Expertise and accuracy

As an affiliate of Bloomberg L.P., we have access to deep professional experience in the field of AI. Bloomberg Industry Group utilizes hundreds of AI researchers and engineers who are constantly exploring state-of-the-art AI tools and technologies to determine how to apply them in the tax domain.

In addition to building and deploying AI models in the company’s solutions, these researchers and engineers also publish multiple research papers each year, putting us at the forefront of AI research and the deployment of AI solutions. As a result of this work, and with the emergence and proliferation of generative AI, we are uniquely positioned to understand both its power and limitations, so we can build safeguards that foster accuracy in our AI models.

In fact, many of our engineers have worked on Bloomberg Tax for over a decade, developing deep domain expertise. Their understanding of the tax domain enables them to understand our customers’ problems and apply the appropriate technology to solving them.

For our customers, two realities are often in conflict: accuracy is essential, but much of their decision-making is based on interpretations of nuanced laws, regulations, codes, opinions, and more.

Our approach to ensuring accuracy in our generative AI tools is equally thorough. We know generative AI can be inconsistent – hallucinations, inaccuracies, and lack of context are all things we work to avoid and correct in our generative AI offerings.

We utilize the industry-standard retrieval-augmented generation (RAG) framework, as well as our proprietary guardrail service that leverages our foundational technology, deep legal domain expertise, and content to ensure an AI-generated content is grounded in our original content or primary source materials, not the internet at large.

We also have an extensive benchmarking process and system. With the help of tax professionals, we evaluate the accuracy of our generative AI at all phases – from development and early internal testing to the beta environment and customer testing, and onward throughout the life of the product. With these guardrails in place, we’re able to continuously tune and fix the model to prevent it from drifting over time or answers becoming less accurate.

Customer-centric approach

It’s in our DNA to be obsessed with providing our customers with cutting-edge technology built on a foundation of trusted news and analysis that makes them more efficient, more effective, and more knowledgeable about the issues that matter most.

When we build AI-powered tools, our primary focus is whether they will actually solve our customers’ challenges. We understand that adopting new technologies and new workflows can be disruptive, costly, and time consuming. Our mission is to reduce friction and create a natural, incremental transition to using AI.

Rather than developing technology that will upend a customer’s entire workflow or create the end work product, we have taken a building-block approach that focuses on addressing one specific task at a time and seamlessly integrating into existing processes. Collectively, these efficiencies add up to a user-empowered workstream with a low barrier to entry.

We want our AI to make your job easier – not do it for you.

One example of this is Bloomberg Tax’s AI Lab, where we develop new AI tools through partnerships with our customers. This allows real users to provide feedback and have direct input on what challenges we address, which solutions we build and how they function, and the degree to which they interact with AI in our product.

When interacting with our generative AI experiments, customers can provide immediate feedback to help us fine tune our models, ensure accuracy, and fit within their desired workflow.

As always, Bloomberg Tax will continue to collaborate with our customers and empower them to drive the direction of our product development to best serve their needs.

Transparency you can trust

At Bloomberg Tax, we believe that transparency is the key ingredient to building trust in our AI.

We are committed to providing our customers with clear documentation of:

• Where and when our products employ generative AI
• The sources used by our AI
• The benchmarking and validation process for our AI

Throughout our history of employing AI in our products, we have never left a user guessing about where information is coming from. For example, our tools that have traditional used machine learning to suggest related content will indicate why that content is being suggested. Likewise, when generative AI provides an answer or suggests text, it includes links to source material and/or an explanation of why it returned that result.

While we understand the power of generative AI, we also recognize its limitations. AI-generated responses are never 100% trustworthy and pose the risk of hallucinations. We do our best to mitigate these risks by beta testing all generative AI tools before releasing them. In addition, our AI doesn’t generate answers on its own – it also looks through our large collection of carefully vetted information to enhance the accuracy and reliability of the responses it provides. In fact, Bloomberg Tax Answers will abstain from answering when there is no available content covering the topic in question or when the search query is outside the tax scope.

Our brains and expertise are still needed as part of the equation, so we advise all our customers to validate the responses with primary sources and to use their best judgment.