AI Interview is generally available: the first conversation, handled with care

AI Interview is generally available from today, after a preview that ran from last September across 41 firms, legal-aid clinics and in-house teams.

The product does one thing. It conducts the first conversation with a prospective client, by voice or by text, and turns it into a structured intake brief that a lawyer can read in two minutes. It does not give legal advice, it does not decide whether to take the matter, and it does not pretend to be a person.

This note covers what ships, what we learned in preview, what to configure before switching it on, and what it does not yet do.

What ships

• Voice and text intake in 38 languages. The client chooses the channel and the language. The brief is produced in the firm’s working language regardless of the language of the conversation.
• A structured brief. Parties, a chronology in the client’s own words with dates normalised, the relief or outcome the client is seeking, documents the client mentions having, and a list of questions the interview could not resolve. Each item in the brief links to the point in the transcript it came from.
• Limitation-period arithmetic. Where the matter type and jurisdiction are identifiable, the brief includes candidate limitation dates, the rule applied, the trigger date assumed, and a confidence note. This is a date calculation from facts the client gave, not a legal opinion, and the brief says so.
• Conflict check at the end of the conversation. Names of parties and related entities are checked against the firm’s matter database before the brief is released to a lawyer. Potential conflicts are shown to the intake supervisor, not to the client.
• Transcript, recording and brief stored under the firm’s retention settings. Zero retention on our side; per-tenant encryption; data residency in the firm’s chosen region.
• Escalation to a human. The client can ask for a person at any point. The interview also escalates on its own when it detects distress, a possible emergency, or a matter type the firm has configured as “always a person”.

What we learned in preview

Clients tell a machine more, and less. In the preview cohort, intake conversations ran longer on average than the firms’ human-conducted first calls, and the chronologies in the briefs were more complete. Several clinic supervisors noted that clients disclosed things to the interview, about immigration status or domestic circumstances, that they had been reluctant to raise with a person on a first call. The reverse was also true: clients were less likely to ask the interview the question they actually came with, which was usually “do I have a case”. The brief now includes a field for that question, and the interview is tuned to invite it.

Pace matters more than vocabulary. The largest source of abandoned conversations in early preview was the interview moving on before the client had finished. We slowed it down. Silence is now treated as thinking, not as an end of turn, for considerably longer than a typical voice system would allow.

The conflict check must be last, and invisible. An early build ran the check as soon as names were given and, in one configuration, told the client the firm could not help. That is the wrong moment and the wrong messenger. The check now runs after the conversation, the result goes to the supervisor, and the client is told only that the firm will be in touch.

Clinics and firms want different things from the brief. A legal-aid clinic triaging housing matters wants eligibility fields and urgency up front. An M&A practice taking a first call from a founder wants the counterparties and the deal shape. The brief template is configurable per intake stream, and the defaults differ by practice type.

What to configure before switching it on

1. Intake streams. Which matter types the interview will handle, and which go straight to a person.2. Escalation rules. The phrases, topics and signals that should hand over to a human immediately. The defaults cover emergencies and distress; firms add their own.
2. The opening statement. The interview introduces itself as an automated intake assistant, says that it is not a lawyer and that nothing in the conversation is legal advice, and explains who will read the brief. The wording is editable within limits; the disclosures are not removable.
3. Conflict-check connection. The matter database the check will run against, and who receives potential-conflict notices.
4. Retention. How long transcripts and recordings are kept, and whether recordings are kept at all.6. Languages. All 38 are available; firms may restrict the list to those their lawyers can follow up in.

Known limitations

• Limitation arithmetic depends on a correct trigger date. If the client is wrong about when something happened, the calculation is wrong. The brief shows the assumed trigger date prominently so the lawyer can check it first.
• Jurisdiction detection is not certain. Where the client’s account could fall under more than one jurisdiction, the brief lists candidates and computes limitation dates for each rather than choosing.
• Conflict checking is only as complete as the matter database. Entities recorded under variant spellings or former names may not match. The check uses fuzzy matching and shows near-misses to the supervisor, but it will not catch what is not recorded.
• Voice quality in some languages is behind others. Recognition accuracy in the 38 languages ranges from very good to adequate, and is weakest for strong regional accents in languages where our training data is thinner. The brief flags passages the recogniser was unsure of.

The first conversation sets the tone for everything that follows. We have tried to make this one patient, honest about what it is, and useful to the lawyer who picks it up next.