Robot lawyer gets Sberbank role

Russia’s largest state-run lender, Sberbank, has created an artificial intelligence (AI) system that enables it to check a company’s legal capacity without manual intervention.

The “robot lawyer”, as the system is known, is part of the bank’s recent reinvention as a tech company eager to compete with the likes of Apple and Google, as well as local tech giants, such as Yandex.

In the eight months that preceded the unveiling of the software, the robot lawyer issued more than 2.5 million legal opinions. It dramatically speeds up business processes and avoids mistakes that are unavoidable when manually processing large amounts of data and verifying contractors, said Sberbank.

The robot recognises and ingests legal information from various documents, verifying that an entity in question is not bankrupt, liquidated or reorganised. Also, by analysing over 10 million other documents, it checks whether the signatory is entitled to a loan.

Although the main question the robot answers is whether a loan should be issued to a specific entity, its findings go beyond that, also providing the applicant’s key risk metrics, which include risk rationale, risk level and evaluation of how the risk could be removed or minimised.

Sberbank says that the robot, already in active use by the company, could also be in demand across various industries where there is requirement for a customer’s or a contractor’s legal capacity to be verified.

The robot lawyer processes documents sequentially, said Alexander Vedyakhin, first deputy chairman of Sberbank’s executive board.

At the preliminary stage, a package of documents is prepared. To make balanced decisions on credit approval, more than 20 different documents are required from applicants, including general meeting results and an extract from the state registry.

“Each document is translated from scanned copies into a machine-readable format and the text of the document and other significant fragments are extracted, for example, stamps and seals,” said Vedyakhin.

Then a separate tool extracts complex facts from the documents, such as information on voting at shareholders’ meetings and on the company’s CEO. That tool is based on deep learning, such as Bert or CRF. Machine learning algorithms find names, addresses, the number of votes, legal entity name in the text, and connect this information together.

“For example, a robot is able to automatically ‘read’ the proceedings of a shareholder meeting and understand which shareholder owns what part of shares and whether a quorum is reached for each issue put to a vote,” said Vedyakhin.

To develop a system that can read large texts in fluent Russian, Sberbank’s lawyers prepared training datasets, he said. For instance, to train models to extract information from organisations’ charters, 12,000 documents of that type were required.

The robot dramatically cuts the time needed for document processing, from about 30 minutes that a human needs to read a document to less than a minute. The analysis of an entire entity takes the robot seven minutes, on average.

The final stage of the robot lawyer’s work is to form a legal opinion based on the facts extracted from the documents.

“Traditionally, a lawyer manually checks relevant information in a package of documents and makes a decision independently, guided by accepted rules and their own experience,” said Vedyakhin. “The result of this activity is a conclusion, which takes into account all possible legal risks.

“To automate this stage, a ‘digital lawyer’ implements a decision-making system that prepares a legal opinion based on the knowledge of lawyers, laws and internal regulations. For example, the program can verify that a CEO, acting on behalf of the client, is really entitled to obtain a loan from Sberbank.”

To support the decision-making system, Sberbank’s lawyers defined rules in a special subject language, which was specifically developed for the robot lawyer project. This makes the system flexible to changing legislation and internal rules.