Skip to content

THE FIELD JOURNAL

Before an AI Agent Reads Your Inbox: Six Checks to Make First

By Local AI Scout · Beginner guide

Illustration of a laptop showing an inbox with an envelope, beside a clipboard with six ticked items and a padlock

Local AI Scout · Cloud / Hybrid, depending on the connected workflow · Guide · Docs-only · Updated November 10, 2026

Short answer: before you connect an agent to an inbox, customer database or booking calendar, write down what it can read, what it can change, who approves actions and how you disconnect it. Start with a narrow, non-sensitive example. Running the model on your own computer does not automatically keep a connected workflow local or remove your responsibilities for customer information.

This guide follows one reader: a small salon owner who wants help drafting appointment replies, but does not want the tool sending messages, changing bookings or reading every customer record on its own.

Why this is worth checking now

The UK's Information Commissioner's Office has a call-for-evidence page about agentic AI. The ICO's call page lists an October 8, 2026 start and says the call stays open until the end of November 20, 2026‡. That is context, not a new law or a reason to connect a tool before you understand it.

The checklist below is our practical starting point, not the ICO's official questions and not legal advice. If you handle sensitive records or cannot explain what a supplier will do with customer information, pause and get appropriate professional advice.

1. Define the job in one sentence

For our salon owner: “Draft a reply from the appointment details I provide; I will check and send it.” That is narrower than “manage my bookings.”

List what is outside the job too: no refunds, no marketing messages, no appointment cancellations and no access to payment or health notes. A clear boundary gives you something concrete to compare with the permission screen and supplier documentation.

2. Ask what it can read, not just what you plan to ask

Can the tool read one selected message, a folder, the whole inbox or attachments as well? Can it see calendar titles, contact details and past appointments? Does a CRM connection include notes you did not mean to share?

Read the actual permissions and ask the supplier where the data is processed, retained or passed onward. If the connection offers only broad access, do not assume a prompt saying “ignore the other messages” narrows that access. Choose a smaller integration or use a manually prepared example instead.

3. Separate reading from acting

A draft is different from a sent email. A suggested slot is different from changing a calendar. Find out whether the tool can send, publish, change or delete, and whether you can require approval before each external action.

For the first task, keep the owner as the sender. Review the recipient, dates, prices and promises in every draft. If the product cannot separate drafting from sending, decide whether it fits this job before giving it live access.

4. Name the person who owns the workflow

Write one person's name next to the trial: who checks replies, who notices a wrong booking and who can switch it off? Keep a simple action record appropriate to the task, without copying unnecessary personal information into a second system.

Do not treat “the supplier runs the AI” as a complete answer about accountability. Ask who handles a complaint or a data incident and where you can get human support. The legal allocation of responsibility depends on the facts and relevant law; this checklist does not determine it.

5. Keep the wrong things out

Do not connect everything just because a connector is available. Keep payment credentials, identity documents, staff records and sensitive customer notes outside a basic reply-writing trial. Use made-up appointment details to explore the workflow; do not copy a real customer's information just to make the example realistic.

Local processing can be useful for a private drafting task, but you still need suitable access controls and a sensible retention policy. A local model calling remote tools can be Hybrid. See when not to use local AI and what local apps may still send online before choosing the label for your setup.

6. Know the off switch before the first connection

Find the documented route to revoke the connection, remove tokens or keys where relevant, and stop future access. Work out who can do it if the owner is away. Disconnecting future access may not delete copies already retained by a service; check deletion and retention terms separately.

Do not invent a menu path from someone else's screenshot. Follow the current documentation for your product and account, and ask the supplier if it is unclear.

Start with a sample, not the whole business

Illustration of blank sample cards and a rubber stamp with a checkmark beside a locked cabinet and disconnected plug
A smaller trial and a human checkpoint. Illustrative artwork, not a product screenshot.

Our salon owner's first example can be entirely fictional: “A customer asks whether a Tuesday afternoon appointment is available. Draft a reply asking which service they need. Do not send anything or change the calendar.” This is an untested example, not evidence that a particular tool will obey it.

Compare the result with the boundary you wrote down. Does the tool invent availability, a price or a promise? Can you tell what it accessed? Does it wait for approval? You should be able to answer those questions before considering a limited live connection.

If you also manage a public business listing, read the Gemini and Business Profile account checks. A public profile and a private customer inbox need different access decisions.

Choose a limited trial if…

  • The job is narrow and the permissions match it.
  • You can keep sensitive information out and review every action.
  • A named person can inspect, stop and disconnect the workflow.

Wait if…

  • You cannot tell what data the connection exposes or where it goes.
  • The tool sends or changes things without the approval you need.
  • You do not have an appropriate basis or permission to share the information.

Related: Is it safe to upload documents to ChatGPT? · ChatGPT GIFs for small business · See AI agent visits with Microsoft Clarity · All local-business guides

Frequently asked questions

Is read-only access harmless?

No. Reading can expose personal or confidential information even without editing a record. Scope and supplier data practices still matter.

Does a local model solve the access problem?

Not by itself. Model location and tool permissions are separate. A local model may still call a cloud inbox or booking service through a connector.

Is this an ICO compliance checklist?

No. It is an editorial beginner checklist. Check the ICO's current material and get professional advice where the task needs it; completing these steps is not proof of legal compliance.

If I connect Gemini to my inbox, can a human review that data?

Yes, according to Google's Help page "About personalisation with Connected Apps," opened November 10, 2026 (Europe/London). Google's en-GB wording states:

"Data from Connected Apps is used and human reviewed per the Gemini Apps Privacy Notice."

Under the heading "How human review helps improve Google services," it says:

"When connected to Gemini, a subset of connected apps data processed by Gemini is reviewed by human reviewers (including trained service providers) to debug our services, help keep them safe and improve them."

The same page adds this warning:

"Important: Please don't connect your apps to Gemini Apps if they contain confidential information that you wouldn't want a reviewer to see."

Google also says it takes steps before review:

"Before trained service providers review interactions personalised using Connected Apps, we take steps to protect your privacy, such as disconnecting interactions from your Google Account and reducing personal information in this data, unless it's to address abuse or harm. If you submitted feedback, or we need to address abuse or harm, human reviewers may also review some of your Gmail inbox or Photos library media."

The heading "What data is included in human review" was present on the page we opened, but the body under it was not visible to us. We have not filled that gap. This is Google's published wording, not a test of a Connected Apps account, and not legal advice.


How we researched this: documentation context and editorial guidance, prepared November 8, 2026 (Europe/London); no inbox, CRM or bookings test. Context: ICO agentic AI call for evidence. ‡ We could not open the ICO's page to confirm the call's dates on November 8, 9 or 10, 2026. Check the ICO page before relying on them. The salon scenario is fictional. This is general information, not legal advice. See our editorial policy. Images are AI-generated illustrative artwork, not product screenshots.

Also opened November 10, 2026 (Europe/London): Google Gemini Apps Help, "About personalisation with Connected Apps". Human-review wording is quoted from that en-GB page. The heading "What data is included in human review" had no visible body when we opened it. We did not connect an inbox or test Connected Apps.

Keep exploring: Apps · Models · All guides

THE SCOUT LETTER

Good finds.
Less noise.

Useful local AI apps, practical guides and the updates worth your time. A thoughtful note in your inbox.

Free to join. Unsubscribe with one click.