I’m the General Counsel at Harvey, and I often get asked how our in-house legal team utilizes the platform. Harvey helps us focus on the work that really matters by cutting down on time-consuming, repetitive tasks. Internally, we estimate our legal team saves around 20-40 attorney hours per week, but as I will discuss below, it can vary somewhat from week to week. This comports with anecdotal evidence we hear from our customers, who consistently report time savings of around 2 to 10 hours a week per attorney.

Before getting into the details, allow me to set the stage. Our company has grown quickly in the last two years — from 2 people to over 200 people. We have offices in San Francisco, New York, and London. We have a large engineering and product team, a large go-to-market team, as well as teams responsible for accounting and finance, people operations, and, of course, our legal team.

Our legal team is organized as you might expect: We have our “PandA” team that handles product counseling, commercial agreements, and procurement (PandA here stands for “Products and Agreements”). We have a Privacy and AI Lead, and we have a Head of Corporate who handles all things corporate, as well as employment, real estate, insurance, and compliance. As is typical for a startup, our “swim lanes” are broad, and folks in our department get exposure to a wide range of issues.

Our heaviest Harvey usage by far comes from our PandA team, who often rely on the platform for creative language when negotiating customer contracts. We have a fulsome clause library for some commonly negotiated provisions, but we routinely encounter requests beyond our usual scope. For example, a customer recently had some cheeky requests around the interplay between indemnification and limitations on liability. After conceptually aligning internally on a way we could get comfortable with the issue, we used Harvey to generate a draft of the revised language. The output required a few tweaks, but was otherwise good to go—and the customer ultimately signed off on it as well. This type of Harvey usage is common in our day-to-day work. In this specific case, it saved us about 10 minutes.

No more first drafts. We use Harvey to help generate drafts of just about everything that we produce. Our template NDA, contractor agreements, and many of our internal policies and guidelines all originate from first drafts provided by Harvey. We are strong believers in clear language (rather than overly legalistic verbiage), so we often preface our prompts for tasks like this with “Draft me [type of document] using easy-to-understand language that doesn’t contain a lot of legalese.” Tasks like this don’t necessarily come up every week, but when they do, Harvey saves us a few hours of legal work at a time.

Below is an example of how we generated a revised version one of the many employee policies that needed to meet both federal and state requirements:

Here is an example of how we generated an email for our sales team outlining best practices for working with legal:

Here, we created guidelines for our marketing team to adhere to when creating customer case studies, allowing them to move quickly while staying within our established parameters.

One of the biggest time savers for our team is Vault. Vault allows us to create persistent projects with up to 1000 documents for each project (with larger limits coming soon). For example, we can upload all our customer contracts to Vault and then run queries against the relevant contracts. Want to know when a particular customer is up for renewal? Just ask.

There was a time when our company was small enough that our lawyers could remember all the details in our head, but those days are behind us. We’ve got hundreds of customers now, and between eval agreements, platform agreements, DPAs, BAAs, and order forms, we are pushing over 1,000 customer contracts. Just being able to quickly query Harvey when someone asks for a specific detail about a specific customer saves us lots of time every day.

Vault can also come to the rescue on bigger projects as well. For example, our engineering team came to legal the other week to discuss some back-end infrastructure changes they were thinking about making. I needed to know if any of our customers had provisions in their contracts with us that would be impacted by the change. In the olden days (last year), I would have spun up a “code red” for our team, created a collaborative spreadsheet of all our customers, and assigned dozens of rows to each lawyer on the team. Each of us would have then taken ~10 minutes to review and update each customer contract in the spreadsheet.

With hundreds of contracts, this process would have taken 15+ attorney hours to complete. But with Harvey, we were able to ask the question of Vault, which created a table of all our customer contracts, Harvey’s responses for each contract, and links to the source documents so that we could verify the results. The Vault query took 3-4 minutes to run, and we then had to review the output, but because Harvey provides pinpoint citations within the relevant document, one attorney was able to complete this task in under two hours. Aside from the huge amount of time saved, it also meant that we didn’t have to pull a bunch of attorneys from their daily work to get this answer for our engineering team. Surprise projects like this occur a couple of times each month, and everytime they do, using Vault saves such surprises from turning into fire drills and saves us days worth of attorney time.

Our security team similarly saves time by utilizing Vault. They’ve created a project containing all our up-to-date security policies. When prospective customers come to us with security questionnaires, we often use Vault to generate a first draft of our responses. To state the obvious, this does not mean that our team isn’t reviewing every question and response, but it does provide a good starting point, saving our security team an incredible amount of time for every such questionnaire.

Vault is a really great example of what generative AI is currently very good at doing: dealing with large amounts of unstructured data. For lawyers and other professionals who regularly deal with large amounts of unstructured data (such as customer contracts, emails, and Slacks for internal investigations), Vault can help bring a sense of order to a sea of data.

Below are a couple of thoughts on how we try to get the most out of Harvey:

Generative AI isn’t the right tool for everything (at least not yet). As an example, we still aren’t where we want to be on Harvey’s usefulness for case law research. While we are highly committed to attenuating hallucinations as best we can, LLM models still can’t guarantee 100% accuracy yet. One of Harvey’s key features is its ability to provide citations for queries involving source documents, so that lawyers can go back and check the work. And while techniques like grounding can help reduce the incidence of hallucinations (that is, when generative AI systems essentially make stuff up), it is important that lawyers do go back and check the work. Harvey (and any legal AI tool) is not meant to be used without lawyer review and oversight, and should not be empowered to make unsupervised decisions.

Generative AI isn’t here to replace lawyers; it is here to augment us and help us reclaim time on certain kinds of necessary but not necessarily fun tasks. We are really excited about what Harvey can help lawyers do, but we remain equally excited about what is to come. The state of the art in AI is moving quickly, and our goal at Harvey is to make it useful and practical to our customers' daily work.