# AI and connected agents

Document version

Version: 2026-10-10.draft-1

What built-in AI and connected agents can receive, how permissions and review work, and what to check before using them with personal information.

AI can help explain work, prepare drafts and suggest a next step. It can also receive the information needed for that request. Your organisation should decide which AI uses and connections it allows.

This draft explains the current application paths. It does not establish a lawful basis for your organisation's processing, approve an external provider or promise a particular account-level retention or model-improvement policy.

## 1. Built-in AI and your connected agents

At a glance: These are separate processing paths and need separate checks.

Built-in assistance, including Looper, uses a server-side model-provider connection. The current implementation supports OpenAI API requests for enabled assistance. The model and features available depend on workspace settings, the plan and how the service is configured.

A connected agent or ChatGPT client uses a connection your organisation or member authorises. The client can ask for information or actions through the tools and permissions it is granted. Its provider receives the material returned by those tools and processes it under that account's agreement and settings.

Connecting a ChatGPT account does not turn it into Tellenze's built-in API account. Do not assume that its chat history, deletion controls, provider terms or data settings apply to the built-in service, or the other way round. The Subprocessors and providers page identifies the draft distinction.

## 2. What built-in AI can receive

At a glance: The request can include your message and the authorised material needed to help.

A Looper request can include your instruction, relevant recent conversation history, authorised context from the page or work you are using, and information returned by available read tools. Enabled features such as draft assistance or Signals processing have their own selected inputs. Information can include names, work content or other personal data supplied by members.

Images and PDF files attached to the current Looper request can be sent as file content for the model to read. File names and the information inside a file can identify people. Review them before submitting the request. A private workspace file can still be disclosed to the model provider through an authorised AI request.

Requests also contain technical information needed to operate the service, such as the selected model, request identifiers and a pseudonymous safety identifier. Responses, provider references and available usage information can become part of the service's local history or operational records.

Use only the material needed for the request. Do not include passwords, secret keys, unrelated customer records or unnecessary sensitive information. If your organisation has restricted a category of information, that restriction also applies to AI messages and attachments.

## 3. What the request controls do

At a glance: The current controls reduce some disclosure and storage, but have clear limits.

The main Looper chat replaces matching email-address text in its outgoing payload. This is a limited text rule. It does not reliably remove names, other identifiers, every possible address format or personal information inside encoded images and PDFs. It does not make the request anonymous.

The request builder sets store to false for model responses. That asks the Responses API not to keep the generated response for later API retrieval. OpenAI's official documentation distinguishes this application-state setting from separate security or abuse-monitoring records and feature-specific retention.

The setting alone does not establish Zero Data Retention, a processing region or an account-wide data policy. The actual agreement, approved account controls, endpoint and enabled features must be checked. This draft has not verified those account-specific details for Tellenze.

## 4. Permissions still apply

At a glance: An AI request does not give a member or connection extra access.

The application checks the responsible member's current workspace access when resolving context and using tools. Connected apps and agents also have explicit grants. A tool being available to a client does not authorise access to every workspace, project, document or file.

The main Looper chat presents supported changes as proposals for approval and rechecks access before applying them. Customer-connected clients, app APIs and configured automations can use the permissions and instructions they have been given; their review flows may differ from Looper's on-screen proposal flow.

Your organisation should decide who can enable a connection, what it can read or change and when a person must review an action. Keep grants narrow and review them when the work or member's role changes. Read access itself can disclose content to an external client, even if it cannot change the workspace.

Work descriptions, documents and material returned by tools can contain instructions written by other people. Treat that material as source content and review the requested action. Permissions and technical controls help limit risk, but they do not prove that every generated suggestion is safe or correct.

## 5. AI email drafts need a separate Send

At a glance: Preparing an AI draft does not send the message.

Email powerups can send your instruction and selected authorised work or conversation context to the model provider to prepare a draft. The preparation request has no sending tools. You must review the sender, recipients, subject, body and any included work information before the separate Send action.

A reviewed send passes the message and addressing information to the enabled Gmail or SMTP connection. The application rechecks the member, relevant work access, current connection and draft version. A provider may retain the sent message, addressing information and delivery records under the connected account's terms.

A delivery record may show provider acceptance or an uncertain outcome. It does not prove inbox delivery. Check an uncertain outcome before sending again. Separate service emails, such as an invitation you trigger, use the transactional-email flow; they are not an AI decision to send a prepared draft.

## 6. Keep people responsible for decisions

At a glance: Check the evidence and use judgement before acting.

AI output can be incomplete, mistaken or misleading. Check facts, source material, proposed changes and recipients before using it. A confident answer or a suggested completion is not evidence that work was completed, tested, approved or delivered.

Do not rely on an AI suggestion as the sole basis for a decision that has legal or similarly important effects on a person. Your organisation must assess the relevant law, safeguards and human review before using automated processing for such decisions.

AI assistance does not supply professional advice or make a sensitive processing activity suitable for the service. Agree additional needs before using it for health, employment, financial, criminal-offence or other high-risk information. The ordinary workspace features are not a completed assessment for those uses.

## 7. History and retention

At a glance: Local history and provider records have different controls.

Tellenze stores conversation and relevant action history so the service can show the work and continue a request. Selected local conversation content and tool records use encrypted fields. Other activity, identifiers or usage fields have their own storage rules. Encryption of a field does not establish its deletion period.

Archiving a conversation, cancelling a pending request or disabling a feature does not automatically delete every local or provider copy. Information already sent can remain in an external account's history, sent mailbox, security records or other permitted retention.

Check the actual provider agreement and account controls for retention and any permitted use of content. The owner must confirm the service's own history, logs, backups and deletion process. Customer-specific return and deletion arrangements belong in the executed DPA; this draft does not promise immediate erasure or a fixed retention period.

## 8. Revoking a connection

At a glance: Stopping access is useful, but does not recall information already shared.

Use the relevant connection controls to remove or narrow an agent or app's future access. Depending on the provider, you may also need to revoke the grant in its account settings. Your workspace's administrator can help check the connection and its remaining permissions.

Revocation does not recall tool responses, erase external chat history or remove an email already sent. Local service records and any provider copies must be handled through their relevant retention and deletion processes.

If a connection has been misused or a credential is exposed, stop the access you control and report the concern. Do not send the exposed credential in your report.

## 9. Your organisation's checks

At a glance: Choose the purpose, recipients and safeguards before enabling a use.

Your organisation must identify a lawful basis, give required information to people, minimise the data and make sure it has authority for the selected processing and disclosures. An access grant, a user request or a draft agreement does not settle those duties on its own.

Review built-in AI and customer-connected accounts separately. Record relevant providers, their roles, storage and access locations, transfer protection and history controls. Where processing presents a high risk, assess whether a data protection impact assessment or additional safeguards are required before starting.

Tell members which uses are allowed and when they must ask for review. Human oversight, careful permissions and clear instructions are useful safeguards, but this notice makes no general claim that every AI use complies with the GDPR.

## 10. Questions and requests

At a glance: Ask your workspace administrator about enabled uses, or contact Tellenze about its service processing.

Your workspace administrator can explain your organisation's enabled AI features, member permissions and connected accounts. For Tellenze's built-in processing, provider review or a data request, contact general@tellenze.com.

For personal information your organisation controls, it normally decides how to respond to a rights request. Tellenze's proposed assistance is described in the Data processing agreement. The separate Privacy notice explains the roles for public website and business activities.

## Sources used for this document

- [Official OpenAI documentation: API data controls, storage and retention](https://developers.openai.com/api/docs/guides/your-data)
- [OpenAI: API service data processing addendum](https://openai.com/policies/data-processing-addendum/)
- [GDPR: responsibilities, rights, automated decisions and impact assessments](https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng)

Contact: general@tellenze.com