Move from AI uncertainty to business advantage

Large-language models

GenAI uses large-language models (LLM) to produce content. An LLM is a deep-learning algorithm that can perform various natural-language processing (NLP) tasks. Because they use transformer models — a neural-network architecture that can automatically transform one type of input into another type of output — and are trained using massive datasets, they can recognize, translate, predict, or generate text or other context based on words that commonly go together.

LLMs have become increasingly popular because they have broad applicability for a range of NLP tasks, including:

• Text generation
• Translation
• Content summary
• Rewriting content
• Classification and categorization
• Sentiment analysis
• Conversational AI and chatbots

The numerous advantages that LLMs provide to organizations and users include:

• Extensibility
• Adaptability
• Flexibility
• Performance
• Accuracy

The most widely known and used LLM-based AI is ChatGPT.

The primary obstacles to AI for accounting

Artificial intelligence is exciting, and its impact on the profession is unprecedented. But, AI poses legitimate concerns for firms. One problem involves “hallucinations.” AI needs to be trained on large, complex, and disparate datasets. It’s trained to find words that often go together and present a rational response, but it’s not necessarily trained to return accurate, recent, or precise responses.

For the user, there’s no way of knowing what pieces of data the model based its response on and no easy way to vet the information the models give, which can lead to some severe problems. For example, a lawyer recently used ChatGPT to prepare a court filing, but the AI bot delivered fake cases that the attorney presented in court. The judge fined two lawyers and the firm for the mistake. For accounting professionals who rely on accuracy and preciseness, the results of fake information can be devastating to their firm and career.

Then, there are reasonable concerns about privacy and security. If employees input company data into the LLM, the AI could incorporate it into its learning model. That information could become a part of its knowledge base, and other users could see responses with proprietary data. This situation happened at Samsung, where workers unwittingly leaked top-secret data while using ChatGPT to help them with tasks. The company allowed engineers in its semiconductor branch to use ChatGPT to help fix problems with source code. To do so, the workers input confidential data, such as source code and data relating to its hardware. That information is now part of the data ChatGPT uses to generate content.

Governments are responding to this concern with new regulations. For example, in January 2023, the National Institute of Standards and Technology issued the AI Risk Management Framework (AI RMF) to provide guidance for using, designing, or deploying AI systems. The framework is voluntary, with no penalties for non-compliance, so state regulations take on an essential role in promoting privacy protections.

Given the potential for input information to be used to train an LLM or otherwise exposed to other parties, firms should take care to maintain client confidentiality and refrain from entering client data into LLM tools.

A third concern is that the AI may have an “unconscious bias” that impacts the accuracy or fairness of outcomes generated by the tool. AI models are susceptible to a range of biases arising both from the datasets they are trained on and the way they’re trained. These include availability bias — over-reliance on more prominent information — and confirmation bias — the tendency to disregard evidence that challenges a pre-established position. In legal, tax, and accounting contexts, any inbuilt bias that is allowed to skew results and outcomes has the potential to undermine the advice given to clients.

Thomson Reuters professionals recommend caution when approaching AI. “As a senior accountant, you would not be likely to just hand off some work to a junior colleague and then let it go out the door without looking at it yourself. It’s the same dynamic with AI,” says Des Brady, Senior Director for Inbound Product Marketing with Thomson Reuters. “We recommend it be used as a first draft, used as your assistant and your guide to the content and the answers you need. But we would never say you can just rely on this and send it to your client without even looking at it. We would always recommend a check.”

Mitigating the risks: The TR approach

There’s a way to eliminate these concerns — using a trusted partner for your AI, one with vetted data and proven content for the LLM to learn from. That’s the case with Thomson Reuters.

The foundational capabilities of Thomson Reuters are based on the skillsets of its people, its trusted content, and its rigorous approach to retrieval augmented generation (RAG). Thomson Reuters has been working with AI since 1991 when Chief Scientist Howard Turtle helped found one of its first R&D groups. Since then, the company has paired advanced data scientists with its large team of tax, accounting, and audit subject-matter experts.

Together, these two groups are rigorously training and testing a variety of large-language models with Thomson Reuters content, marrying authoritative insight with the LLM so people can trust the responses they get from Thomson Reuters Checkpoint Edge, a full-text database of federal, state, and international tax laws, regulations, and secondary research sources.

When Checkpoint GenAI launches in 2024, the RAG approach will restrict the LLM to working solely with this expertly created content, with the rigorous testing process reducing inaccuracies arising from bias or other sources.

“When somebody asks a large-language model a question within a Thomson Reuters environment, it’s not going out to pull information from across the web,” Brady says. “It’s pulling information from the strictly delimited areas of content that we have told it to answer a question from.” This ability is where the Thomson Reuters pedigree as a publisher and reputation — built on expert and trusted content — comes into play.

Adds Hawkins, “The quality that our users depend on today — the accuracy, the completeness, the currency — that’s the foundation of why generative AI within Checkpoint can also be trusted.”

Furthermore, as a leading provider of tax and accounting workflow solutions, Thomson Reuters draws on long experience in building information security into the tools it offers to professional customers, eliminating data privacy concerns.

Conclusion

AI is here to stay. Accounting firms that choose to deploy it can deliver significant benefits for themselves and their clients but must be careful and intentional in their approach. The productivity and customer-service benefits are genuine, but so are the concerns about the privacy of information and reliability of results.

Thomson Reuters is developing the tools to help firms use AI effectively. Checkpoint Edge already uses a highly sophisticated AI-powered algorithm to return targeted research results based on full-phrase, natural-language questions. Following the deployment of generative AI in 2024, it will be able not only to return search results but also to summarize findings, prompt follow-ups, and serve up next steps to the user in a single motion, radically re-imagining the tax research experience.

To get more information about the development of generative AI in tax and accounting research, enroll in the Thomson Reuters Checkpoint Edge generative AI engagement program. There, you can see early prototypes, stay regularly informed on progress, and influence the development of GenAI.

Checkpoint Edge

Create unique plans for yourself or larger groups of users with a Checkpoint Edge subscription. Contact us and we’ll help customize a solution that meets your specific needs.

Thomson Reuters

Thomson Reuters leads at the confluence of content and technology, delivering trusted acumen to complex arenas of industry, including legal, tax, accounting, risk, and fraud — all while reaching billions through today’s most global news service, Reuters.