What is an agent?
An agent in this context is software that acts for you, not just gives you answers. For example, it might move files, send emails or update records without you doing each step.
← The Human Bit Weekly
4 October 2026 · 6 min read
If you wouldn't paste it into an email without checking, don't let an agent act on it unchecked.
The 60-second version
Whenever an AI suggests or acts, name where the suggestion came from and check its reason before sharing or automating.
- What changed
- Worth watchingChatGPT’s ad test has expanded across Europe. Check paid placements separately.
- Worth watchingMicrosoft describes controls for supported agents and connected tenants.
- Worth watchingSnowflake describes permissions and audit trails for gateway traffic.
- What to do
- Label any AI-originated or agent-enabled decision or suggestion before it goes into email, tickets, or your handover.
- What stays yours
- You say out loud (in a comment, a forwarding note or a shared doc) which suggestion or action is AI-sourced, and whether you trust its basis.
New to this?First time with AI?
What does “Sponsored” mean?
“Sponsored” identifies a paid placement. OpenAI says ChatGPT ads are labelled and displayed below answers, and that advertisers cannot influence those answers. Check any advertised option against your own criteria before including it in a recommendation.
What is an audit trail?
An audit trail is a record of what was done, by whom, and when. It helps you work out who changed something, important if you let an agent or tool act automatically.
Start here
Three things that changed
01Worth watchingNewly announced
ChatGPT’s ad test has expanded across Europe. Check paid placements separately.
OpenAI has expanded its ad test across 31 European markets. It says ads are separate from ChatGPT’s answers. If you include an advertised option in a shortlist, identify it as paid and check its claims against your criteria.
OpenAI: European ad expansion02Worth watchingNewly announced
Microsoft describes controls for supported agents and connected tenants.
Microsoft describes a shared view of agents across supported platforms and connected tenants. Its controls depend on enabled integrations, platform APIs and administrator permissions. Check which agents your organisation has connected and who can review or stop them.
Microsoft: Agent 365 update03Worth watchingNewly announced
Snowflake describes permissions and audit trails for gateway traffic.
Snowflake describes permission controls and action records for agent traffic passing through Cortex AI Gateway, with the listed controls in private preview. Your team still has to decide what agents may do and check that the relevant traffic is covered.
Snowflake: agent governance
The one that mattersWhenever an AI suggests or acts, name where the suggestion came from and check its reason before sharing or automating.
You should treat AI anyone-can-use tools and agents as new sources of actions and suggestions, not just helpers. AI-generated recommendations, paid placements and agent actions each need their own checks. This matters most for any list of options or decision you might forward, use, or automate. Before sending a client email or letting an agent update your spreadsheet, identify the source, check the evidence, and confirm what you have authorised. Your main job is to pause and say: 'where did this come from, and does it deserve to move forward.'
The full take: what is new, the makers' claim, where it fits, what is still unknown
What's genuinely new
These updates bring different checks into the same workflow: identify any paid placement you choose to use, verify the evidence behind an answer, and confirm an agent’s permissions and recorded actions. A dashboard or gateway is useful only when you understand what it covers.
What the makers claim
Microsoft describes administrator controls for supported agents and connected tenants. Snowflake describes central permissions and audit trails, with Cortex AI Gateway controls in private preview. OpenAI says sponsored placements are separate from answers and offers ad explanations and controls. We have not independently tested how clear the labels are or how completely these agent controls cover a real workflow.
Where it helps
You’re collecting options or drafting next steps and ask ChatGPT, then plan to copy what it gives into an email, ticket, or plan, or you use an agent to act on your behalf in a workflow.
Where it doesn't
You’re just brainstorming privately without copying suggestions anywhere, or you only use AI tools with no outside data or actions.
OpenAI has expanded its ad test across 31 European markets. It says ads are separate from ChatGPT’s answers.
Before sending a client email or letting an agent update your spreadsheet, identify the source, check the evidence, and confirm what you have authorised.
- How clear ad labelling is to a tired user copying suggestions into real documents.
- Whether agent audit gates capture everything when agents are spun up by other tools or scripts.
Try this
One thing to try this week
Before using an AI recommendation or agent action, check its source, evidence and permissions.
AI drafts answers, options, or actions for your real work, turning a meeting into draft next steps, or summarising a budget.
You decide what gets shared, assigned or acted on after checking the evidence, identifying any paid placement you included, and reviewing the agent’s permissions and actions.
- 1
Set criteria for a decision using information you are permitted to share, then ask for a shortlist. Verify each candidate against sources. If you also choose a candidate from a visible sponsored placement, label it as paid.
- 2
If you use an agent to automate part of your job (posting updates, sending emails, moving files), review the activity log or summary to confirm what was done, when, and by what type of agent.
- 3
Before sharing, note which items came from AI, which came from any paid placement, and why you kept them. For agent work, record what changed and how you checked it.
Human checkpoint
Label any AI-originated or agent-enabled decision or suggestion before it goes into email, tickets, or your handover.
The Human Lens · what stays yours
Your reputation is built on why you trusted a suggestion, not just what you picked.
A machine can propose, sort and even act, sometimes better than a person. Still, it’s only you who’ll be asked why a particular recommendation or change happened, especially if the client or team disagrees. Take a minute to state the source, check for bias or sponsorship, and decide if you’d stand by that action in a meeting. This is the part that makes your work yours.
One question to carry into your week
In your next shared doc or client note, which option or decision will you flag as AI-sourced, and what led you to trust it?
See you next Monday.