Signal from the noise
As LLMs play an ever more present role in research workflows, data vendors are rethinking how expert content is sourced, created, structured and delivered to meet clients’ evolving needs.
For years, expert insights reached investors and decision-makers as a document: a transcript, a note, a PDF to read once and file away. That legacy model is giving way to something more demanding. Investment teams now want expert insight to flow into the systems where they screen, model and decide, and they want it structured, traceable and current. The shift is less about content volume, and more about how it can be made more valuable as a dynamic source that sits directly inside a research workflow.
Dan Thomas, Head of Partnerships at Third Bridge, joined Sara Beykpour of Particle and Lucas Ou-Yang of CEOInterviews.ai for a panel on unstructured data. Hosted by Mikheil Shengelia of Eagle Alpha, the discussion covered where research insights are coming from, what makes them usable, and how clients now want to receive data.
We’ve distilled the key insights from Dan here.
How do you gather intelligence at Third Bridge?
At Third Bridge, we produce primary research and data on companies, sectors, supply chains. The largest dataset is the Library of structured interviews with former executives, competitors, customers and suppliers of investable assets. Each interview is recorded, transcribed and enriched so that investors can search and analyze what experts actually say.
Every transcript begins with a real conversation. Third Bridge analysts, or a client, interview a subject matter expert, a company executive, or a former competitor, customer or supplier. Finding, onboarding and retaining the right experts is an ongoing task, and it still depends on people to deliver this. AI has helped the team reach a much larger pool of experts, which has enabled us to produce far more content than before.
How is Third Bridge’s data prepared for Large Language Models (LLMs)?
Third Bridge does the enrichment itself, so the content reaches clients exactly as it was said by the expert. Each expert conversation becomes a structured JSON transcript with unique IDs, a consistent schema, Q&A pairs and speaker IDs that hold across the whole dataset. Companies are mapped to tickers, ISINs, FIGIs or the identifiers a client requests, and GICS sector classifications and semantic chunking are added. As Dan put it: "The content is data verbatim. None of it is summarized or paraphrased."
What are the key challenges you’ve encountered when it comes to making your data set AI ready?
Two challenges stand out. The first is gathering intelligence in the first place, which depends on people as much as technology. Sourcing, vetting, and retaining the right experts is an ongoing task. AI has widened the pool of people we can reach, but it has not removed the need for judgment about who is worth talking to.
The second is making it machine-readable in a way clients can trust and use. Third Bridge started producing interview content in 2013, well before anyone expected large language models (LLMs) to ever read it. Therefore to make the content machine-readable, metadata, tags and labels were added retroactively. As Dan explained: "We've had to retroactively apply a lot of that to our content." That work created point-in-time integrity challenges, which the team has had to resolve record by record. A schema is never finished. Dan described it as a process that keeps going, calling it "a never-ending iterative process." The team works through requirements with each client and shares the schema openly, so that the structure stays consistent as new content types arrive.
How do clients want to receive expert insights data?
Delivering data at scale is now largely a solved problem. Warehouses such as Snowflake and Databricks, along with the hyperscale cloud providers, handle bulk transfer well. Clients now ask for everything from SFTP to a warehouse, a terminal, a newer AI-native platform, or an assistant such as Copilot Studio, Cowork or Codex. Dan framed the real challenge this way: "can you be multi-domain or multi-channel in terms of your delivery?"
Third Bridge meets clients where they are, providing multiple routes to data delivery.
- Client portal. The Third Bridge web application gives analysts direct access to the transcripts and the search tools behind them, with no integration work required.
- Data feeds. Bulk delivery into a client's own systems, through SFTP or into warehouses such as Snowflake and Databricks. Feeds suit firms that run their own research databases and want the full dataset in-house.
- Third-party platforms. Clients can reach the content through platforms such as the Bloomberg Terminal, and through newer AI-native providers including Hebbia or Rogo, so the data sits inside tools they already use.
- MCP connectors. Analysts and agents can query Third Bridge directly through its native MCP connector, so each request returns current content. Agents have their own entitlement paths, separate from human users.
The right channel depends on the workflow. A team that needs a daily snapshot inside its own research system will usually take a data feed. A team that wants to query content as it is published will prefer MCP. Large institutional clients using enterprise LLMs often connect straight to the Third Bridge MCP. Third Bridge's aim is to match the channel to the client's use case, not the reverse.
Where do you see unstructured data adding the most value?
Clients use expert insights to understand the operational realities of an investable asset, such as what a supplier, customer or former executive sees on the ground. Dan explained that these views often surface before the same signals appear in public filings or earnings calls. Many clients compare expert sentiment against management sentiment and sell-side consensus.
Provenance, compliance and trust are of course a considerable part of the value. Dan noted that the metadata around a piece of text, such as who said it and where it came from, tells a client how far to trust the text. Clients increasingly ask for that provenance.
How have LLMs changed discovery?
We see different user groups using LLMs to process expert content in different ways.
Quants used to queue datasets for months before they had time to test them, now they process and ingest large volumes of data faster than ever, and raw, unvectorised content remains the most valuable input for systematic work. For these users, Third Bridge focuses on the integrity and consistency of its schema, so that the ingestion process is as easy as possible.
Fundamental users don’t always have production-ready ingestion pipelines, but they want curated insight on the specific assets they cover. The Third Bridge MCP combined with their LLM of choice provides that out of the box, letting them retrieve the relevant content on demand.
As each group borrows more from each other, the line between systematic and fundamental research is blurring.
What are the legal considerations of sharing data with LLMs?
Broad model training rights worry every data vendor. Extraction risk is high when content is not protected by technical and legal guardrails, and that threatens businesses built on proprietary content. Locking everything behind a web app would of course protect the content, but simultaneously restricts the value that the investing community can obtain from it.
All risks must be carefully balanced to deliver maximum client value, while protecting IP.
How are agentic AI workflows changing things?
Agents are becoming a major consumer of expert content. Clients run agents that check each new publication against their own theses, often in near real time, and citations of Third Bridge content are rising steeply.
Third Bridge has set separate entitlements for agents, since they may need more data in some areas and tighter limits in others. "The first user to engage with our content in the not distant future is highly likely to be an agent that's acting on behalf of a user or a firm to execute a specific research task." noted Dan.
Conclusion
AI workflows are only as powerful as the context and data they are connected to. Making expert insights most impactful takes people who can source trusted and credible voices, a structure that holds up as content grows, delivery that fits how each client works, and licensing that protects the content while still letting clients use it. Agents are raising the bar on all four.
If you are considering how Third Bridge’s primary research and expert content fits into your research process, talk to our team. We can help you identify the route that suits your team's workflow.