Thomson Reuters has been innovating with artificial intelligence (AI) for a number of years. “We launched our very first natural language question answering and search algorithm back in the 1990s and we have been incorporating AI tech into products like Westlaw [a legal research platform] and different search algorithms,” says the company’s chief product officer, David Wong.
Wong joined Thomson Reuters in the midst of the pandemic in 2020. He works alongside the head of engineering and the chief technology officer, who own the engineering and technical teams, as well as Thomson Reuters’ labs. Apart from the Reuters news business, Wong’s responsibilities cover all of the company’s software and content products. Prior to joining Thomson Reuters, Wong worked at Facebook as a product manager.
The company has adopted a multicloud strategy and uses Microsoft Azure, Amazon Web Services (AWS), Google and Oracle. “We try to be as open as possible and work with different technology vendors,” says Wong.
According to Wong, AI – and more specifically generative AI (GenAI) and natural language processing – has been a core competency within Thomson Reuters.
Recalling the evolution of GenAI, he says: “It is an interesting story. If we play the tape back, GPT 3, if you remember, feels like a million years ago. It was launched in August 2020 and it was ‘interesting’.”
Then GPT 3.5 and ChatGPT were introduced, and then GPT 4. “All came out in quick succession at the end of 2022 and into 2023,” he adds.
For Thomson Reuters, these developments were significant because, as Wong points out, GPT 4 in particular has the ability to interpret language and answer questions at post-graduate level. “It marked a point when we realised that the technology is sufficiently sophisticated that we think it can be pointed at problems we solve for our customers,” he says.
Product development with embedded GenAI
These problems fall into two categories. They either help with information retrieval or research problems. Wong uses Westlaw as an example, which offers a detailed database of proprietary content that is used by legal professionals to do their job right. The product, he says, is used to get an answer to a specific legal question.
“[GPT 4] marked a point when we realised that [GenAI] is sufficiently sophisticated that we think it can be pointed at problems we solve for our customers”
David Wong, Thomson Reuters
“The second problem we solve for our customers is generally the production of written work, such as helping people to write a contract, prepare a tax return or a regulatory filing. GenAI happens to be very good at those two problems.”
Moreover, he says not applying this technology would mean “missing out on a huge opportunity”. There is also the risk that unless Thomson Reuters innovates with GenAI, other companies will produce better products.
Wong sees strong alignments between the GenAI capabilities Thomson Reuters is developing and the GenAI functionality that is being introduced in Microsoft 365 (M365), with its Copilot offerings. “We serve many of the same customers and we both saw the opportunity in the same way. Lawyers, accountants and risk professionals live in Microsoft products. So there was a strong alignment,” he says.
As such, Wong says Thomson Reuters saw an opportunity to integrate with Microsoft 365 Copilot, expanding the plugins and apps it already builds on top of the Microsoft ecosystem: “We were one of the very first companies to get access to M365 Copilot as part of their third-party extensibility programme.”
This enabled Thomson Reuters to develop functionality that allows the M365 Copilot to interact with Thomson Reuters’ own datasets, as well as other AI agents. “We focused on building out integration with Microsoft Word,” he adds.
Intellectual property concerns
One of the risks associated with the use of GenAI is data leakage. Wong says Thomson Reuters has a robust privacy policy, which existed before GenAI.
“We’ve established a clear relationship of how we handle our customers’ data, which applies to the way that we operate with AI. When it comes to generative AI, we’ve tried to take our clients’ confidentiality very seriously and make that a principle in the way that we operate.”
As an example, he adds: “We do not train the AI on our customers’ data. We do not enrich any large language models with our customers’ information unless they explicitly ask us to or they want to undertake some kind of joint collaboration for AI innovation.”
The majority of use cases at Thomson Reuters rely on the AI inference engine, where a large language model (LLM) is accessed via an application programming interface (API). The AI-enhanced products developed by Thomson Reuters takes answers from the LLM. “We are not training or enriching the systems with the data that our customers provide,” he adds.
When working with commercial LLM technology providers, Wong says the company sets out a contractually binding agreement stipulating that the LLM cannot consume any of the input data from Thomson Reuters or data from its customers. The contract also states that Thomson Reuters is able to audit the LLM provider to ensure its intellectual property or data from its customers is not being used for training.
An example of how the integration with M365 Word Copilot works is that it can help a user create a new sales agreement based on a draft agreement, which can be modified using content that Thomson Reuters provides. The user may use Thomson Reuters to find a clause that will be suitable and enforceable in California, for example, which can then be pulled into the sales agreement document.
“When we think about the problems that we applied generative AI to, we’re typically not entrusting the AI to make decisions. We’re really focused on how we provide research and analysis and then automation of a process.”
From an auditing perspective, the AI system is being used for retrieval augmented generation. Wong says: “We’re able to reference all of the source information.”
Augmentation, a reality check
The way Thomson Reuters is working with Microsoft and embedding GenAI in its products is, to use industry terminology, AI augmentation of a labour-intensive task. The AI can provide fast access to domain expertise, which will save people time.
Wong does not believe AI is yet at a stage of development when it exhibits superhuman intelligence. The answers it can provide are not accurate and trustworthy enough yet, so there are caveats to using current technology limiting its use in applications that require very deep analysis. But that does not mean such AI-augmented systems cannot help people complete tasks such as writing a contract document.