Skip to main content

Reviewing contracts using AI

Harnessing machine learning to reduce risk

A legal change required a huge contract estate to be reviewed fast. We combined human and artificial intelligence to complete the task quickly and accurately - even 'teaching' the AI to work in different languages

When a European government passed a law to tackle bribery and corruption in the healthcare sector, we were called in to advise a leading player on its potential exposure. The legislation created two new criminal offences and extended the scope of criminal liabilities to cover the undue provision of benefits to all healthcare professionals.

Our client had 90,000 pages of distribution contracts that needed to be reviewed. Many had ‘evergreen’ provisions, which meant they were automatically renewed each year, and with criminal liabilities potentially attaching to all agreements concluded after the law was passed, it was vital to identify and amend as quickly as possible those that presented a risk.

Combining human and artificial intelligence

The sheer volume of the agreements – and the variation in the data they contained (there were multiple different formats and many contained provisions amended by hand) – would have taken too long to review manually. We chose therefore to employ an innovative combination of human and artificial intelligence.

Our lawyers reviewed the contracts to identify the relevant ‘boilerplate’ and individual provisions that might trigger a breach and then used a cutting-edge software platform to extract them into a tailor-made report for individual review by lawyers.

Teaching algorithms different languages

Over time we taught the system’s machine-learning algorithms (initially configured to work only in English) to identify the potentially risky provisions in the relevant language. We also trained the system to distinguish between 20 types of agreement, and to extract data such as the parties to each contract and its term length and start date. 

The review of the first 4,000 contracts took less than three months. We were able to categorise and prioritise the unorganised dataset very quickly, allocate contracts to reviewers and run real-time analytics to provide the client with a clear picture of its exposure. 

We have applied the same approach to teach the algorithm to work in other languages. We have also worked with the software’s developers, Kira Systems, to use the outputs from the review to deliver what is effectively a contract management system that will significantly reduce our client’s future risk. The company can now, for example, identify all agreements with potentially problematic provisions, organise and distribute them according to remediation needs and prioritise the necessary amendments.