Five decisions AI should never make during legal intake.

1. Final conflict decisions

An intake system can collect names, entities, relationships, opposing parties, insurers, witnesses, and prior counsel. It can normalize that information and place it into a conflicts workflow. It can even flag an apparent match for review. The final conflict determination belongs with the firm.

Names are incomplete, relationships are not always obvious, and a match can require legal and factual analysis. Treating an automated search result as a final clearance risks both false negatives and false positives. The correct output is a reviewable record, not an invisible conclusion.

2. Whether the caller has a viable legal claim

Administrative screening rules are useful. A firm may choose not to handle matters outside a practice area, jurisdiction, incident date, or minimum factual threshold. That is different from allowing a system to decide that a caller has no claim.

A caller's first account is often incomplete. The person may use the wrong legal vocabulary, omit a decisive fact, misunderstand who caused the harm, or be unable to identify an exception. Software can route a matter under firm-approved criteria. It should not present a legal conclusion as fact.

3. What the matter is worth

Case valuation is not a lookup table. Liability, causation, damages, insurance, venue, credibility, evidence, collectability, litigation posture, and timing all matter. Early intake rarely contains enough reliable information to support a responsible value estimate.

A system may collect valuation inputs and surface missing information. It should not give the caller a settlement estimate, suggest a guaranteed range, or use a score as though it were an attorney's considered opinion. The sales benefit is small, while the expectation and professional-risk cost can be substantial.

4. Legal advice or strategy

Intake can answer administrative questions: office hours, consultation format, what documents to bring, and what happens next. Advice about deadlines, rights, evidence preservation, communications with another party, medical treatment, or litigation strategy requires an approved attorney response.

The workflow should recognize advice requests and move them to a defined escalation path. If no person is available, the system should acknowledge the request, avoid answering it, capture the issue precisely, and state the next administrative step.

5. Whether the firm represents the caller

A scheduled consultation is not representation. A completed intake is not acceptance. A positive screening result is not an attorney-client agreement. The system must not blur those distinctions in an effort to sound reassuring.

The firm should approve exact language for consultations, declines, referrals, document requests, and follow-up. Representation begins only through the firm's authorized process. The intake record should show what the caller was told and what remains pending.

The boundary is straightforward: AI may complete controlled intake tasks. Attorneys retain conflicts decisions, legal conclusions, advice, valuation, and representation decisions.

A better division of work

Narrow scope is not a weakness. It is what makes automation testable. Give the system defined questions, approved disclosures, routing conditions, transfer rules, scheduling logic, and a required output. Test normal calls, incomplete answers, caller distress, language changes, advice requests, and failures.

Then give a named person responsibility for quality review. The firm should compare summaries with recordings, examine escalations, correct the workflow, and document material changes. AI can make the front end faster and more consistent, but responsibility remains with the people deploying it.