AI review

AI review helps people and AI understand where work stands, what needs attention, and what to do next. It brings the case’s situation, evidence, and proposed actions into one report.

The value carries across evaluations. Record a result, reevaluate, and see whether the earlier conclusion still holds. The next report considers the previous report, handled proposals, your feedback, and the current evidence.

Start with a case

Open a case in Web and select Evaluate in AI review. Case editors can request an evaluation; people with case work access can read the saved report. Generation runs in the background. Reading a report never runs AI.

An evaluation can consider:

• The case’s purpose, input, and current situation.
• Tasks, their instructions and recorded outcomes.
• Non-deleted records and their structured data.
• The pinned playbook’s instructions and expected outputs, when readable by the case audience.
• Available saved text from linked external sources, excluding external sources on public cases.

These inputs are bounded. A linked URL or attachment does not mean Review opened it, inspected an image, or executed a test. Missing and unavailable evidence remains unverified.

Check the work against its criteria

Review reviews the supplied results and records against the case’s purpose and, when available, the pinned playbook’s instructions and expected outputs. It looks for supporting evidence, gaps, and contradictions that matter to the decision. A missing cost estimate, for example, leaves a recommendation to accept a customer request unverified; a recorded estimate that exceeds the expected revenue provides evidence for reconsidering it.

The report should distinguish requirements stated in the case or playbook from criteria suggested by AI. Inspect the cited evidence and any stated limitations before relying on the conclusion.

Read the report and its evidence

A report puts important information first: the current situation, meaningful changes, open issues, and useful next decisions. Its summary also appears in work lists.

The assessment distinguishes:

Follow cited records, tasks, saved source text, or the pinned playbook to inspect the basis for a claim. An AI judgment supports a decision; it does not replace a designated reviewer’s formal approval.

Turn a proposal into work

A report may propose four kinds of work. Review the full saved contents before confirming; creation uses them unchanged and accepts no title, content, or definition overrides. Creation is an explicit user or agent action; evaluation alone creates no work.

• Create a task to gather evidence, resolve a problem, or advance the next decision.
• Create a playbook suggestion to propose a repeatable improvement. The normal suggestion workflow decides whether it becomes a new playbook version.
• Start a case for a separate purpose, responsibility, or completion criterion. It creates a directed handoff from the reviewed case without closing it.
• Create a playbook with its first version for an evidenced recurring pattern and likely future reuse. A completed case or theoretical reuse alone is insufficient; handled Playbook-creation recommendations are not proposed again for the same case.
• Dismiss a proposal when it does not fit. Optional feedback informs the next evaluation.

Suggestion text is shared with the playbook audience, which may differ from the case audience. Review it for private case details before submitting it.

New cases and playbooks use your active billing home with restricted creator access. Source sharing and external sources are not copied. Confirm using only reportId, proposalId, and action; optional feedback is for dismissal. Retries use the proposal identity, so do not supply idempotencyKey.

Record the result, then reevaluate

For example, a customer request has clear commercial interest but no delivery or support cost estimate. Review points out the missing evidence. The team creates a task, checks the costs, and saves its findings as a record. A new evaluation can then reconsider whether the request makes sense and propose a useful next step.

A closed task is a signal, not proof that the underlying issue is resolved. Keep the actual result and evidence in the case. Handled proposals and feedback help later evaluations avoid repeating work already adopted or dismissed.

A freshness notice appears when meaningful evidence has changed since evaluation. It does not run AI automatically. Select Reevaluate to update the assessment. A notice is advisory: inspect the saved report and the newer evidence before acting.

Control, privacy, and credits

Evaluations currently run on demand.

The case owner’s AI settings control built-in AI permission and web search. Custom home guidance and per-request prompts are not supported. Generation uses token-priced credits from that owning account and charges only when a report is saved. Task, suggestion, or playbook creation costs 5 credits. A new case costs 5 credits plus 2 for its handoff. These actions charge the acting caller’s active billing home; dismissal is free.

A public case link does not expose its review report or summary. Sharing changes hide reports generated for the earlier audience; ownership transfer clears derived reports and action links. Durable conclusions that must outlast a report belong in records.

Use Web, CLI, or MCP

The same saved report and proposal decisions are available across surfaces.

epismo case review generate "$CASE_ID"
epismo case review get "$CASE_ID"
epismo case review act "$CASE_ID" \
  --report-id "$REPORT_ID" --proposal-id "$PROPOSAL_ID" \
  --action create

Generation returns job metadata. Poll get until job.status is complete, failed, or skipped. For MCP use epismo_case_review_generate, epismo_case_review_get, and epismo_case_review_act.

Reference

• AI settings: permissions and settings.
• CLI: arguments and retries.
• MCP: tool arguments and usage.