs its own licence per user, either as an add-on to some Microsoft 365 plans or bundled into its top tier.
Put together, an agent is becoming less like a feature and more like a member of the team: it has an account, it holds access, it does work while you’re not looking, and someone needs to answer for it.
Why it matters for how teams work
Work gets handed off, not just assisted. With a chat assistant, a person stays in the loop for every step. An always-on agent can be given a job, such as chasing overdue supplier confirmations or preparing a weekly summary, and left to get on with it. That frees time, but it also moves the risk. A mistake might not be noticed until someone reads the output, or until a customer replies.
Access becomes the main control. If an agent has its own mailbox and files, what it can reach is set by the permissions you give it, not by the person who asked. Most small and mid-sized businesses already struggle to keep staff access tidy. Agents add more accounts to keep track of, and they don’t leave the business in an obvious way.
Agents multiply quietly. Staff can already build their own agents in many tools. The new usage report exists because organisations need to find out what’s running. If you don’t know how many agents your team uses today, that’s the first gap to close.
How NZ businesses could use it
The useful jobs are ones that are routine, well defined and easy to check:
- Distribution and wholesale: watch a shared inbox for order confirmations and delivery changes, update the tracking sheet, and flag anything that affects a promised date.
- Agriculture: pull compliance records, contractor timesheets and supplier invoices into one folder each week, so the end-of-month pack is ready to review.
- Construction: keep a running list of open RFIs and variations from project email, and draft a weekly update for the project manager to approve before it goes to the client.
- Professional services: prepare a first draft of the month-end client report from the files and emails already on record, for a staff member to check and send.
- Manufacturing: summarise quality notes and supplier emails into a daily exceptions list for the operations manager.
In every case the agent prepares and a person decides. None of these involve the agent paying, committing the business, or sending anything final to a customer on its own.
Risks, costs and governance
Privacy. An agent with its own mailbox will see personal information in customer and staff emails. Under the Privacy Act 2020 you remain responsible for how that information is used, stored and kept accurate, and for keeping it secure. Limit the agent to the folders and mailboxes the job needs, and check where its data is processed. This is general information, not legal advice.
Security. An agent account is a target, just like a staff account. Its access should be the minimum needed, it should be covered by the same sign-in rules as staff, and someone should review what it has done.
Accuracy. Agents can misread an email or act on an out-of-date file. Build checks into the job: a person approves anything external, and the agent reports what it did, not just what it produced.
Cost. Microsoft’s governance features for agents sit behind an extra licence, and some agent features are billed on usage. Before rolling anything out, ask your IT provider what you’d pay per user and per month at the volume you expect.
Human sign-off. Decide in advance which actions an agent may never take alone: payments, changes to bank details, contract terms, staff matters and anything sent to a regulator.
What managers should do differently
Treat each agent like a new hire. It needs a job description, a manager, the access that role needs and no more, and a review after its first month. When the job ends, the agent’s access ends too.
Make one person accountable for each agent. Microsoft’s tools now support a named owner. Use that, and make sure the owner is someone who understands the work, not only IT.
Measure the work, not the activity. Ask how many hours the agent saves, how often its output needs fixing, and whether anything went wrong. If the answer isn’t clear after a month, stop it.
Try this in the next 30 days
- Find out what’s already running. Ask your IT provider for the agents usage report, or a list of agents and AI apps connected to your Microsoft 365 or Google tenant.
- Name an owner for every agent that’s in active use. Retire the ones nobody claims.
- Write a one-line rule for actions agents can never take without a person approving them.
- Pick one routine job that takes a staff member two or more hours a week and is easy to check.
- Scope the access that job needs: which mailbox, which folders, read or write.
- Run it for two weeks with a person reviewing every output, and record the time saved and the corrections made.
- Check the cost, including licences for governance features, before scaling up.
- Update your AI use policy to cover agents that act on their own, not just staff using AI tools.
Where to start
If you already use Microsoft 365, start by finding out which agents your team has built and who owns them. Our AI and agents work covers choosing the right first job, setting access and approvals, and measuring whether it pays off. If nobody in your business owns decisions about systems and access, a fractional CTO can. Or get in touch to talk it through.