Dastavez AI: building a contract review assistant that lawyers were willing to trust

Dastavez AI's managing partner told us about a first-year associate who had reviewed the same vendor agreement template forty times in a month — each time checking the same fifteen clauses against the firm's playbook, each time late at night because the client wanted comments by morning.

The firm had tried two generic AI tools. Associates stopped using them within a week. The tools summarised contracts fluently, but they did not know the firm's positions, could not say where their answers came from, and occasionally invented clauses that did not exist. For a law firm, that last point ended the conversation.

Our first workshop was with three partners and two associates. We asked them to describe what would make them trust an AI review. The answers shaped every screen: show the exact clause text, show which playbook rule applies, show similar clauses the firm has negotiated before, and never state anything the document does not support.

That became our non-negotiable rule. Every statement Dastavez AI makes links to a location in the contract or in the firm's own documents. If it cannot cite a source, it says so.

The firm's playbooks lived in Word documents: preferred positions, acceptable fallbacks and walk-away points for each clause type, written for humans. We worked with a senior associate to restructure them into clause-by-clause rules with examples, which the review engine could apply and cite.

This work had a side effect the partners valued: for the first time, the playbooks were consistent across practice groups, and gaps — clause types with no agreed position — became visible.

Legal AI is as much about evaluation as about generation, so we staffed a dedicated evaluation role alongside the AI engineers.

Confidentiality and accuracy drove every architectural decision. Each was reviewed with the firm's partners and IT lead.

Everything that shipped in the first firm-wide release.

Before any associate used Dastavez AI on live work, we built an evaluation set from contracts the firm had already reviewed, with the associates' actual comments as the answer key. Every change to prompts, models or retrieval was scored against it: did the review find the issues the associate found, did it flag anything the associate had not, and did it ever cite something that was not there.

Only when the tool reliably found what associates found — and never fabricated a citation on the evaluation set — did we open it to one practice group, then the whole firm.

Associates use Dastavez AI as a first pass, not a final answer. They open a contract, run the playbook review, read the findings with their sources, and then write their own comments — often starting from the suggested redline. Partners reviewing the work can see which issues the tool found and which the associate added.

The most-used feature was one we almost left out: precedent search, which lets a lawyer ask how the firm has negotiated a particular clause for similar clients before.

In professional services, citations are the product. The firm adopted Dastavez AI because it could show its work, not because of the model behind it.

The most valuable data was already inside the firm. Its playbooks and past matters made the tool specific, and no generic model could have supplied that.