Perspective 4 min read

Why we do not build a chat interface for contract review

Conversational interfaces abstract away the text. In transactional practice, the exact text is the entire point. Why we keep our tools in the document.

2026 · 10 · 13·admin

It is a frequent question during procurement. A firm will evaluate Review and Draft, note the integration with their playbook, and then ask where the chat window is. They want a sidebar where an associate can ask questions like, “Is our liability capped?” or “Does this agreement contain a non-solicit?”

The answer is that we do not have a chat interface for active contract review, and we have no plans to build one.

The software industry has spent the last three years putting conversational sidebars into every application. For many tasks, this makes sense. But a contract is not a general-knowledge repository. It is a spatial, highly structured legal instrument. When you introduce a chat interface to a drafting workflow, you introduce abstraction. In transactional practice, abstraction is dangerous.

Here is why we keep our models out of the sidebar and in the document.

The spatial nature of drafting

A contract is a machine built of words. Its meaning depends entirely on location, hierarchy, and explicit cross-references. Clause 9.1 might cap liability, but its effectiveness relies on the definitions in Article 1, the carve-outs in Clause 9.3, and the survival provisions in Clause 12.

A chat interface sits outside this structure. It forces the lawyer to look away from the text, ask a question in a vacuum, and receive an answer in a vacuum. The cognitive load of context-switching between a conversational summary and the actual drafting is high. More importantly, it breaks the visual connection to the document’s architecture.

When you read a contract, your eyes need to track the interplay of paragraphs. You need to see that the indemnity is physically separated from the limitation of liability by three pages, which might indicate a drafting error or an intentional disconnection. A conversational interface flattens this geometry into a linear stream of text.

The danger of the summary

Large language models are tuned to be helpful. By default, they synthesise, simplify, and summarise. This behaviour is useful when you are trying to understand a recipe or troubleshoot a database error. It is a liability when you are reviewing a Share Purchase Agreement.

If an associate asks a conversational sidebar, “Are there environmental representations?”, the model is likely to reply, “Yes, the Seller makes environmental representations under Section 4.12.”

This is factually true, but professionally inadequate. The summary fails to mention that the representations are qualified by a materiality standard defined in an entirely different section. It ignores the knowledge qualifier attached to the specific environmental sub-clause. It does not cross-reference the disclosure schedules that effectively gut the representation.

When a model summarises a contract, it inevitably strips away the ‘notwithstandings’, the ‘subject tos’, and the precise phrasing of the qualifiers. In legal review, those conditions are not friction. They are the substance. Relying on a summary encourages a false sense of security, leading lawyers to accept the model’s conclusion without interrogating the text.

Where the eyes are

This is why Review operates entirely within the Word document. We put the intelligence exactly where the lawyer’s eyes already are.

When ZAAN-7B Counsel reads a draft against a firm’s playbook, it does not wait to be asked about a missing carve-out. It reads the document, flags issues by severity (High, Medium, or Note), and places those flags directly on the deficient clauses using the native comments API.

If opposing counsel has inserted an asymmetrical fee-shifting provision, Review highlights the exact sentence. It proposes an inline redline to neutralise it, and it provides a direct citation to the playbook rule that was violated. There is no conversation. There is only the text, the rule, and the proposed edit.

With the release of Review 4.2 earlier this year, we refined this further. The model is now much better at knowing when to say nothing. If a clause deviates from the playbook in phrasing but achieves the exact same legal outcome, Review stays quiet. It behaves like a senior associate leaving a targeted margin note, rather than a chatbot eager to demonstrate how much it knows.

When conversation actually works

This does not mean conversational interfaces have no place in legal software. We use them heavily, but only where the task itself is conversational.

AI Interview uses voice and text dialogue because client intake is an extraction exercise. You are pulling unstructured facts from a human being in distress, attempting to build a structured brief and run statute-of-limitations math. That requires a responsive, empathetic dialogue in the client’s native language.

Similarly, our Trial & Arbitration Simulation relies on conversation to model the examination of witnesses and the interjections of the tribunal. Recall allows plain-language search across a firm’s matter history because querying a database of past decisions is an exploratory task.

But drafting and reviewing are not exploratory tasks. They are execution.

The interface dictates the workflow. If you give a lawyer a chat window, they will ask for summaries, and they will start to trust the summaries over the text. If you give them inline redlines and margin notes, they will read the contract. We prefer to keep the lawyer in the document.

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.