AI purpose statement
Intended purpose
Handled uses configured AI to help place work under a subject, summarize a conversation or draft a reply on request, suggest responsibilities during reviewed company setup, prepare advisory notes for configured subjects, research authorized Handled and connected-service context on request, identify possible commitments through separately enabled Outlook checks, and prepare results for explicitly delegated agent work. It is not designed for employee monitoring, performance evaluation or allocating work based on individual behaviour.
The human stays in charge
- Routing selects a subject using the post and ownership map. People can review and correct the result.
- Conversation assistance does not send its draft. Setup suggestions do not change live ownership until an administrator explicitly applies reviewed changes.
- Agent execution can update its progress and prepare evidence, but it cannot approve its own result. A named human supervisor reviews the final result. Subject-help tasks share only the authorized brief and return advisory results.
- Configured subject agents can automatically add advisory notes to delivered Public Requests. Personal research returns private findings and sources to the requester; it does not automatically post a reply, claim work, complete it or contact anyone.
- External agents run outside Handled. Their operators are responsible for any external tools and processing they configure; registering an agent does not itself start an AI worker.
What the models read
- Routing: post text; subject names, aliases and brands; and recent human correction examples. A private post may be processed for routing, although private work is excluded from company-wide correction learning.
- Conversation assistance: the original post and bounded recent reply text the requesting person is authorized to read. Attachment contents are not automatically included.
- Organization suggestions: selected active profiles’ names, departments, directory status and reference IDs, plus current subject names and aliases. Email addresses are excluded. Names alone are insufficient evidence; proposals require review and do not establish actual responsibility.
- Hosted agents: authorized delegated work and replies, applicable supervisor review notes, scoped prior approved answers, and permitted connected-service results. Subject agents receive the permitted task brief rather than automatic access to the parent’s whole conversation.
- Automatic advisory notes: the delivered Public Request, relevant completed Public work in the same subject, and limited recent messages from administrator-selected Slack channels. Notes follow the Handled item’s audience rather than Slack channel membership; administrators must select channels appropriate for that audience.
- Personal Slack research: the person’s question and bounded message excerpts and thread context retrieved from approved channels using their own Slack authorization. Available older history depends on Slack access, plan and retention. Results include source links and may be partial. The service encrypts the personal Slack credential and does not send it to the model.
Ask AI and Microsoft research: Global research reads only Handled work visible to the requester in the selected company. If separately configured and enabled, Outlook uses the person’s explicitly connected work or school mailbox; Teams uses approved channels within their current access. Microsoft consent describes the provider-level permissions. Handled filters sources before sending evidence to the model. Bounded message content can be used for findings and suggested next steps; attachment files are not automatically retrieved. Personal account credentials are not passed to models or company agents.
Supplied content is treated as untrusted data. These safeguards reduce risk; they do not make AI answers reliable or remove the need to review them. Feature availability depends on service and company configuration.
Optional Outlook checks and preparation
Where available, each person can separately opt in to scheduled checks of their own connected Outlook mailbox. Connecting Outlook alone does not start checks. The model receives bounded recent Inbox and Sent Items excerpts and message context to suggest possible commitments or requests. Attachments are not downloaded. Suggestions can be incomplete or mistaken; newer replies and source access are checked before review. This is not a complete audit of the mailbox.
Suggestions and supporting excerpts are encrypted and stored for review by that person only, with a 30-day expiry. Keyed dismissal and expiry fingerprints contain no message text and last up to 90 days. These records are separate from saved private Ask AI research history. A requested reply draft uses the suggestion’s supporting excerpt and context, remains temporary and is not sent automatically. Creating company work requires reviewing the normal composer and audience; private mailbox evidence is not automatically attached. You can pause or turn off checks and disconnect Outlook. See the Privacy Policy for export, deletion and retention details.
Optional research for subject agents
An administrator can permit research sources for a subject agent. Each responsible person separately authorizes their own connected sources for that agent and the displayed subjects. This can start bounded research for eligible new work assigned to that person. Connecting an account or configuring an agent supervisor does not grant access to another person’s mailbox. Research uses currently authorized Handled context and permitted Slack, Teams or Outlook excerpts; account credentials are not given to models or external agent operators.
Findings and supporting excerpts are stored as encrypted private preparation for the responsible person, separate from shared agent results and ordinary Recent searches. Current ownership, consent and source access are checked before review or sharing. The person can edit and explicitly share text into the request; only that shared text becomes company work. Preparation does not automatically send messages or complete requests. Turning authorization off or losing the required access invalidates private preparation. These checks share the hosted execution’s existing spending and token limits and may return partial findings.
Provider and transfers
The model provider is the Anthropic API. The applicable contracting entity depends on the account and governing terms; current commercial terms identify Anthropic Ireland, Limited for EEA customers. This does not mean processing stays in Ireland or the EU. Account-specific agreement verification is pending. The commercial terms exclude training on customer data. Account-specific transfer arrangements and any required assessment remain under verification, as explained in the Data Processing Agreement; public provider terms alone do not establish completion. Hosted-agent model spending is bounded by per-run and per-company daily ceilings that we operate. Optional Outlook checks and draft preparation use separate per-person limits and a service-wide daily budget.
Accountability
Questions, on-demand personal research findings and retrieved excerpts are saved in encrypted private history for up to 30 days, with at most 50 searches per person and company. Only the requester can reopen or delete them, subject to current source access. Expired or inaccessible research is removed; content whose access cannot be verified is withheld. Saved history does not become shared workspace content or an agent’s background knowledge. A person can review exact text and its intended audience before saving it into a request or reply. That explicitly shared content follows normal work-record retention. A private item is visible to its author and current subject owners, not only the person who performed the research. Automatic advisory notes and agent execution records may retain excerpts. Model-provider processing and retention still apply as described in the DPA.
To reduce repeated processing, Handled can reuse recent analysis when a new authorized source check retrieves the same evidence for the same question. Internal analysis records follow the saved research retention and access controls above. The model provider may temporarily cache repeated prompt content under its API processing terms; the DPA's provider-retention terms continue to apply.
Routing and agent work retain the operational records exposed by the service, including delegation, proposals and reviewed outcomes. Not every infrastructure or internal execution event appears in a conversation. Authorized administrators can export available records as JSON from Settings; attachment metadata is included, with file contents downloaded separately. Contact us where those tools do not fulfil an applicable return or access request.