If there’s one thing that I’ve learned along the way, it’s that we have the most amazing users! By sharing some of these more intimate details I hope to paint a vision for the future of legal work (and the adoption of AI as a whole) that hopefully excites you as much as it excites us. Because we will need your help to achieve it.
But before I talk about the future, let me take a step back and explain how Donna evolved.
A company we hated…
It all started in early 2017 when Anton Gerdes and I co-founded a company that tried to build a structured database from “fuzzy company data” (somewhat like Diffbot and Clearbit). We did things like crawl company websites to detect people profiles on “Team” pages, then turned this snapshot into “Company A has 3 developers and 5 salespeople” in our database. We were tracking hundreds of these signals and we indexed over 2 billion company websites.
On a technical level it was very stimulating; Anton got to apply his love for NLP (Natural Language Processing). And thanks to Google Cloud (who offered us €100.000 in server credits) I got to play around with massive server clusters. But one thing was missing, we didn’t have a clear end-user, someone that we could make happy and talk with. We wanted to build a product for people, not a data-layer. There was no “product-founder” fit. So in late 2017 we decided to look at some alternative uses for our technology, to see if we could pivot (I recently did a talk on this story for Sony, so feel free to email me if you would like to hear more). We spent an afternoon brainstorming ideas and gave ourselves a week to try them out.
Seed of an idea…
An experience I had in my childhood/teenage years stuck with me. I wrote my first “export control & license agreement” at age 12. My mother was an IP lawyer at Shell at the time and I was desperate for a summer job. So I spent the summer turning Word Perfect agreements into Microsoft Word templates.
I was given a list of common mistakes that creep into legal agreements. Often typos occur when you create placeholders for the
[Company Name]. For example, if the agreement was for “Amazing Inc”, you would just use find-and-replace everywhere, right? Unfortunately, because lawyers often copy-and-paste clauses from other agreements, it's easy for another entity name such as “Amazing LLC” to sneak in. Or to even leave an entirely different name such as “Mediocre GmbH” in the original agreement. When you’ve been working on a 200-page draft for weeks, it’s so easy to read over such little mistakes, but its mistakes like this can turn out to be quite costly.
Jumping back to 2017, while brainstorming ideas for our pivot. We had built a piece of technology that would extract information from the Terms & Conditions pages on company websites. T&C documents contain a lot of valuable information including the official company name and where the company operates from. But when we first used the new tool it kept extracting incorrect company names. Initially we thought this was a bug, but we soon realised that the Terms & Conditions actually contained the wrong company name. In fact 20% of websites on the internet had the wrong company name somewhere in their Terms and Conditions, simply because people had copied-and-pasted them from somewhere else and missed a name. We began to build a tool that could recognise this mistake and correct it.
Falling back on what I had learned from my summer as a law intern, we connected the idea that we could use this system to help lawyers with essentially similar mistakes. To test and validate the idea we launched a landing page that said
Lawyers, send us your agreements and we will highlight undefined terms.
To our surprise, we received several highly confidential documents and it became very clear that we were onto something.
Finding our Core…
Over the next few months we got to re-use a lot of our core technology and experience in NLP and we married it with our insights and connections into the legal industry.
Donna started out with a very simple rule based system. But we quickly realised that these more traditional machine learning methods led to too many false positives. Our early users had to dig through a lot of uninteresting suggestions to find a few good pointers. It’s a bit like having an office intern that constantly runs into your office confused about how to deal with trivial matters.
We wanted to offer our users a more executive experience…
“These are some important things for you to look at. I’ve fixed these for you, can you double check that they are correct. And these things you might want to look at if you have time. Oh and also I’ve made sure to send your client a little gift because they closed that deal you drafted for them last week”.
So we started to build deeper and more advanced neural networks to teach Donna how to prioritise and evaluate her own comments. We were essentially using the same algorithms used by Facebook and Google to optimize clickthrough on ads. But instead of using those algorithms to make you buy more things, we use them to help you be more productive. A real breakthrough came when a few early users requested we help them to identify some very complex and subtle issues. They wanted to make sure that certain key dates happened before certain events.
Our earlier tests made it clear that this was not a trivial problem as we would have to teach Donna a lot of "common sense", something which AI is generally very bad at.
We did manage to teach Donna how to recognise important details, such as key dates. But, instead of having Donna figure out if there was a problem we used Donna's understanding of the agreement and had her create a simple overview for the lawyer in the form of a suggestion pane. And the response was overwhelmingly positive.
We realised that the future for Donna was as Assistive Intelligence. Just like the TV character I jokingly mentioned before, we could use Donna's understanding of the contract to provide lawyers with relevant information at the right time. Although Donna is a simple tool to use, she saves a lot of time and scrolling back and forth in the document. It's absolutely magical!
We already saw that for every 10 issues that Donna finds, 6 get accepted by the lawyer. But we also saw that by providing the lawyer with more relevant information, they themselves find another 6 issues that would have been impossible for us to find.
That is the true power of machine-human collaboration.
Interestingly, the feedback we received didn't just focus on the time saved, or the improved quality...Donna genuinely made people happy!
On a Mission…
Lawyers spend a lot of time having to do robot-like tasks such as proofreading, formatting documents and copying-pasting snippets. This is not the strategic, and engaging work that most lawyers enjoy. It's frustrating, error prone and time consuming work, that is often not billable either. We don't think our AI will or should ever replace a lawyer. Our mission is to replace the robot within™ a lawyer. By taking care of robotic tasks we free up a lawyer's valuable time to creatively design solutions and provide a better service to their clients.
Donna isn't about saving time, it's about helping lawyers invest their time into work that they find engaging that will set them apart from their competition.
We hope that one day Donna can serve as a role model for machine-human collaboration in the same way that Salesforce became an example of the SaaS business model.
Over the coming months we will continue to improve the Donna add-on for Microsoft Word. But we can’t do this without you, so please keep sharing your feedback with us. As we bring on a few new team members, we can focus on updating Donna based on your feedback more frequently.
In the slightly longer run we plan to leverage Donna's ability to read and understand agreements to bring some exciting new knowledge sharing features. We hope to make information, snippets, and best practices more accessible to everyone within your team or firm.
And of course we will be working hard on sharing Donna with as many lawyers as possible. We already have users in more than 37 countries mainly because of your recommendations! We would love your continued help in sharing Donna with colleagues or friends. If you know someone that might (or should) be interested you can always make an intro to firstname.lastname@example.org and we'll make sure to give them VIP treatment on your behalf.
Thank you for being a part of our journey.