Autonomous AI agents: decide where they must stop
Facts checked
Autonomous AI agents use information to decide what to do and work toward a goal without a person approving every step, Microsoft says. Before you assign a task, choose what a person must check and which actions need your approval. Write those limits beside the task, then ask how each one will be met before you use the work.
What can AI agents do?
Microsoft says autonomous AI agents take in data, choose steps and act on a goal without human sign-off for each step. It says to put clear AI choices and checks first. It also says to review AI choices often and use human checks where needed.
For your task, ask what information the agent would read, what it would decide and what it would change. Write each part down. Don't let a broad goal such as "handle customer questions" stand in for those choices.
Choose a small piece of work to start with. Use the one-page task inventory to pick a task you can describe and check.
Ask where the task should end. Should the agent prepare a draft for you, or are you considering letting it act on that draft? If you want it to act, name the exact action and ask how you'll check that it stays within your instructions.
Set a rule for checking factual claims before the work is used. Use the draft claim check to compare a draft with the records behind it.
How do autonomous AI agents complete tasks?
For a supply-order trial, provide the stock list, ask for a draft order and have a person check it before anyone buys. Ask the person setting up the task to show that it stops at the draft.
To try the supply-order task:
- Choose the input. Supply your stock list with the items marked for reorder and the supplier notes you want used.
- Define the decision. Ask for an order draft using only the marked items. Ask for a list of gaps if the notes don't settle what to order.
- Limit the action. Request an order draft for review, and keep placing the order as a task for a person.
- Assign the check. Name who will compare the draft with the stock list and check the supplier.
- Approve the use. Read the checked draft yourself, then decide whether the person should place the order.
Example
Use this instruction as a starting point for the draft task:
Draft a supply order using only the items marked for reorder on this stock list. Use the supplier notes provided. List any missing quantities or unclear item names for me to check. Leave the order as a draft. Ask me before adding an item or placing an order.
Before a trial, ask the person setting up the task to show what happens when an item has no quantity. Also ask them to show what happens if a request calls for placing an order. Keep the trial out of live purchasing until you've checked how those cases are handled.
BreadHead says AI systems do much of the work, and people direct it, check it and approve it.
Book a discovery call with BreadHead to discuss your task and whether BreadHead fits. Booking signs you up for nothing.
What risks do autonomous AI agents introduce?
For the supply-order trial, check for unneeded items, the wrong supplier and an order you haven't approved. Hold the order until you've resolved each issue.
Keep the same task in view. Ask what should happen if the stock notes conflict, if the draft adds an item or if someone asks the agent to place the order. Write down who should take over in each case.
Use this check:
- Which stock list and supplier notes may the agent use?
- Who will check a detail that the notes don't settle?
- Does the task stop at a draft, or include a change outside it?
- Which actions need my approval before they happen?
- Who will take over when the request goes beyond the task?
- What must I see before I approve the work for use?
Fill in a line for each action you're considering:
| Action | Allowed source | Where to stop | Person who checks | Approval needed before use |
|---|---|---|---|---|
| ______ | ______ | ______ | ______ | ______ |
Example
For the supply order, write down where the task must stop:
Draft only: refer any missing quantity to the person who checks stock. I must approve the items, quantities and supplier before a person places the order.
Ask whoever sets up the task to explain how they'll meet that boundary. Have them show you a draft with a missing detail, then compare what happened with the stop you asked for. If you can't confirm the limit, hold the task and decide what needs to change.
Check those choices again when you add a new source or action. Ask whether the same person can still check the work and whether your earlier approval covers the added use.
With BreadHead, you review and approve the work before it goes public or into use.
Set your limits before choosing a platform. See the best AI workflow automation tools when you're ready to compare tools.
Bring the task and its stop points to a discovery call with BreadHead to talk about your work and whether BreadHead fits. You can book without signing up for anything.
Sources
Microsoft: Introduction to Autonomous AI Agents
BreadHead's own pages: Home, AI Service Disclaimer, Discovery call booking page, AI Honesty Test watch page, Calendar.
Statements from BreadHead's founder.
This page states no outside numbers.
What changed
- : First version. Facts checked.