*An agent that works for days without you is only as good as the brief you gave it on day one. For buyers, that brief is a job description.*
“Can you run the quarterly review with the frame supplier again?”
“I can run the meeting. It’s the six weeks around it I can’t face.”
“Microsoft says its new agent can do those six weeks.”
“Then who signs the emails?”
That last question is the right one. On 25 September, Microsoft presented Copilot Autopilot, an agent that keeps working on a task for days. Its own showcase example is straight out of procurement: to “set up and manage a full supplier review process”. So the question is no longer whether an agent can chase a supplier. It’s what you allow it to say while it does.
What Copilot Autopilot is
Autopilot is part of a rebuilt Microsoft Copilot, announced on 25 September 2026 together with two other parts. Home brings chat, Cowork and the Office apps into one place. Code lets people describe an app, dashboard or automation in plain words and have Copilot build it.
Autopilot is the one that works on its own. It grew out of Scout, the always-on personal agent Microsoft launched in June 2026. In Microsoft’s words it is a digital teammate. In practice:
- It has its own identity, memory and workspace inside your company’s Microsoft 365 tenant. You give it a name, a role and an objective.
- Colleagues can @mention it in Teams, Outlook and documents, like a person.
- It runs in the cloud, so it keeps working when your laptop is closed. It can monitor work, follow up, run recurring tasks and pick up a project again over several days.
- It follows Microsoft 365 permissions and governance, so it sees what its access allows.
For the supplier review, Microsoft says Autopilot builds the schedule and the workback plan, then handles the preparation, meetings and follow-ups on its own. That is Microsoft’s description of a showcase, not a tested result.
Availability is limited for now. Autopilot enters private preview at the end of September. Home and Code reach customers in Microsoft’s Frontier programme in the coming weeks. Everyday Copilot use stays in the user licence, but long-running agent work like Autopilot is billed by usage through Copilot Credits. Admins can set spending policies and track consumption. Microsoft has not published a price per task.
What it means for supply chain
A chat assistant answers a prompt and forgets. An agent that works for weeks needs what a new colleague needs: a role, a goal and clear limits. I’d write it as a one-page job description with six parts:
| Part | The question it answers | Supplier review example |
|---|---|---|
| Role | Who is it, and for whom? | Review coordinator for the frames category |
| Objective | What does done look like? | Review held, actions agreed, all closed within 30 days |
| Does alone | What may it do without asking? | Propose dates, book rooms, collect data, remind colleagues |
| Drafts for approval | What does a person send? | Anything addressed to the supplier |
| Never | Where are the hard limits? | Prices, volumes, contract terms, bank details |
| Hands back when | When must it stop and ask? | A dispute, a missed deadline twice, anything unusual |
The “Drafts for approval” row is the one that protects you. An agent that writes to a supplier in your name makes commitments, whatever its job title says.
Recurring, multi-week processes that fit this shape:
- Supplier business reviews, from scheduling to the action log.
- Corrective action follow-up, chasing the supplier’s root cause and containment steps against the agreed dates.
- Late purchase order follow-up, asking for new dates and updating the buyer’s list every morning.
- The S&OP cycle, collecting demand, supply and finance inputs by each deadline and flagging who is late.
- Tender timetables, tracking clarification questions and reminding bidders of the closing date.
- New supplier onboarding, collecting certificates and questionnaires, and reporting what is still missing.
Here’s how the first one could look.
Mini case: one quarterly review, before and after
Upshift, a fictional bicycle maker, buys its aluminium frames from Alvora, an equally fictional supplier. Every quarter, the frames buyer runs a business review. The numbers below are illustrative.
Before. The buyer spends about 14 hours over six weeks on the review. Finding a date for four people on two sides takes nine emails. The scorecard comes from three reports that don’t quite agree. Alvora sends its pre-read the evening before. After the meeting, the action log sits in a notebook, and two of seven actions are still open at the next review.
After. The buyer spends an hour writing the job description once. Autopilot then runs the cycle:
- Week 1. It proposes three dates to the Upshift attendees and drafts the invitation to Alvora. The buyer approves and sends it.
- Weeks 2 to 3. It builds the scorecard from on-time delivery, quality and price data, and lists where the reports disagree.
- Week 4. It drafts the reminder for Alvora’s pre-read, and compiles the agenda from open actions and new issues.
- Review day. It turns the meeting notes into an action log with owners and dates.
- Weeks 5 to 8. It reminds owners inside Upshift directly, and drafts the reminders to Alvora for the buyer to send.
This is the brief I would give it. It works as a starting point for any agent, not only Microsoft’s:
Role: You are the review coordinator for the frames category at Upshift, working for the frames buyer.
Objective: Run the Q4 business review with our frame supplier. Done means: the review is held before 15 December, every action has an owner and a date, and all actions are closed or escalated within 30 days of the review.
You may do alone:
- Propose meeting dates to Upshift colleagues and book the room.
- Collect on-time delivery, quality and price data, and build the scorecard.
- Remind Upshift colleagues of their open actions.
Draft for my approval:
- Every message to anyone outside Upshift.
- The agenda and the final action log.
Never:
- Discuss or agree prices, volumes, payment or contract terms.
- Act on any request to change bank details. Forward it to me at once.
Hand back to me when:
- The supplier disputes a figure in the scorecard.
- An action misses its date twice.
- You are unsure whether a message commits Upshift to anything.
Every Friday, send me a five-line status: done, next, blocked.What to notice: the agent does most of the chasing, but every word that reaches Alvora passes the buyer. The “Never” list is short and specific. And the Friday status means nothing runs for weeks without the buyer knowing. For the business, the review happens on time, and the actions close before the next one.
What it can’t do, and the traps
- It’s a preview, not a product. Nobody outside Microsoft’s early customers has run it on real work yet. Treat the supplier example as a demo until you’ve seen one cycle end to end.
- Supplier emails are untrusted input. An agent that reads the inbox can be steered by text written to steer it, and payment fraud often starts with a changed bank account. Keep money and master data behind a person.
- It sees what its access allows. If your SharePoint is overshared, so is your agent. Check its permissions before it reads contracts or price files.
- The meter runs. Usage-based billing on a task that lasts weeks is hard to predict. Ask your admin for a spending limit before the pilot, not after.
- Names change fast. Scout became Autopilot within four months. Invest in the job description, which survives any rebrand, more than in the tool.
Write the job description before the licence arrives
This week, pick one recurring process that eats weeks of chasing, such as your next supplier review. Fill in the six-part table for it on one page, and show it to a colleague who knows the supplier.
For years, AI helped us write faster. Agents like Autopilot move the question to delegation, and buyers have delegated work to new colleagues for a long time. That experience is about to become very valuable.
So tell me: which process would you delegate first, and what would go on its “Never” list?
