AI suggestions and their limits
What AI suggestions are built to do in LegalBook, why a person decides every one, what AI refuses, and how the firm controls it. Not available yet.
For: OwnerAttorneyStaff
On this page
Not available yet. LegalBook's AI layer is built and tested with a test provider, but no AI provider is configured for firms yet, and call transcription is not available. Everything in LegalBook works without AI. This page describes how AI is built to work once it is available.
AI in LegalBook is an assistant to the people who do the work. It never gives legal advice, never makes a legal decision and never changes a record on its own.
The features
| Feature | What it would do |
|---|---|
| Call summary | Summarize a completed call and suggest a disposition |
| Thread summary | Summarize an email thread or voicemail for staff |
| Text extraction | Read the text of a scanned document |
| Document classification | Suggest a document type, or a notice type |
| Fact extraction | Suggest candidate facts from a document |
| Correspondence classification | Classify an incoming email |
| Inconsistency detection | Point out contradictions, implausible values and missing facts across verified facts |
| Duplicate detection | Find near-duplicate documents (done inside LegalBook, with no outside provider) |
| Message draft | Draft a client message from an approved template, never sent by AI |
| Routing suggestion | Suggest a queue for work that fell back to default routing |
| SOP recommendation | Suggest relevant approved procedures for a work item |
| Translation | Help with translation, labelled as machine translation |
| Chronology | Build a case chronology from verified facts only |
| Intake answers | Suggest intake answers from a call or a document |
| Knowledge answer | Answer from the firm's approved guidance, with citations |
A person decides every suggestion
An AI result is a suggestion. A person accepts, edits or rejects it, with a reason when rejecting. Only then does the module that owns the record apply it.
- Extracted facts arrive as candidates and must be verified by a person before they become facts.
- A drafted client message goes into the notes; Use draft copies it into the reply for a person to edit and send.
- A call summary in the wrap-up is labelled Suggested, not approved.
Review is required for every feature except a few display-only internal aids (thread summaries, procedure recommendations, knowledge answers, duplicate detection and text extraction).
Suggestions wait in AI suggestions for staff, and in AI suggestions for attorney review for attorneys.
What AI refuses
AI requests and answers are checked against boundary rules in English and Spanish. Requests in these categories are refused or flagged:
- legal advice,
- eligibility decisions ("does my client qualify"),
- conflict waivers,
- case strategy,
- dismissing a deadline or appointment,
- finalizing a filing,
- changing a canonical fact,
- ending a representation.
Eligibility decisions, conflict waivers, deadline dismissals and ending a representation are refused for every audience. Advice addressed to a client, or a prediction of a case's outcome, is refused in anything client-facing. A refusal can be escalated to a person. Firms see the rules and refusals in AI boundaries.
Redaction
Input is redacted before it is sent to the provider, at the level the firm sets for each feature (none, personal information, or strict). The source's sensitivity and the matter's controls are checked first, and attorney-only material is not sent unless the firm allows it.
Firm controls
- AI settings (Firm owner or Firm administrator): switch each feature on or off, with practice-area overrides, redaction level and sensitivity limits. AI is off for a firm by default, and the default redaction level is strict.
- AI sign-offs (attorneys): any feature whose output is drafted for a client, and any unredacted use, needs an attorney's sign-off before it can be switched on.
- AI matter controls (attorneys): turn AI off for a single matter.
- AI prompt templates: approved, versioned instructions, with two-person approval.
- AI activity: every run, with the model, template and policy version, the redacted input as sent, the output, who reviewed it and the cost estimate.