Della
Compliance supportContract: analytics/drafting/review
dellalegal.com
Contents
- What does it claim to do?
- Substantiation of claims & potential issues
- How might the end-user assess effectiveness?
- What form does it take?
- Is it currently in use?
- The creators
- Jurisdiction
- License
What does it claim to do?
Della is a platform that allows users to automate the document review process through the creation of checklists of questions which are answered by the underlying Question Answering system.
Claimed essential features
- Review and summarise legal documents.
Using question answering, “we help lawyers to review documents as quickly as possible, and secondly we help lawyers to summarise long and complex documents, presenting the key information in customisable reports.” (How Della is transforming the way lawyers interact with legal documents, archived).
“The expected answer outputs for these questions will be a variety of dates, numbers and yes/no answers.” (ibid.)
- Apply playbooks and checklists from different legal domains.
- Assess documents across languages.
Checklists are available for specific types of document review, e.g. “Commercial Lease Agreements, Data Processing Agreements (DPA), Loan Agreements, Non Disclosure Agreements (NDA), Software-as-a-Service (SaaS) agreements, and many others, including distribution and supplier contracts.” (Demo request; archived)
“Select from 20+ pre-defined checklists for specific document types, for example Commercial Leases, M&A and Due Diligence, or easily create your own in minutes; Gain competitive advantage. Many clients now request AI tech to be used on deals and negotiations.” (Della brochure page; archived)
“Due diligence: Della is pre-trained on standard M&A, due diligence questions and potentially problematic clauses; Contractual Review: Della extracts essential terms and can highlight any deviations from your expectations; Compliance: Ask Della to use your compliance playbook to analyse client contracts and spot potential issues; Lease analysis: Della can find standard lease provisions and metrics across a wide contract portfolio; Force Majeure: Della can detect contractual risks during the Covid-19 crisis; GDPR: Della can analyse Data Processing Agreements (DPAs) to help determine GDPR compliance.” (Della for law firms; archived)
“Della’s multilingual capabilities overcome challenges involved in cross-border deals by allowing teams to upload documents in one language and ask questions in another” (Della Platform; archived)
Claimed rationale and benefits
- Enable firms of all sizes to assess contracts accurately and quickly.
- Reduce time spent by lawyers on contract review.
- Make contract review easier to perform.
“Our mission is to democratise the use of AI. We want to make it easier for firms of all sizes to review contracts with their own checklist or playbook as fast and as accurately as possible.” (Della mission statement; archived)
“Della’s AI accelerates the contract review process so lawyers can spend more time on legal work rather than paperwork: Easy to get started; Create personalised checklists; Review contracts even faster.” (Della homepage; archived)
“While traditional contract review focuses on clause detection and data extraction, Della focuses on the questions legal professionals want to ask of their contracts. Using proprietary AI, Della allows users to ask questions about their contracts in plain English (and other languages), and get clear answers.” (Della Platform; archived)
Claimed design choices
- Uses large :language models and transfer learning to enable multilingual contract review.
“Della can analyse documents in English, French and Dutch and can answer questions in more than 30 other languages; Della is affordable. Through Della’s unique ‘transfer learning’ system, smaller law firms can leverage the tool for sophisticated contract analysis without having to train their own models, saving both time and money; Della becomes smarter over time. Della uses cutting-edge massive :language models to power its AI, and an innovative ‘transfer learning’ system, which allows users to share training securely to ensure that Della gets smarter every time it is used.” (Della brochure page; archived)
“Della has already been trained on 20+ languages” (Della Platform; archived); “The core :language model is trained in more than 100 languages” (FAQs question 12; archived)
“Della is built on massive :language models which are pre-trained with enormous amounts of data… This means that our AI does not require large amounts of legal data and human supervision… …and it is smarter because you are building upon an existing layer of intelligence.” (Della Platform; archived)
- Uses natural language question answering.
“While traditional contract review focuses on clause detection and data extraction, Della focuses on the questions legal professionals want to ask of their contracts. Using proprietary AI, Della allows users to ask questions about their contracts in plain English (and other languages), and get clear answers.” (Della Platform; archived)
“Other available legal AI solutions focus on detecting clauses… At Della, we take a different approach; we allow you to ask your own questions, in your own words.”
“When you ask Della a question, it can give you a simple ‘Yes’ or ‘No’ answer and raise possible red flags”
“Della does not learn to detect specific language patterns but learns to understand questions and their associated answers”
“The platform continually improves, ensuring all users benefit in a safe and secure manner”
- Use of transfer learning means a low number of training documents is required.
- Models are retrained weekly.
”[Training requires] generally between 5 and 30, depending on the document” (FAQs question 3; archived)
Della retrains “usually weekly”, allowing for removal of documents from the model e.g. if a client deletes them or closes their account (FAQs question 10; archived)
“Our models learn directly from your corrections so you can review documents faster” (Della Platform; archived)
Substantiation of claims & potential issues
- Della’s use of multilingual transfer learning may elide important differences of linguistic, cultural, and legal context, particularly between the legal jurisdictions whose languages the system is claimed to work in (legal systems in the various English-speaking, French-speaking, and German-speaking jurisdictions vary significantly).
There are no scientific or technical publications available that describe the backend of Della. The website features some white papers and a FAQ section.
Data
- Della is suitable for contract review, the example on the website suggest it is possible, for instance, to review lease agreements in multiple languages.
- The extensive list of languages is not provided, however from the website of the company we assume that it can at least handle English, German, and French.
- They claim to use massive :language models for their systems, and claim that it allows to not have to use much legal data. This suggests that the models themselves are not trained on legal data (https://dellalegal.com/della-platform-legal-contract-ai/#video).
The system
- The blog post ‘Leaving the sandbox: from school to the real world’ (archived) mentions the use of :Named Entity Recognition system but there is no information on how this is used for the Question Answering system.
- From the marketing material present on the website, it can be inferred that Della uses a multilingual Transformer model with an additional layer for Question Answering.
-
There is no information available on the methodology or training material used to refine the :language model to the legal domain and what the coverage is for different legal subdomains.
-
The case study described in ‘White Paper – A Pragmatic Approach to AI for Legal Departments’ (archived), at p. 13, mentions a review phase in which employees at Della finetune the system on client specific data in three stages: “The Della team, reviewed 3 batches of document iteratively. This helped to train the system and measure the accuracy of the data extractions”.
Evaluation metrics
- Question 8 of the FAQ refers to performance metrics: “We use a number of metrics internally, for example, what percentage of the answers are right on our first suggestion, what percentage of our answers are correct with one of our suggestions and F1 which is a more sensitive metric.” (FAQs; archived). These numbers are not presented.
The GDPR subprocessors page (archived) refers to the OCR engine used:
- “ABBYY OCR SDK. We use the OCR SDK API provided by ABBYY to extract the text of scanned documents”. ABBYY is a cloud-based optical character recognition (OCR) service: https://www.abbyy.com.
How might the end-user assess effectiveness?
Reference is made to security certification under ISO27001 (Information Security Management) (FAQs question 7; archived). It relates to the secure storage of clients’ data in the cloud: https://dellalegal.com/della-obtained-iso-27001-security-certification/ (archived).
With respect to performance in its core task of document review, see accuracy tests mentioned under Q2 (but these do not go into great depth). The FAQ also refers to performance, stating that “We use a number of metrics internally, for example, what percentage of the answers are right on our first suggestion, what percentage of our answers are correct with one of our suggestions and :F1 which is a more sensitive metric.” (FAQs question 8; archived). No details of the latter were available at time of research.
Top What form does it take?
Form
Platform
Details
The system comes with sets of pre-defined questions that a corpus of documents can be tested with, but it also allows for further question ‘checklists’ to be defined by the end-user. The model can also be trained (transfer learning) using the end-user’s documents.
Top Is it in current use?
The company appears to be in a startup phase, and has recently received a round of funding (‘Della raises $2.5M funding to transform legal review’ (Della blog, 24 March 2022) (archived)). The marketing literature refers to it being ‘trusted’ by various parties, and some of the resources refer to concrete case studies where firms have imported their documents into Della.
Some of the marketing literature is very new. The FAQ, for example, is from October 2021.
Top The creators
Created by
Legal tech company
Details
Della was founded in 2020 by Nicolas Chauville (CTO) and Christophe Frèrebeau (CEO), who both appear to have a background in software engineering. The company has headquarters in London and Paris.
Top Jurisdiction
Background of developers
France
Target jurisdiction
“Multijurisdictional”. They also claim to have trained the language model on over 20 languages, including the possibility of assessing documents across different languages (Della Platform; archived; NB the FAQs refer to the “core model” being trained on over 100 languages (FAQs question 12; archived)). There is no explicit mention of particular jurisdictions being targeted; the broad claim seems to be that any legal workflow that fits the product and that uses one of the trained languages will be amenable to use of the application.
Target legal domains
Primarily private law.
Areas where there is a canonical text (often a ‘playbook’) against which to compare a draft document, often an agreement of some kind. They list “contract reviews, repapering projects across departments, data privacy, employment, pensions, M&A” (How we help Law Firms; archived) as their main areas of application.
Top