AI Suggestions

AI suggestions is the administrator's inbox for the durable knowledge two AI Assist features propose but never apply on their own: guidance suggestions drafted from closed cases, and question-mining themes drawn from what people ask Ask AccessPoint. Nothing here changes your configuration until an administrator explicitly accepts it.

AI suggestions is part of the optional AI Assist capability. Both features that feed it — Capture learnings and Question mining — only run when AI Assist is deployed. See the Features page.

Where to Find It

Open Settings from the app toolbar and choose AI suggestions in the Foundations group, alongside the Jurisdiction playbook it can propose entries into.

Guidance Suggestions (Capture Learnings)

When a case closes, a Capture learnings card offers administrators the durable knowledge it revealed — organization facts and playbook guidance, each accepted or dismissed explicitly. The same run can also file a guidance suggestion: an append-only refinement to a catalog entry's AI guidance text — the request type, exemption, or extension reason the case actually used.

Guidance suggestions behave deliberately conservatively:

• Accepting adds, never rewrites. Accepting a suggestion appends the proposed text to the target entry's existing AI guidance — it never replaces or overwrites what's already there.
• Edit before accepting. You can revise the proposed wording in the inbox before accepting it, so what lands in the catalog is what you actually want grounding the assistant.
• Dismissed never returns. Once dismissed, a suggestion is gone for good — Capture learnings doesn't re-propose the same guidance from a later case.
• A deleted target is refused politely. If the catalog entry a suggestion targets has since been deleted, accepting it fails gracefully with an explanation rather than creating an orphaned edit.

Question Mining

Question mining clusters what people actually ask Ask AccessPoint into aggregate themes, so administrators can see where the assistant is fielding the same question repeatedly — a signal that a playbook entry or organization fact is missing. Run it from this panel; where a theme reveals a gap, it can draft the missing playbook entry or organization fact directly, ready for one-click acceptance into the same curated stores Capture learnings writes to.

What Question Mining Sees (and Doesn't)

Question mining is built around a deliberately strict privacy contract:

• Only questions are mined — never answers. The assistant's responses are never part of the input.
• Paraphrases and counts only. A theme is a generalized description and how often it came up — never verbatim transcript text, and never who asked.
• Single-question themes are dropped. A theme has to represent more than one occurrence to surface at all, so mining can never expose an identifiable individual question.
• It refuses to run below a volume threshold. With too little question history, mining declines outright rather than producing themes thin enough to be reversible back to a single asker.

Related Pages

• Jurisdiction Playbook — where accepted playbook-entry drafts land, and the process guidance the assistant grounds "how do we handle X" answers on.
• Ask AccessPoint (Case Assistant) — the assistant whose questions Question mining aggregates.
• Organization — home to the organization fact library accepted organization-fact suggestions land in.
• Features — where AI Assist and its individual features are enabled tenant-wide.