Release 4 min read

Interview 1.8: dynamic probing for incomplete chronologies

Clients rarely recount events in a linear sequence. Interview 1.8 builds a real-time temporal graph during intake, identifying missing dates and asking targeted follow-up questions.

2026 · 11 · 07·admin

When a prospective client calls a legal-aid clinic or a firm’s intake line, they do not recite a chronological list of legally significant events. They start with the most emotional moment—the termination, the injury, the breach—and work backwards.

We launched AI Interview in May 2025 to handle client intake and produce a structured brief for the reviewing attorney. The voice and text agent performed well at capturing the names of opposing parties, identifying primary grievances, and running initial conflict checks against a firm’s matter database. But chronologies remained fragmented. A client might say, “They stopped paying the invoices after the summer, which was right around the time the first shipment arrived late.” For a human associate, this requires a follow-up question. For earlier versions of our agent, it resulted in a vague date entry in the final brief.

With Interview 1.8, shipping today across all 38 supported languages, the agent now builds a real-time temporal graph of the conversation. When it detects a missing dependency—a date required to calculate a limitation period, or an undefined gap between two material events—it formulates a clarifying question before the call ends.

The problem with human memory

A structured intake brief needs anchors. If a commercial tenant says they received an eviction notice, the exact date of receipt dictates the response window. If an employee alleges retaliation, the time elapsed between the protected disclosure and the adverse action is the core of the claim.

Previously, our system relied on static logic. It knew it needed a date of incident. If the user provided one, it moved on. If the user was evasive or unsure, the system often accepted a vague approximation to avoid frustrating the caller. This defeated the purpose of the automated statute-of-limitations math we built into the platform. A limitation period cannot be calculated from “sometime around Easter”.

Constructing the event graph in real time

We use our in-house model, ZAAN-7B Counsel, to parse the transcript as it unfolds. Every material event mentioned by the caller is treated as a node. The model attempts to assign an absolute date or a relative time constraint to each node.

If the caller states they signed the NDA on 14 October, and claims the counterparty launched a competing product “three weeks later”, the system assigns a relative value. It does not need to interrupt. Parsing time expressions across 38 languages introduces significant complexity here. Idiomatic time references in Mandarin or Arabic do not map cleanly to Gregorian calendar dates without context. ZAAN-7B Counsel normalises these inputs into a standard ISO 8601 format in the backend, preserving the original phrasing in the trace view so the attorney can verify the translation.

However, if the caller mentions a breach of confidentiality but provides no temporal anchor at all, the graph registers an unresolved node. The system evaluates whether this missing date is strictly necessary for the intake brief. If the missing date relates to a peripheral detail, the agent stays silent. If it is the trigger for a statutory deadline, it queues a clarification.

Calibrating the interruption

Interrupting a client who is recounting a stressful event is poor practice. We spent the last four months working with our Open Bar users—specifically legal-aid clinics in London and our residency cohort in Lagos—to refine when and how the agent asks for dates.

The agent is instructed to wait for a natural conversational boundary. Instead of cutting the user off, it holds the unresolved node in memory. When the caller concludes their thought, the agent circles back: “You mentioned the supplier sent a revised statement of work. Do you remember the exact date you received that document?”

This required adjusting the prompt weights for our empathy layer. The agent must balance the strict requirements of civil procedure with the conversational tone expected by a distressed caller. In our beta testing across active accounts, this delayed-probe approach increased the completion rate of chronologies from 62 per cent to 89 per cent, without increasing the caller drop-off rate.

Feeding the limitation-period calculator

The primary utility of a complete chronology is triage. Firms use Interview to route high-priority matters to the top of the queue.

Once the temporal graph resolves, the system runs the limitation-period math based on the jurisdiction identified. For instance, a breach of contract claim under English law generally carries a six-year limitation period from the date of the breach, whereas specific employment tribunals require filings within three months less one day. By securing the exact date of dismissal or breach during the first interaction, the triage paralegal knows exactly what they are looking at before they open the file.

If the math indicates that a statutory deadline expires within the next thirty days, the resulting intake brief is flagged as High severity in the firm’s dashboard. Because the rules governing limitation periods vary wildly, the system does not attempt to provide legal advice to the caller. It merely structures the dates and flags the calculation. The timeline is presented in the final brief alongside the exact transcript citations, allowing the associate to verify the logic instantly.

A polite, precise follow-up question during the initial call saves an associate three emails later.

See it on a contract you have already reviewed.

Send us a draft your team has already redlined and we will show you what ZAAN catches, and what it misses.