Client Story
Bridging Three Systems for One Answer with AI
A research publisher built on Salesforce
Our client covers the people, issues, ideas, and events that move institutional real estate and infrastructure investing. Its core subscriber product is a Salesforce database of contacts, organizations, investment programs, and commitments.
Our team was already supporting the company’s Salesforce implementation, so we were well-situated to help them get more out of that data.
Good data, too many places to look
Subscribers had their work cut out for them. The company’s subscriber database works, but its filters and dropdowns take time to learn and navigate. Additionally, 15 GB of past publications sat in a hard-to-search PDF archive, and news lived on a separate WordPress site. To find recent news for a given investment firm, a subscriber had to click through the database and then read back through older articles. When they had questions, the company’s team had to answer them, creating a communication backlog that built up quickly.
The CEO wanted a natural language interface across all of the company’s data, a use case that took shape in a Rapid Generative AI Assessment that our AWS team ran with the client. They looked at Salesforce Agentforce, which they found very expensive, as well as another vendor. They chose Soliant due to our extensive experience not only with their data and business but also with the various technology options in play.
A team of AI specialists
The AI research assistant works as a team of agents, each with one job and the tools to do it. An orchestrator agent reads each question and sends it to the specialists that can help: one for the Salesforce database, one for the publication archive, and one for WordPress news. It then writes a single response from what they return.
Citations were critical, as subscribers making investment decisions need to see where an answer came from. We also worked with the client’s outside web agency and prompted them to open up API features that were on their roadmap but not yet live.
Tackling data first
Useful and reliable AI solutions depend on good data. Much of the effort in connecting to Salesforce went into the client’s data. An agent that writes database queries is only as accurate as the records it works with, and issues like duplicate records and human typos can make the same answer look different in two places. Our team spent two months tracing mismatches and making corrections before we could ensure the Salesforce answers were trustworthy.
Going live for subscribers
The AI research assistant launched in 2026. A subscriber can now ask, “Are there investment opportunities that focus on multifamily?” and get one answer, with a link to each source. Each question runs under the subscriber’s own login, so the assistant can only surface what that person already has access to. The client agreed to judge success on five measures: accuracy against predefined questions, adoption, response time, export usage, and thumbs up or down ratings.
A partner with the right experience and expertise
As a long-time partner, our team understood the client’s data model long before we built on it. Once plain questions had to become precise queries against imperfect CRM data, it became clear how much that mattered.
We built each agent so it could be tested on its own, which let the team choose the right AI model one agent at a time. The system runs on Amazon Bedrock, so the client can switch models with a configuration change as models improve or prices shift.
And the work isn’t finished yet. We keep uncovering ways for AI to make an impact in this business. Next up is a writer agent that builds answers directly from source passages and adds inline citations.
Under the hood
We built the agents with Strands Agents and run them on Amazon Bedrock AgentCore. The Salesforce agent turns each question into a SOQL (Salesforce Object Query Language) query, using industry terms mapped to the client’s fields, and runs it against Salesforce. The publication and news agents use retrieval-augmented generation: before answering, they pull relevant passages from vector indexes built on Amazon Bedrock Knowledge Bases and Amazon S3 Vectors, which scheduled pipelines keep current. All models run inside AWS, so the client’s data stays in their own environment and is never used for training. Several of these Amazon services were so new during the build that the team spent months working from launch blog posts, rather than official documentation.
Start with the data and systems you already have
Many companies have years of valuable information sitting in systems that were never built to work together. We can help you sort out what your data needs first and where AI will actually pay off in your existing systems. Reach out to our team to learn more about how we can help you make the most of your full data picture.