AI Lease Abstraction

The backing of CoStar Group puts us in a unique position to set a new standard for accuracy by investing heavily in the testing, validation, and iteration process. It’s not a one-and-done project – our approach is an ongoing investment in people, processes, and technology that we’re making to ensure we can be highly confident in our outputs – even as the AI model changes and as our prompts evolve. 

All customer data usage is governed by our existing privacy policies and SOC audit requirements as well as our contractual obligations.   

• We will never use your data within any CoStar product without your express written consent. If you do not provide us data rights, we will not use your data.  

• If we have permission to use your data, it is always aggregated and anonymized – and subject to our data anonymity policies to ensure it is never identifiable.  

• We are not using your data to train an AI model. Customer data is never stored or exposed to the public cloud. Instead, our internal teams use customer leases to tweak our prompts, and we have an enterprise agreement in place with AWS Bedrock that allows us to access world-class AI models while isolating customer data with enterprise-grade data security.  

We use customer data to help our internal team tweak our prompts – not to train the model. Customer data is never stored or exposed to the public cloud.  

Instead of exposing data directly, we have an enterprise agreement in place with Amazon Bedrock (AWS Bedrock) a service that allows us to use existing AI models without having to build or host them ourselves – or exposing any customer data. It provides the infrastructure to be able to access world-class LLMs keep customer data isolated and secure.   

By this point, everyone is using ChatGPT (or something like it) and lease abstraction feels like a low-hanging fruit use case. Here’s where doing it yourself might introduce some risk:   

Accuracy – AI prompts need extensive feedback to effectively interpret the nuances of lease agreements, and even if you hone the prompt for one document, every lease is structured differently.  

Data Security – Unless you have an enterprise agreement in place, any data you enter into the LLMs is being used to train their models and available for public consumption. In our approach, your data is never exposed to the AI model, and you’re protected by our security audit

Cost Savings – Lease abstraction costs money, even if you’re only doing it with AI. OCR costs money. AI processing costs money – and we’re able to support it at scale.  

We’re doing much more than just dropping your lease in Chat GPT: 

• Initial OCR Scanning ensures proper document classification to make sure your document is run through the right prompts. You don’t want a UK lease going through a US LLM prompt – or an equipment prompt being applied to a real estate lease. 

• Our second OCR scan processes the entire document, including details like table formatting and image reconstruction – details that provide critical context for something like building out a rent schedule that an LLM is not particularly good at parsing out of a PDF alone. 

• Our proprietary prompts are built by lease abstraction experts. They’re designed to account for the complexities that lead to misclassifications and miscalculations in existing AI lease abstraction solutions like complex structure and terminology, murky responsibility splits and accounting and cashflow ambiguities.  

• Our validations aren’t just basic lease checks to ensure completeness – it’s a sanity check against industry best practices and norms we’ve learned after decades in the leasing space and CoStar Data to flag potential issues even experienced lease abstraction teams might miss without context like rent anomalies, square footage mis-matches, and TI allowances that are far outside industry standard.   

• Perhaps most critically, we have a dedicated team of engineers and lease abstractors focused on ongoing validation and testing to make sure that our prompts continue to produce accurate results as the technology changes and we learn more from the data.