Skip to article

What Can an AI Agent Touch, and What Can It Not?

An AI agent reaches only what it is connected to and permitted to do. Here are the published rules for reading, drafting, writing and sending, and the screens that show them.

By Jacqueline V. TwillieNovember 20, 20266 min readGuide
Quick answer

An AI agent touches only what you connect and permit. BNEDai's rule is that every AI agent starts on draft, never send: it drafts, and you decide what goes out. In the published builds each instruction block names the sources it may read, and writes stop at an approval card.

Before an organization lets an AI agent near its work, the questions are the same ones: what can it reach, what can it change, and who says yes. This page answers with what BNEDai has published.

Four actions are worth separating, because each has its own setting. Reading is the AI agent seeing something: a pasted text, a sheet, a calendar. Drafting is the AI agent producing a document or a message that waits for a person. Writing is the AI agent changing something in another system, such as adding a task or creating a file. Sending is the AI agent putting something in front of someone outside. Most of the published builds linked on this page stop at drafting. Reading, drafting, writing and sending are separate decisions, and an approval step can sit on any of them. A reviewer who asks about all four gets a fuller answer than one who asks only whether the AI agent is safe, and the table below shows where each published build puts the line.

Each published build says what the AI agent may read

Each published build's instruction block says what the AI agent may read, and the guide has you test it on a real case before relying on it. The limit is written in plain words inside the instructions, so a reviewer can read it without opening any settings screen, and the same wording can be pasted into your own tool. Three examples follow, each taken from a guide that is already published. The old and new text comparison build tells the AI agent to read only the text you paste. The quiet deals build reads one pipeline sheet and never searches the web. The client onboarding packet build reads one client's signed scope and two templates and uses nothing from any other client's folder.

Drafting ends where a person decides

At the rule BNEDai starts every AI agent on. Jacqueline V. Twillie: "Every AI agent I set up starts with the same rule: draft, never send." Its default is that it drafts, and you decide what goes out. An AI agent that runs on a schedule still needs the same review habits. Keep it drafting and preparing until you have watched its output long enough to trust it, then loosen deliberately, one action at a time, for one AI agent at a time.

Reading, drafting, writing and sending each have their own setting

Each action has its own setting. The table shows the four, with a published build that sets each one.

| Action | How the published builds treat it | A build that shows it | |---|---|---| | Read | Limited to the named source: a pasted text, one sheet, one folder | Quiet deals, client onboarding packet | | Draft | Output arrives as a draft you read and edit | Discovery call recap | | Write | Stops at an approval card; the connector's approval setting is changed from Always allow to Ask for writes and deletes | Meeting notes to action items, client onboarding packet | | Send | The instructions say send nothing, and no sending connector is added | Client onboarding packet, renewal call debrief |

The meeting notes build shows the write case plainly: the AI agent returns the action items, you approve the list, and only then does it add the approved items to your task tool. The platform default in those guides is Always allow, and each guide has you change that setting.

These are the settings in BNEDai's published builds. They are not a claim about every platform, and an organization with its own policy sets its own line.

Ask the person who built it to show you

Ask the person who built it to show you. In the October 7, 2026 session Jacqueline V. Twillie taught with Women Applying AI, the connector's tool management screen separated read-only tools from write and delete tools, and she showed restricting each group. Removing access sat behind three dots beside the connected account. She also said that if you change a password in a connected system, the AI agent can no longer reach it until you come back and reauthorize.

She keeps a 15 minute meeting on her calendar every three weeks to read her own AI agents' instructions, because an AI agent with self-updating instructions needs a person who reads them. The 15-minute review page gives the reading order: the instructions, each skill, the connectors list and the tool settings under each connector.

Three Index dimensions bear on access

Three of its eight dimensions bear directly on this question. The AI Readiness Index is BNEDai's free 18-question self-assessment.

  • The Yes Gate. Does a named human click yes before any AI-assisted action that reaches the outside world, is draft-versus-send an explicit written rule, and can anyone say which actions each tool takes alone.
  • Scoped Access. Do AI tools use scoped, revocable credentials instead of anyone's real password, is each tool limited to the minimum for its one job, and could you revoke a misbehaving tool in under five minutes.
  • Reversibility. Are AI tools barred from deleting or overwriting without a human yes, and is there a known undo path.

What to ask the vendor, and what this page cannot tell you

The vendor's own documents answer questions about the platform: its privacy policy, its security page and its compliance materials. Jacqueline V. Twillie said in the October session that she hands enterprise questions to the vendor's team and shared the links to those pages with the group. She also said she can't tell anyone what their company allows. Take your company's data rules to the person who decides.

A practice account is a reasonable first step. She told the group that a dummy account, such as the one where promotional email lands, is a good way to get comfortable with the level of access you are giving a system.

Frequently asked questions

What can an AI agent touch?

Only what it is connected to and permitted to do. In BNEDai's published builds the instruction block names the sources it may read, and the connector settings decide whether a write stops at an approval card.

Can an AI agent send email on its own?

In the published builds its instructions tell it to send nothing, and the default is that it drafts and you decide what goes out. Whether a given platform allows sending is something to check in its tool settings.

Who approves what an AI agent writes?

In the published builds, the person who built it. The connector's approval setting is changed to ask for writes and deletes, so a write waits for a yes on an approval card.

Reviewed on November 20, 2026.

More on Ownership for Corporate Executives