A title examiner in a mid-size county used to spend hours stitching together state guides, county rules, and probate-or-tax-ID edge cases before a single property could close. Rocket Close, the Detroit-based title insurance and settlement arm of Rocket Companies, built an AI agent called Supercharger to handle that research. Six months in, the examiner's job looks different, though not gone.
The bottleneck was real and specific. Title operations became a squeeze point in the homebuying process as loan demand grew, because every state, and often every county, has its own rules on how a title has to be examined, recorded, and cleared. An examiner checking a single property might have to read a state guide, look up a county's probate procedure, reconcile a tax-ID discrepancy, and confirm recording requirements, then stitch the answers together before a deal could move. The work is research-heavy, fragmented across systems, and unforgiving of error, since a missed lien or a wrong probate step can delay or kill a closing.
Rocket Close's answer was a custom agent, not a generic chatbot. In an AWS Machine Learning Blog post co-published with the company, the team describes Supercharger as an agentic AI system, a term for software that plans and takes multi-step actions using tools rather than answering a single prompt. Supercharger combines title and closing knowledge, interacts with internal operations teams in natural language, and routes questions to the right sources, including state guides, county rules, and internal documents.
The stack is the interesting part for anyone trying to picture what the agent actually does. Supercharger runs on Strands Agents, AWS's open-source software development kit for building agents. The reasoning model is a large language model served through Amazon Bedrock, AWS's managed model service. The agent reads from Amazon Bedrock Knowledge Bases, which are curated document stores the model can query, and it calls external systems through the Model Context Protocol, a standard that lets an agent invoke tools and data sources. The combination lets the agent answer a question like, "What is the probate procedure in this county, and does this tax-ID mismatch change the recording path?" by pulling from the right knowledge base, calling the right tool, and returning a structured answer, instead of asking a human to do the same research.
The behavior change claims, though, come from a vendor case study, not independent reporting. The AWS Machine Learning Blog post describes the build, the architecture, and the goal: centralize knowledge, automate research-heavy examination tasks, and reduce time spent searching for information. It does not publish independently audited numbers on time saved, error rates, or deal throughput. Any claim that the agent cut research time by a specific percentage, or that it caught a specific class of errors the team used to miss, is vendor-reported. That distinction matters because a title examiner's work is also a legal record, and the gap between a vendor's stated gain and a measurable one is the gap between a marketing post and a regulated outcome.
What the agent does not do is also part of the story. The human examiner still owns the judgment, the signature, and the legal liability on a title decision. The agent is a research assistant that can read thousands of pages of state and county guidance in seconds and surface a probable answer. The examiner still has to decide whether to trust the answer, whether the cited authority applies to this property, and whether to sign off. In a workflow where a wrong call can delay a closing or expose the agency to liability, the human step is not a soft afterthought. It is the part that converts the agent's output into a title opinion.
There are open questions the build does not answer. The audit trail, meaning exactly what the agent read, which tools it called, and which version of which knowledge base it pulled from, is not described in detail in the public post. The ongoing build-and-maintenance cost of a custom agent in a regulated workflow is also not disclosed, and it is a real number any title agency considering a similar build would want to know. The question of replication is open as well. A mid-size title agency without an AWS partnership, without an in-house machine-learning team, and without a curated knowledge base of state and county rules would be starting from a different place, and the build Rocket Close describes is not obviously portable.
The honest frame, six months in, is a tradeoff. Rocket Close bought time for its examiners, centralized a body of fragmented knowledge, and put a multi-step research assistant into a workflow that used to depend on an individual staffer's ability to find the right rule. It also took on a custom build, a vendor dependency, and the obligation to maintain, audit, and defend the agent's output in a regulated setting. The agent's value is not that it replaced the examiner. It is that it changed what the examiner spends their day on, from searching to deciding.