Mountains of Documents
Before Jevin started building AI for legal work, he’d spent years working in machine learning and deep learning across other industries. What drew him to legal tech when he joined Clio around 2023 was a specific technical challenge that had resisted good solutions for a long time.
“There’s so much nuance in a legal document,” he says. “The answer is in there, but it’s often really hidden within sentences.” Fund documents are dense, cross-referential prose, where the meaning of one provision often depends on language buried somewhere else entirely. Traditional software couldn’t handle that reliably, and the volume made it worse. A single engagement might involve hundreds of documents, each with its own structure and its own set of buried answers.
When LLMs arrived, Jevin recognized the technical fit immediately. “It was like a perfect solution to the problem we’d always had in legal tech: how do we deal with very large documents, make sense of them, and draft from them?” Large language models are very good at exactly this kind of work, reading dense text at scale, finding the relevant passage and surfacing what matters.
Vibe-Driven Development
The most common failure mode in legal AI, Jevin argues, is that developers who don’t understand legal work can’t tell whether an AI answer is right.
“What a lot of companies do is paint some lipstick on an LLM and call it their own legal tech solution,” he says. “All that’s really happening is they’re forwarding text from a document into ChatGPT and asking it questions. And if it looks like it kind of makes sense, especially for a developer who’s not a lawyer, they ship it to production.” He calls this, half-mockingly, vibe-driven development. “The vibe of this sounds legal, therefore it should be good.”
The problem is that legal work is genuinely nuanced in ways that aren’t visible to a non-lawyer. A system can be correct on easy questions and wrong on the ones that matter. Testing it once or twice against a simple document proves nothing about how it will perform across the full range of real-world complexity.
At Covenant, the lawyers and engineers work inside the same firm, and that changes how the product gets built. Covenant’s legal team works directly with clients and brings deep domain experience to every deal. Jevin’s engineering team relies on those lawyers to define and refine the product, and when the engineers aren’t sure how to guide the models on a particular piece of legal work, the lawyers step in to shape the approach.
How a Review Works Now
Take a fund document review. A lawyer used to read the document from start to finish against the client’s checklist of terms: fees, MFN rights, transfer limits, reporting duties, key person provisions and the rest. Each answer was typed into a memo or spreadsheet with the clause reference beside it. When a new draft arrived, the lawyer compared it to the last one and re-checked every term, including the ones that hadn’t changed.
Covenant’s system now answers the client’s full list of questions against the documents. Each question goes to several AI models separately, and a separate “judge” model compares their answers. Where the models agree, the answer is accepted. Where they disagree, or the answer is unclear, the question goes to a lawyer rather than leaving the system to guess. Every answer links to the passage it came from, so a lawyer can check the source in one click. When a new draft arrives, the system identifies what changed since the last reviewed version and re-asks only the questions those changes could affect, carrying over the answers for untouched terms. After the lawyer signs off, it flags issues against the client’s standards and drafts the summary the client reads.
The questions the models couldn’t agree on are usually the hard, ambiguous clauses, which is where legal judgment counts most. Lawyers check and correct the AI’s answers before anything reaches the client, and nothing is released without a person approving it. They also decide what the client should push back on. The system can flag that a term misses the client’s standard, but whether it’s worth negotiating depends on the relationship and the deal.
From Text to Data
The change Jevin credits with saving lawyers the most time is one most users never see: turning document text into structured data.
An AI answer to “what’s the management fee?” comes back as a sentence, something like “2% of commitments, stepping down to 1.5% after the investment period.” A lawyer can read that, but it can’t be sorted, charted or compared across 200 deals. So Jevin’s team gave every question a type: a percentage, a dollar amount, a date, a yes-or-no, or a choice from a fixed list. Once a lawyer approves an answer, a second step converts it into a typed value: 2%, 1.5%, step-down: yes. If a question’s type changes later, the system re-converts every past answer to that question, keeping the whole portfolio consistent without anyone re-reviewing old deals.
That single change turned a client’s history of reviewed documents into a database of terms, and everything Covenant built afterward reads from it. Clients can chart any term across their deals, such as fees against fund size. They can see what’s normal for each term across their own portfolio and how far a new deal sits from it. Outlier-finding agents query the data, identify terms that are rare or off-norm for that client, and flag them for the lawyer with their reasoning. Exports give clients spreadsheets with typed columns they can work with directly.
What’s coming
Jevin estimates that engineers are about two years ahead of most knowledge-based professions in how they’ve integrated AI into daily work, and the pattern there is instructive. Senior engineers are benefiting most because they already operate at an abstract level. “I spend a lot of my time just thinking and planning on how features are going to go,” Jevin says. “I do very little hands-on coding now, because the LLM can usually do it faster and with more completeness. I spend a lot more time reviewing the code and editing plans.”
He sees a parallel forming in law. Senior lawyers will increasingly spend their time on negotiation strategy, position strategy and building playbooks for different deal scenarios. “They get a document in, they tell the LLM to go and change it based on the strategy plan they’ve worked on together, it goes and does it, and then they quickly review the redlines and send it off.”