Blog AI access
AI access
How Much Access Should an AI Assistant Have to Email and Client Data?
An assistant that only shows a draft button can sit on access that also sends, changes or deletes. Start from the action the workflow needs, then check what the account can really do.
We read the official documentation for Gmail, Microsoft Graph and Exchange Online, HubSpot, Google OAuth and Microsoft Entra, plus OWASP and Canadian Centre for Cyber Security guidance on AI systems that take actions. The common thread: the name of a permission may not match what people assume it allows. What we offer is a starting access table for five workflows, an approval record and a 20-check test you can run in your own test setup.
5
separate decisions: read, draft, write, send and delete
20
checks in the proposed test, to run in your own test setup
about 1 hour
default lifetime of a Microsoft Entra access token (Microsoft Entra docs)
The thesis
Decide the one action a workflow needs, then check what the connected account can really do. Close the gap with technical limits, an owner and a test, not with an instruction in the prompt.
The scene
A draft button, and what sits behind it
A teaching example.
You want an assistant to draft replies to new inquiries. The setup screen asks you to connect Gmail, you click allow, and the assistant shows a "Create draft" button. It looks like draft access.
Google's own documentation describes the gmail.compose scope as "Manage drafts and send emails". If that is the scope behind the button, the connection can also send. Nothing on the screen tells you that.
That gap between what a workflow needs and what the account allows is what this article is about. Read, draft, write, send and delete are five different decisions, and each one should be made on purpose.
How we looked
Documentation from six platforms and two security guides
What the providers document, fetched October 27, 2026.
We read the official pages for Gmail API scopes, Microsoft Graph permissions and access types, Exchange Online application roles, HubSpot scopes, Google OAuth token revocation and Microsoft Entra access revocation. We also read OWASP's entries on prompt injection and excessive agency and the Canadian Centre for Cyber Security guidance on agentic AI. All were fetched on October 27, 2026.
- What this shows: what each permission is documented to allow, and where the documentation itself warns about gaps.
- Limits: this compares what each provider's documentation says. How a specific account is set up, and how long a revoke takes, show in your own setup.
What the docs say
The permission name does not always tell the whole story
Six documented details worth checking before you connect.
What six permissions really allow
| Platform | What the documentation says | What to check |
|---|---|---|
| Gmail | gmail.compose is described as "Manage drafts and send emails" | Do not call a compose connection draft-only |
| Gmail | The full mail.google.com scope can read, compose, send and permanently delete all email | Whether anything really needs it |
| Microsoft Graph | Mail.ReadWrite can create, read, update and delete mail, and does not include sending | Blocking Send still leaves change and delete |
| Microsoft Graph | Mail.Send can save a copy to Sent Items without Mail.ReadWrite | No need for broad write access just for the sent copy |
| Exchange Online | Application role grants are independent of unscoped grants in Microsoft Entra ID, and consents are additive | Every grant in every place, not one roles screen |
| HubSpot | Scopes open sets of API endpoints and the data behind them | The scopes each endpoint you use needs |
Documented behaviour, fetched October 27, 2026. Not a ranking.
Microsoft also separates delegated access, where an app acts on behalf of a signed-in user, from app-only access, where it acts with its own identity and no user. Do not assume an app-only connection is limited by what the person who installed it can see.
Three layers
What was allowed, what the account can do, what your system lets through
All three have to match.
- Business permission. What did the owner allow? The workflow, the records, the actions and who approves.
- Account access. What can the token and identity really do? The exact scopes, roles and grants for this account.
- Runtime limit. What does your own system let through? Narrow tools, server checks, refusals and a log.
A separate account for the assistant makes its actions easier to trace, but on its own it limits nothing. And counting scopes is not enough: one scope can hold several operations, so compare operations, records, identity and time.
The method
A starting access table for five workflows
A starting policy to adapt to your own systems.
Proposed starting access
| Workflow | Read | Draft | Write | Send | Delete |
|---|---|---|---|---|---|
| Sorting incoming email | Limited | Yes | No | No | No |
| Drafting a reply | Limited | Yes | No | No | No |
| Updating the CRM | Limited | Yes | Approved | No | No |
| Approved follow-up | Limited | Yes | No | Approved | No |
| Suggesting a clean-up | Limited | Yes | No | No | No |
"Limited" means selected records and fields. "Approved" means one specific action a person approved. A Teamulate working method.
The safest start for drafting is a draft inside your own app, so the assistant needs no write access to the mailbox at all. For the CRM, show "before and after" for every change. Also test whether a CRM field change can trigger an existing automation that sends an email: an assistant without a send tool can still cause one.
Approval
Approve the exact action, not a yes/no box
If anything changes after approval, the approval no longer covers it.
Say a person approves a follow-up to one contact. If the recipient, the text or the attachment changes afterwards, the old approval should not let the new version through.
What the approval record holds
| Field | Why it is needed |
|---|---|
| Account, identity and operation | Stops reuse for another account or another kind of action |
| Record or thread ID and version | Ties the action to the item and the state the person saw |
| Recipients, subject, body, attachments | Covers everything the person approved |
| A fingerprint (hash) of the content | Shows whether anything changed after approval |
| Approver, expiry, number of uses | Says who allowed it, until when and how often |
A Teamulate working method.
The approval is stored on your own server. It must not be a field the assistant can fill in, and text inside an incoming email can never grant permission. OWASP's guidance on excessive agency also recommends a person approving high-impact actions before they happen.
Incoming content
An email is data, not an instruction
Limit what can happen, not only what the assistant decides.
OWASP describes indirect prompt injection: an AI system reads outside content, such as a web page or a file, and that content changes how it behaves. It notes that retrieval and fine-tuning do not fully prevent it. An email that tells the assistant to ignore its rules and send a client list is exactly that case.
So the defence cannot depend only on the assistant spotting the trick. OWASP gives a direct example: an extension that summarises a mailbox may only need to read email, so it should not also be able to delete or send. Keep tokens out of the assistant's context, use narrow tools with checked inputs, and check the account and record on every request. The Canadian Centre for Cyber Security makes the same point for agentic AI: least privilege, limited to the exact resources, operations and time needed.
Stop and revoke
Pausing the assistant is not the same as removing access
Plan and test both.
- Pause in your app stops new work and holds waiting jobs.
- Revoke at the provider changes what the issued credentials can still do.
- Confirm by trying the old credentials again in a test and recording the real refusal.
Google notes that after a successful revoke response it might take some time before the revocation has full effect, and that revoking removes all scopes previously granted to the project. Microsoft notes that Entra access tokens last 1 hour by default. A sent email cannot be unsent by revoking access, so measure the time to block locally and the time to a confirmed refusal separately.
Check your own
A 20-check test before you connect
Use test mailboxes and made-up records only.
The proposed test set
| Group | Checks | What it covers |
|---|---|---|
| Scope | 5 | Allowed read, a record outside the list, another account, a blocked field, delete |
| Approval | 7 | Valid action, no approval, changed recipient, body or attachment, expired, reused |
| Life cycle | 5 | Pause, waiting job, revoke, timeout, out-of-date state |
| Edges | 3 | Prompt injection, another tool with the same access, a CRM change that triggers an email |
A practical starting set, not a full list of combinations.
Write down the provider, identity, real scopes and grants, app version and expected result before each check. If a blocked path is not refused, that action stays switched off. If a result is unclear, record it as unknown, not as a pass.
Support
What a marketing team with AI support can keep running, and what stays with you
Start with one workflow and the tools you have.
Start small: one workflow, the accounts it touches, a person who can change those accounts, and drafts reviewed by a person. If a simple template or a shared inbox rule solves it, that is a good answer too.
What stays with you: deciding what the tools may touch, approving anything that sends, changes or deletes, and the accounts themselves.
For B2B teams, this kind of access review is part of the marketing operations our AI marketing team runs alongside you, with every step visible. You can see how our workflows run. To talk it through, request a demo.
Checklist
Checklist: before you connect an assistant
Run it on one workflow.
- One workflow is chosen, with an owner.
- The actions it needs are written down: read, draft, write, send, delete.
- Every real scope, role and grant is listed, including old ones.
- The gap between needed and granted is closed, or accepted in writing by the owner.
- Sending, changing and deleting each need a stored, specific approval.
- Pause and revoke have both been tested in a test setup.
- Every action that changes something is logged, including refusals.
Key terms and facts
The short version
- Scope
- A named permission an app asks for when it connects to an account, such as gmail.compose. One scope can cover several operations.
- Delegated and app-only access
- Delegated access acts on behalf of a signed-in user. App-only access acts with the app's own identity, without a signed-in user.
- Approval record
- A stored record of exactly what a person approved: account, item, recipients, content, expiry and number of uses.
- Revoke
- Withdrawing access at the provider, as opposed to pausing the assistant inside your own app.
Key facts
- TEAMULATE REVIEWofficial documentation for Gmail, Microsoft Graph, Exchange Online, HubSpot, Google OAuth and Microsoft Entra, plus OWASP and Cyber Centre guidance, fetched October 27, 2026.
- DOCUMENTEDGoogle describes the gmail.compose scope as "Manage drafts and send emails".
- DOCUMENTEDMicrosoft says Mail.ReadWrite can create, read, update and delete mail, and does not include sending.
- DOCUMENTEDMicrosoft Entra access tokens last 1 hour by default.
- TEAMULATE METHODa 20-check test (5 scope, 7 approval, 5 life cycle, 3 edge cases).
Takeaways
What to do with this
A button is not a permission
A draft button can sit on access that also sends, changes or deletes. Check the scope behind it.
Three layers must match
What the owner allowed, what the account can do and what your system lets through.
Approve the exact action
If the recipient, text or attachment changes, the old approval no longer covers it.
Test the refusals
A blocked path that is not refused keeps that action switched off.
Questions and answers
Common questions
Is "draft only" access to Gmail really draft only?
Not if it uses the gmail.compose scope. Google describes that scope as "Manage drafts and send emails". The safest start is drafting inside your own app, with no write access to the mailbox.
Does blocking Send in Microsoft 365 make an assistant safe?
Not on its own. Microsoft documents that Mail.ReadWrite can create, read, update and delete mail without sending, so change and delete still need their own limits.
Is a separate account for the assistant enough?
It makes actions easier to trace, but it limits nothing by itself. The limits come from the real grants and from checks in your own system.
Can an email trick an AI assistant into doing something?
OWASP describes this as indirect prompt injection. The defence is to limit what the assistant can do, treat incoming content as data, and require a person to approve high-impact actions.
If we revoke access, does the assistant stop immediately?
Not always. Google notes revocation might take some time to take full effect, and Microsoft Entra access tokens last 1 hour by default. Pause in your own app first, then revoke, then confirm with a test.
Sources
Named sources
Teamulate review of official documentation and security guidance, fetched October 27, 2026:
- Google for Developers: Choose Gmail API scopes: https://developers.google.com/workspace/gmail/api/auth/scopes
- Microsoft Learn: Microsoft Graph permissions reference: https://learn.microsoft.com/en-us/graph/permissions-reference
- Microsoft Learn: Authentication and authorization basics (Microsoft Graph): https://learn.microsoft.com/en-us/graph/auth/auth-concepts
- Microsoft Learn: Role Based Access Control for Applications in Exchange Online: https://learn.microsoft.com/en-us/exchange/permissions-exo/application-rbac
- HubSpot docs: Scopes: https://developers.hubspot.com/docs/apps/developer-platform/build-apps/authentication/scopes
- Google for Developers: Using OAuth 2.0 for Web Server Applications: https://developers.google.com/identity/protocols/oauth2/web-server
- Microsoft Learn: Revoke user access in an emergency in Microsoft Entra ID: https://learn.microsoft.com/en-us/entra/identity/users/users-revoke-access
- OWASP Gen AI Security Project: LLM01:2025 Prompt Injection: https://genai.owasp.org/llmrisk/llm01-prompt-injection/
- OWASP Gen AI Security Project: LLM06:2025 Excessive Agency: https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
- Canadian Centre for Cyber Security: Careful adoption of agentic AI: https://www.cyber.gc.ca/en/guidance/careful-adoption-agentic-ai