AI-assisted curation: planned, optional, and human-governed
Last updated: 6 August 2026
We plan to add an optional feature that helps administrators turn selected terminology-scan findings into reviewable glossary drafts. It is not enabled in the current Marketplace release, and it will never approve, publish, or overwrite a glossary term automatically.
What the feature would do
- An eligible administrator runs the existing Confluence terminology scan.
- The administrator selects a bounded set of findings and chooses Generate draft candidates.
- The feature groups the selected terms and proposes a canonical term, aliases, a plain-language draft definition, and source-page references.
- The administrator reviews, edits, selects, or discards each candidate.
- Selected candidates become ordinary glossary drafts.
- The existing four-eyes workflow remains required before a term can become approved.
The intended benefit is reduced manual drafting effort after a scan. It is not a legal interpretation service, a compliance decision-maker, or a guarantee of a correct definition.
What it would not do
- It would not run automatically or scan an entire Confluence site for AI processing.
- It would not approve, publish, or silently overwrite terms.
- It would not generate regulatory mappings in the first release.
- It would not provide model confidence scores or generated legal reasoning.
- It would not replace the customer’s own reviewers, terminology owners, or approval process.
- It would not affect the availability of the existing scanner, manual draft creation, or four-eyes approval workflow if the AI service is unavailable.
Planned data flow
For a request deliberately initiated by an eligible administrator, the app would fetch minimal supporting sentences from the selected Confluence pages and send the request to Atlassian’s Forge LLM capability. The app does not use a DailyMind-operated model endpoint for this feature.
| Data category | Planned use | App persistence |
|---|---|---|
| Selected term names, occurrence counts, and page identifiers | Bound the request and link output to the administrator’s chosen findings. | Evidence metadata may be retained with the candidate and audit event. |
| Minimal supporting page sentences | Provide context for a draft definition and aliases. | Not stored in curation records, audit events, logs, exports, prompts, or raw model responses. |
| Draft term, aliases, source-page references, and model/version metadata | Enable human review, traceability, and normal draft creation. | Retained according to the approved production retention rule; that rule is not yet authorised or published. |
We will publish the exact processor, retention, and data-flow position before enabling this capability. Until then, we do not make a “no data egress,” data-residency, model-quality, service-availability, response-time, capacity, or cost commitment for AI-assisted curation.
Third-party dependency
The feature depends on Atlassian’s Forge LLM service. DailyMind cannot control or guarantee its availability, model/version behaviour, moderation behaviour, usage limits, pricing, or Marketplace upgrade process. If the service is unavailable, the AI feature will fail safely and the existing non-AI workflows remain available.
Atlassian documents Forge LLMs and their upgrade requirements in its Forge LLMs API documentation. Forge invocation constraints also vary by function type; see Atlassian Forge invocation limits.
What must happen before release
- We must record staging evidence for successful and failed requests, moderation, retry behaviour, evidence freshness, usage, cost, and the upgrade experience.
- We must approve and publish the data-flow record, processor/retention position, privacy and Marketplace disclosures, support guidance, enablement rule, capacity limits, and pilot criteria.
- Atlassian Marketplace must process the major-version release that adds Forge LLM capability.
- Each customer’s Confluence administrator must approve that upgrade for their tenant.
- Only then may eligible customer administrators choose to use the optional feature.
Customer controls
- Administrators choose the findings submitted for a draft request.
- Administrators and reviewers can edit or discard candidate drafts.
- The creator or last editor cannot approve the resulting term.
- A different eligible reviewer must perform the normal four-eyes approval.
Related documents
- Compliance for Confluence — approved terms, page scanning, and audit evidence for regulated teams in Confluence
- Privacy Policy — current Marketplace-release data handling.
- Security Whitepaper — current Marketplace-release security boundary.
- App Limitations — product and platform constraints.