AI Agents

Agents That Keep Working After You Log Off: What to Hand Over First

Microsoft's Autopilot makes the always-on agent mainstream. A two-question test for which jobs to give it

By STREaMLiNE

Hand over the work you can check in a minute and undo in a click. Keep the rest.

On 25 September Microsoft rebuilt its Copilot app around three things, and the headline one is Autopilot: an agent you "give a name, a role and a goal" and which then keeps working inside Microsoft 365 after you have signed off. It has its own identity, memory and workspace, with permissions and an audit trail behind it, and colleagues can mention it in a chat like a teammate. It is moving to private preview now, with the agent features billed by usage.

Whatever you think of the product, the shape is here to stay. The question for a small business is no longer whether an agent can do a task. It is which tasks to hand over first, and on what terms.

How does an always-on agent actually work?

Strip away the branding and every long-running agent follows the same loop. A person sets the goal and the limits. The agent works through your tools while you do something else. It stops at a checkpoint when a decision is needed, records what it did, and a person reviews the record.

Long-running agents · The loop behind every one

The work loop of a long-running agent, from goal to weekly reviewAn owner sets a goal and limits. The agent, with its own identity, works through tools such as email, files and calendar while the owner is away. At a checkpoint it stops to ask before sending or changing anything that matters. Every action is written to a log. The owner reviews the log each week and adjusts the goal and limits, which starts the loop again. OWNER sets goal, limits AGENT own identity TOOLS email, files CHECKPOINT asks before sends LOG every action REVIEW owner, weekly works while you are away; stops when a decision is needed

The loop is only as safe as its weakest link, and in practice that link is the review. An agent nobody reviews is not an assistant. It is an unsupervised employee with your passwords.

Which jobs should you hand over first?

Ask two questions about any task. How easily can I check the result? And how easily can I undo it if it is wrong? The answers sort your work into four groups.

Delegation · Two questions sort every task

Which tasks to hand to an agent, by how easy they are to check and to undoA two by two grid. Easy to check and easy to undo: hand over first, for example sorting and labelling mail, weekly report drafts, meeting notes and actions. Easy to check but hard to undo: the agent drafts and a person sends, for example client emails, invoice reminders and quotes. Hard to check but easy to undo: spot-check weekly, for example categorising transactions, tidying CRM records and tagging documents. Hard to check and hard to undo: keep it human, for example client advice, lodgements and payments, and hiring decisions. DRAFT, YOU SEND client emailsinvoice remindersquotes HAND OVER FIRST sorting and labelling mailweekly report draftsmeeting notes and actions KEEP IT HUMAN advice to a clientlodgements and paymentshiring decisions SPOT-CHECK WEEKLY categorising transactionstidying CRM recordstagging documents harder to undo easier to undo easy to check hard to check

Hand over first: easy to check, easy to undo. Sorting and labelling email, drafting the weekly summary, turning meeting notes into actions. If the agent gets one wrong, you notice in seconds and fix it in one click.

Draft, you send: easy to check, hard to undo. Client emails, invoice reminders, quotes. Once sent, they are sent. Let the agent prepare everything and a person press the button.

Spot-check weekly: hard to check, easy to undo. Categorising hundreds of transactions or tidying CRM records. You cannot read every one, so you sample: twenty a week, and a count of how many you corrected.

Keep it human: hard to check, hard to undo. Advice to a client, lodgements, payments, hiring decisions. An agent can gather the material. A person decides.

A useful rule of thumb: if checking the agent's work takes longer than doing it yourself, the task is in the wrong box.

What should be in place before an agent runs overnight?

Microsoft's framing is right on one point: an agent needs an identity, so that everything it does is recorded against it rather than against a person's login. Beyond that, three things matter for a small business.

  1. A named owner. One person who set the goal, can change it, and reads the log. Not "the team".
  2. Limits written down. Which systems it can use, what it may send without asking (ideally nothing, at first), and what it must stop and ask about.
  3. A budget. Usage-billed agents cost more the longer they run. Set a monthly cap before the first run, not after the first invoice.

Start with one job from the "hand over first" box. Run it for a month. Then decide, with the log in front of you, whether it earns a second.

Sources: Microsoft, Introducing the new Copilot with Home, Code and Autopilot, VentureBeat.

Key takeaways

  • Two questions sort every taskHow easily can you check the result, and how easily can you undo it? Start where both answers are easy.
  • Drafts before sendsFor anything that leaves the business, let the agent prepare and a person press send.
  • Always-on needs an ownerAn agent that keeps working needs a named person who set its goal and reads its log each week.

Frequently asked

Can small businesses use Autopilot today?

Not generally. Microsoft said on 25 September that Autopilot is expanding to private preview at the end of the month, and its agent features run on usage-based billing on top of the Microsoft 365 Copilot licence. The two-question test works for any agent, whichever product you end up using.

What is the difference between a chatbot and a long-running agent?

A chatbot answers when you ask and stops. A long-running agent is given a goal and keeps working on it across hours or days, using your email, files and calendar, and checking back when it needs a decision. The second is far more useful and needs far more thought about permissions.

What is a good first job for an agent in an accounting or bookkeeping practice?

Work that is easy to check and easy to undo: sorting and labelling incoming client email, drafting the weekly work-in-progress summary, or turning meeting notes into an action list. Client advice, lodgements and payments stay with a person.

How do I know if the agent is doing a good job?

Decide before you start. Pick one measure, such as time saved per week or the share of drafts sent without edits, and read the agent's log weekly for a month. If you cannot tell whether it is helping, it is not ready for more responsibility.

AI Agents

An AI Agent Got Into Medicare's Statistics Portal. Here Is How to Fence In Yours

An OpenAI agent routed around blocks into a Medicare statistics portal. What happened, and five controls that fence in your own agents.

Read

Automations

AI Agents Are Moving to Pay-Per-Task. Cost One Before You Switch It On

Microsoft now bills agent work by usage. How to cost an agent task before it runs, and why review time usually outweighs usage.

Read