Third Bridge discusses rebuilding expert research libraries for AI agents via MCP and knowledge graphs
Summary
Dan Thomas of Third Bridge explains the firm's shift from brokered services to a content-forward library optimized for AI agents using the Model Context Protocol (MCP). He details the technical overhaul required to make legacy transcripts agent-ready, including chunk-level metadata and inline citations to reduce hallucinations. Speakers discuss how AI is shifting the bottleneck to expert supply, with voice agents handling high-volume channel checks while senior experts remain reserved for high-touch interactions. The episode highlights a convergence of fundamental and quant research, where knowledge graphs map ecosystem risks and portfolio managers directly access insights to bypass junior analysts.
Episode sections
- 0:00 – 2:40
Intro and Dan Thomas Background
Brett introduces Dan Thomas, Head of Partnerships at Third Bridge, noting the firm's activity in the MCP ecosystem. Dan outlines his 10-year tenure, starting as a content analyst covering European high yield and distressed credit, including niche sectors like nickel smelters and ferry companies, before moving into product strategy and partnerships.
- 2:40 – 5:02
Third Bridge's Shift to Content and Library Scale
Dan explains Third Bridge's founding in 2007 as a European alternative to GLG and its evolution from a brokered services model to a content-forward library. He notes the library contains hundreds of thousands of conversations, with tens of thousands added annually, and mentions early adoption of publishing transcripts starting around 2012-2013.
- 5:02 – 7:17
Content Generation Mix and Client Contributions
Dan discusses the mix of internally directed content versus client-contributed transcripts. He highlights that while internal analysts ensure unique coverage, client-contributed models allow for cost-effective access to more experts, a trend that has grown significantly since its introduction.
- 7:17 – 11:54
Early Adoption of MCP for Agentic Distribution
Dan details Third Bridge's strategic decision to embrace the Model Context Protocol (MCP) early, recognizing that the principal consumer of data is shifting to agentic applications. He argues that vendors must optimize for agent consumption via robust schemas and tool descriptions rather than focusing solely on human front-end experiences.
- 11:54 – 16:54
Rebuilding Transcript Libraries for Agents
Dan describes the technical overhaul required to make legacy transcripts (dating back to 2011) agent-ready. This involved upgrading metadata schemas, chunking documents by speaker pairs and semantic meaning, and providing claim-level metadata to improve retrieval accuracy and efficiency for AI models.
- 16:54 – 19:14
Handling Nuance and Lossiness in Transcripts
The speakers discuss how LLMs now better capture nuances like hedging and equivocation that were previously lost in traditional NLP. Dan notes this allows systematic investors to process expert nuance at scale, turning a previously 'lossy' data format into a structured factor for analysis.
- 19:14 – 22:17
Citations, Trust, and Hallucination Mitigation
Dan emphasizes Third Bridge's commitment to inline citations linking to vetted experts, which builds trust and reduces hallucination concerns. He observes that as users become more comfortable with data sources, click-through rates on citations decrease, indicating growing trust in the integrated workflow.
- 22:17 – 31:10
MCP vs. Raw Data Feeds for Backtesting
Dan addresses the challenge of managing multiple MCP connectors versus a single data pipe. He notes that while large firms prefer raw data feeds for rigorous backtesting to control embeddings, smaller firms and those building proofs of concept (POCs) are adopting MCP for faster deployment and lower resource requirements.
- 31:10 – 32:43
Scaling Expert Networks with Voice Agents
Speaker 1 notes that vendors are using AI for high-volume, consistent channel checks to generate signals over time. However, he remains skeptical that agentic calls will fully displace high-touch, senior-level expert engagements, as many experts prefer direct interaction and will not engage with agents.
- 32:43 – 35:53
Supply-Side Bottlenecks and Expert Flexibility
The speaker argues that AI shifts the bottleneck to the supply side, as experts may become overwhelmed by unsolicited requests from various channels. Firms must improve the lives of contributing experts by offering flexibility, such as asynchronous voice agent interactions, to maintain relationships and capacity.
- 35:53 – 39:36
Voice Agent Maturity and Training Data
Voice agents are now in production and no longer a curiosity, though they will not replace all peer-to-peer interactions. Third Bridge's advantage lies in its established library of interview transcripts and deep bench of sector analysts who can evaluate and fine-tune these models against a 'golden set' of responses.
- 39:36 – 43:10
Shift to Portfolio Managers and Risk Personas
Interest is growing among Portfolio Managers (PMs) and risk professionals who use semantic retrieval to map expert insights directly to portfolio holdings. This allows for programmatic alerts on red flags or interdependencies, bypassing the traditional 'game of telephone' through junior analysts.
- 43:10 – 45:05
Convergence of Fundamental and Quant Research
Speaker 1 predicts a convergence between discretionary fundamental research and quant research, as AI makes fundamental resources useful at scale for systematic decision-making. Differentiation will come from generating profound, differentiated insights on specific companies or ecosystems rather than accessing commoditized data.
- 45:05 – 51:44
Ecosystem Modeling and Knowledge Graphs
Fundamental investing is shifting from modeling individual companies to mapping entire ecosystems, leveraging expert insights from suppliers, competitors, and customers. Third Bridge is using knowledge graphs to link claims across content, enabling systematic flagging of issues in branches of a supply chain that affect the core thesis.
Claims worth verifying
Third Bridge was founded in 2007 by two Bain & Co. consultants.
Unidentified speaker (speaker_1) ·2:40
How to verify: Third Bridge corporate history on their official website or SEC filings if publicly traded (verify via company investor relations page). For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Primary sourceThird Bridge has hundreds of thousands of conversations in its library.
Unidentified speaker (UNKNOWN) ·4:42
How to verify: Third Bridge product documentation or sales materials describing library size. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Third Bridge added tens of thousands of transcripts every year.
Unidentified speaker (UNKNOWN) ·4:42
How to verify: Third Bridge annual reports or press releases detailing content growth. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Third Bridge started publishing transcripts to a library in 2012-2013.
Unidentified speaker (UNKNOWN) ·3:38
How to verify: Third Bridge historical product announcements or press releases from 2026-2026. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Dan Thomas has been with Third Bridge for 10 years.
Unidentified speaker (speaker_1) ·1:21
How to verify: Dan Thomas's LinkedIn profile or Third Bridge press releases announcing his role. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Third Bridge was founded as a European alternative to GLG.
Unidentified speaker (speaker_1) ·2:40
How to verify: Third Bridge 'About Us' page or company history section. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
MCP acronym started bubbling up in the middle of last year (relative to 2026 release, implying mid-2025).
Unidentified speaker (speaker_1) ·8:29
How to verify: Public announcements or documentation of MCP (Model Context Protocol) release dates. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Third Bridge started publishing content in 2011-2012.
Unidentified speaker (speaker_1) ·12:25
How to verify: Third Bridge historical product announcements or press releases from 2011-2012. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Third Bridge has offices all over the world.
Unidentified speaker (speaker_1) ·3:18
How to verify: Third Bridge office locations listed on their official website. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Third Bridge has as much business in Europe as in the US and Asia.
Unidentified speaker (speaker_1) ·3:18
How to verify: Third Bridge financial disclosures or press releases detailing regional revenue splits. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Vendors in the expert network space are using AI for channel check type products to generate high-end, consistent signals over time.
Unidentified speaker (speaker_1) ·31:10
How to verify: Review vendor press releases or case studies from Third Bridge and competitors detailing AI-driven channel check products; verify via company websites or investor presentations. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Voice agents are now in production across Third Bridge and its peers, and are established components of many clients' processes.
Unidentified speaker (UNKNOWN) ·35:53
How to verify: Check Third Bridge's official website for product updates or press releases announcing the launch of voice agent capabilities; verify via company filings or news archives. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Primary sourceThird Bridge owns all IP for its interview transcripts and uses them to develop and fine-tune AI models.
Unidentified speaker (speaker_1) ·37:49
How to verify: Review Third Bridge's terms of service or intellectual property disclosures in any available public filings or legal documents; verify via company legal pages. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Primary sourceThird Bridge has a deep bench of sector analysts on payroll who evaluate AI responses against a 'golden set' of expected responses.
Unidentified speaker (speaker_1) ·38:09
How to verify: Check Third Bridge's career pages or organizational structure descriptions; verify via company website or LinkedIn profiles of employees. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Third Bridge has been conducting interviews systematically for over a decade with a specific playbook.
Unidentified speaker (speaker_1) ·38:53
How to verify: Review Third Bridge's corporate history or 'About Us' page for founding dates and methodology descriptions; verify via company website. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Portfolio Managers are using semantic retrieval to map expert insights directly to portfolio holdings for programmatic alerts.
Unidentified speaker (speaker_1) ·40:21
How to verify: Review case studies or white papers from Third Bridge or portfolio management software providers detailing AI integration in PM workflows; verify via industry reports. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
AI has made fundamental resources useful at scale for systematic decision-making, leading to a convergence between fundamental and quant research.
Unidentified speaker (UNKNOWN) ·43:38
How to verify: Review academic or industry research on AI in finance; verify via publications from financial technology firms or hedge funds discussing AI adoption. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Primary sourceExpert libraries may become 'beta' (commoditized), increasing demand for bespoke, proprietary research with senior experts.
Unidentified speaker (UNKNOWN) ·47:43
How to verify: Monitor trends in expert network pricing and product offerings; verify via industry reports from McKinsey, BCG, or similar consulting firms on alternative data trends. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Mosaic theory holds as expert calls rarely provide direct alpha signals but deepen comprehension over time.
Unidentified speaker (speaker_1) ·49:38
How to verify: Review investment literature on mosaic theory; verify via CFA Institute materials or academic papers on fundamental analysis. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Claude and ChatGPT have made significant progress in Excel modeling and three-statement models.
Unidentified speaker (speaker_1) ·46:39
How to verify: Review product updates from OpenAI and Anthropic; verify via their official blogs or documentation on AI capabilities. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.
Do your own research
- Verify Third Bridge's current library size and growth rate against their official website or investor materials.
- Check the technical documentation for the Model Context Protocol (MCP) to understand schema requirements for agent integration.
- Review Third Bridge's public filings or press releases regarding their transition from brokered services to content-forward models.
- Assess the reliability of AI-generated insights by comparing them against original transcript sources for hallucination rates.
- Investigate expert supply constraints by reviewing industry reports on analyst availability and engagement trends.
- Look for case studies or whitepapers from Third Bridge demonstrating the efficacy of knowledge graphs in supply chain risk mapping.
Full transcriptexpand
- 0:00
- Unidentified speaker (speaker_0): Thurbridge has been a name that's been coming up over and over, just very active in the MCP ecosystem. There's hundreds of thousands of conversations in there, tens of thousands of transfers being added every year. What has been your experience with voice models and AI-driven expert network calls thus far? Voice agents, they're not a curiosity anymore. They are
- 0:19
- Unidentified speaker (speaker_0): in production across a number of our peers, ourselves included. One of the big challenges of building a very effective digital researcher is just finding the right context.
- 0:40
- Unidentified speaker (UNKNOWN): All
- 0:40
- Unidentified speaker (speaker_0): right, great. Thank you, and welcome to the next episode of Invest with AI, where we explore For the intersection of fundamental investing and artificial intelligence. I'm really excited to have Dan Thomas here. who is head of partnerships at Third Bridge. Thirdbridge has been a name that's been coming up over and over, just very active on the partnership space in the
- 0:59
- Unidentified speaker (speaker_0): MCP ecosystem.
- 1:01
- Unidentified speaker (UNKNOWN): And
- 1:01
- Unidentified speaker (speaker_0): if you talked in other pods, one of the big challenges of building a very effective digital researcher is just finding the right context that you can uh pump into the system. So thanks, uh thanks for being with us, Dan. Uh maybe just give us uh to start a little bit of your background and tell us a little a little bit about what Thur Third Bridge does.
- 1:21
- Unidentified speaker (speaker_1): Yeah. Um Brett, good to good to speak again. Thanks for the uh the invite. Um my name's Dan Thomas. As as Brett said, I'm head of partnerships at Thurbridge currently, but um have worn many hats at the firm over the 10 years that I've I've been here. So I actually I actually started out
- 1:39
- Unidentified speaker (speaker_1): as one of the kind of the content analysts producing interviews. That was 10 years ago. Now I started on high yield and distressed credit in Europe. So some some pretty funky, kind of hairy issuers from like uh nickel and zinc smelting to like Mediterranean ferry companies and and and what have you. Um over my time
- 1:59
- Unidentified speaker (speaker_1): pivoted more into a commercial role, um, then spent a number of years as a head of head of product strategy, and then pivoted into partnerships as those became a much more important part of our overall business and strategy. So I guess in the the 10 years that I've been with the firm, I've had a decent amount of perspective on how the content gets
- 2:19
- Unidentified speaker (speaker_1): produced
- 2:20
- Unidentified speaker (speaker_1): from speaking directly with the with the experts themselves, produced and published over a thousand interviews with with my name and then moved into obviously product strategy. So like to think I have a decent uh sense of of how the the content and the data gets gets packaged. Um and obviously increasingly in my in my in my partnerships role now, um, a pretty unique
- 2:40
- Unidentified speaker (speaker_1): vantage point in terms of how how that content um ultimately gets distributed and makes its way to to the end user. Um in terms of what Third Thurbridge, Bridge, so a bit about Thurbridge. Third Bridge for those that obviously aren't familiar with us. Um we were founded in 2007 by two Bain and co-consultants. Um
- 2:58
- Unidentified speaker (speaker_1): they they you know they took a look at the GLG you know business, more more North American focused at the time and and felt, you know, why don't we sort of build a uh a European alternative? Um and you know, it turns out it was a great idea and had a tremendous amount of success. We're now you know twenty twenty years old, fully global in terms of our coverage. We have just as much
- 3:18
- Unidentified speaker (speaker_1): business in Europe as we do in the US as we do in Asia. Um we have offices all over the world. But what was um you know, I think the most interesting transition for for for us and for me over the 10 years is how much the business has changed from a uh you know, fundamentally services-oriented business, where the core bread and butter of of what
- 3:38
- Unidentified speaker (UNKNOWN): uh
- 3:38
- Unidentified speaker (speaker_1): um I guess what Thurbridge was was about was facilitating these like broker style interactions between a client and an expert in you know in in in private on a on a on a on a call. The models shifted now where we're much more kind of content focused and content forward. Um because you know there are now obviously
- 3:58
- Unidentified speaker (speaker_1): as we probably will talk about the systems that can take in and ingest and make sense of of large amounts of content. Um but also we were just seeing a demand for for clients to to to to get that information much more quickly than would otherwise be allowed through those those kind of brokered conversations where you know there was scheduling, there were you know calendar
- 4:18
- Unidentified speaker (speaker_1): conflicts, you get on the call, there's you know, all all all of this kind of stuff. And and so we actually were, you know, one of the the first to to to really well these the first to to start publishing these transcripts to to to a library back in twenty twelve, twenty thirteen. Um and and I joined the firm twenty sixteen, twenty seventeen, just as we were really kind of starting
- 4:38
- Unidentified speaker (speaker_1): to to to lean into that. Um and now transcript library, you
- 4:42
- Unidentified speaker (UNKNOWN): know
- 4:42
- Unidentified speaker (speaker_1): there's hundreds of thousands of conver conversations in there. Um, you know, tens of thousands of of transcripts being added every year. Um and that that number kind of continues to scale now that we're also looking at ways to produce additional content with voice agents and AI interviewers that you know that Thurbridge is is is directing.
- 5:02
- Unidentified speaker (speaker_0): What's what's the mix today of client-generated transcripts versus it sounds like a Third Bridge one, of the uh many of the transcripts are generated by internal in-house. interviewers. What's the mix of mix of uh generation?
- 5:18
- Unidentified speaker (speaker_2): Yeah, that's I mean that's right. And and I yeah, not going into um
- 5:23
- Unidentified speaker (UNKNOWN): a
- 5:23
- Unidentified speaker (speaker_1): ton of detail on the on the on the splits. We we started as a kind of curated content business. Uh so that content was being directed by internal third bridge analysts and sector specialists in their own right, many you know, former buy side, former sell side, storied current. careers investing in their own right.
- 5:43
- Unidentified speaker (speaker_1): They lead, define, kind of dictate what's what gets covered, which gives
- 5:49
- Unidentified speaker (speaker_1): the library a unique bent in the sense that you don't just get kind of deno flavor of the month cross-section of the of the index that everybody else is kind of investing in doing work on.
- 6:01
- Unidentified speaker (UNKNOWN): You
- 6:01
- Unidentified speaker (speaker_1): know, we can ensure there's a the sort of a regular cadence of coverage through the analyst team. Um, but we also, you know, we we do the um the client-contributed model as well. That was a newer um a newer thing for for us. There were other people in the space who um kind of proved that that um
- 6:18
- Unidentified speaker (speaker_1): that model. And I think the the the rest of the industry at the time was a little bit skeptical as to whether or not that would kind of catch on, whether people would be sort of willing to to to have those conversations, albeit in a anonymized uh format published for the benefit of of of their competitors to read. Um but that's also now a you know a very meaningful portion of our of
- 6:38
- Unidentified speaker (speaker_1): our library. Um and we would we would expect that to maybe grow as you know folks are always looking to find novel ways, I guess, to um
- 6:49
- Unidentified speaker (UNKNOWN): to
- 6:49
- Unidentified speaker (speaker_1): speak to to more experts. more cost effectively. Um and and actually that model is a very is a very good way of of of of doing that.
- 6:57
- Unidentified speaker (speaker_0): Yeah. Yeah. Yeah. Some of the bills that I ran up at some of my large hedge fund seats in the past and then trying to do that in a more uh sort of lower resource hedge fund wrapper. It's like, yeah, maybe that's not a not a bad idea to participate in some of these uh some of these arrangements. One thing that's notable about
- 7:17
- Unidentified speaker (speaker_0): Third Bridge sort of first was introduced to me you know a number of months ago by a friend um is just the early embrace of the MCP structure when a number of other uh uh players on the expert networks uh side have been And more resistant to that integration can you just uh talk to me as head of
- 7:37
- Unidentified speaker (speaker_0): partnerships about just the philosophy
- 7:39
- Unidentified speaker (speaker_0): of wanting to be present in the agent path via the MCP structure?
- 7:45
- Unidentified speaker (speaker_2): Yeah, I d I think there are fundamentally sort of three three
- 7:49
- Unidentified speaker (speaker_1): three key strategic kind of implications that vendors really need to to wrestle with um today. One of those is obviously having a proprietary prioritary information advantage. Like is is the content actually differentiated? The other is is do you you know maintain an organizational advantage in terms of how you how you package and and and and how you make that that
- 8:09
- Unidentified speaker (speaker_1): content available. But I think that the the one that you know you're alluding to is a distribution advantage um because one of the bigger challenges obviously that the models have in terms of of overall accuracy and usefulness is actually the availability of the the right context at the right time. And so we move moved very, very early um as far as you know MCP being available, it was
- 8:29
- Unidentified speaker (speaker_1): middle of of last last year that the acronym started to even sort of bubble up, really. And that's when we were talking and planning over the summer how do we do this and how do we bring it to market as quickly as we as we can. Because
- 8:43
- Unidentified speaker (speaker_1): it was very, very clear to us that engagement and demand was moving to these sort of frontier models, whether they are packaged through the Claude app or through the Chat GPT app, or whether that LOM is embedded in an AI specific or a vertical specific AI application, whether that's like a
- 9:03
- Unidentified speaker (speaker_1): Hebier, a Rogo, a Model ML, it was very clear to us that we needed to not only design Um
- 9:10
- Unidentified speaker (speaker_1): a backend and distribution that was suitable for human consumption. it was very clear that we also needed to make architectural decisions that were optimized for human, human with agent in the loop and kind of a fully agentic use case as as well. Because I think the data
- 9:30
- Unidentified speaker (UNKNOWN): on
- 9:31
- Unidentified speaker (speaker_1): the client side has does have its own gravitational pull to to to an extent. And I think one of the
- 9:37
- Unidentified speaker (speaker_1): being dogmatic about owning the front end and the user experience can put you in a weird spot as a as a vendor, potentially. It can take you down some some some weird weird areas that that don't necessarily play to you play to your strengths. Because I think one of the things that we lose sight of is I in the near future, not even the too
- 9:57
- Unidentified speaker (speaker_1): distant future, the the principal consumer of, of the data is going to be an application. It's going to be agentic, and they're going to be fairly unsentimental around what the front end actually looks like. It's, oh, how rich are your tool descriptions? How robust is your schema? And how yeah, how accessible is that
- 10:16
- Unidentified speaker (speaker_1): endpoint? Can the agent actually grab it? Is it entitled to? Have you licensed it Sit in the right channels for it to be available, useful, and relevant?
- 10:24
- Unidentified speaker (UNKNOWN): And
- 10:25
- Unidentified speaker (speaker_1): so
- 10:25
- Unidentified speaker (speaker_2): I think by by
- 10:26
- Unidentified speaker (speaker_1): really focusing on like where the demand was was was moving, we were able to launch that product into the market earlier than
- 10:37
- Unidentified speaker (speaker_1): our peers. Um, just because I think we also benefit from the mindset of being like a content first, led expert network. Um and so trying to put that content asset in front of as many people as possible. through um as many channels as possible to serve as like the top of funnel, which you can then obviously start
- 10:57
- Unidentified speaker (speaker_1): to to provide your services through as as well over time, was the you know, was was the play that made the most most sense to us.
- 11:04
- Unidentified speaker (speaker_3): It's uh it's it's interesting and I I tot totally agree about the unsentimentality of uh the front end. Um I
- 11:11
- Unidentified speaker (speaker_0): don't know if it's that's a word. Um
- 11:14
- Unidentified speaker (UNKNOWN): one
- 11:14
- Unidentified speaker (speaker_0): of the one of the things that's been interesting to watch is just the the evolution of context engineering around pulling in that the data, the unstructured data. I felt like 12 to 18 months ago the the bearer people were hitting it was still OCR and hey you can't drop in more than five PDFs into Chat GPT or lose
- 11:34
- Unidentified speaker (speaker_0): attention. Um And now you know there are tens of thousands of transcripts, you know, running through an MCP structure. Can you just talk us about what has had to happen on the extraction level, sort of in your systems, what's happened in the ecosystem broadly to allow an agent now to pull the
- 11:54
- Unidentified speaker (speaker_0): right transcript at the right time at inference?
- 11:58
- Unidentified speaker (speaker_1): I think I think it's a it's a a very important question. If you'd ask
- 12:02
- Unidentified speaker (UNKNOWN): me
- 12:02
- Unidentified speaker (speaker_1): asked a few years ago
- 12:05
- Unidentified speaker (speaker_1): whether or not we needed to invest heavily in in this as a capability, i.e. an enrichment pipeline, um would I have given you the same answer as I gave you today? I don't know. Um it was very clear that the context windows were expanding rapidly. And yeah, we I think the important context to remember is we we started publishing
- 12:25
- Unidentified speaker (speaker_1): content in t two thousand. 2011, 2012, and we weren't structuring our data, our metadata, our tags, our labeling with a view to it being consumed by uh an agent by MCP. No, nobody was nobody was doing that. So we had to we had to look at our our existing schema and how we needed to kind of upgrade and evolve that That
- 12:45
- Unidentified speaker (speaker_1): and for a kind of a truly uh agentic agentic world. And we made an a number of kind of, I guess, key strategic decisions there in terms of not only providing a metadata at the document level, but obviously starting to distill each document down into its chunks. So focusing on speaker pairs and the
- 13:05
- Unidentified speaker (speaker_1): semantic meaning of passages, um and providing then the the kind of the metadata at the the chunk and the claim level as opposed to the document level and realize realizing that you know our transcript library is it's not uh uh a repo of of transcripts that stand alone as as individual documents, but actually
- 13:25
- Unidentified speaker (speaker_1): each one is a composite of you know thousands of different observations and claims. And each of those need to be accounted for in terms of your enrichment pipeline because you know, even though the context window that you models have now is expanding and probably infinite at some point, right? You'll be able to run it over infinite
- 13:45
- Unidentified speaker (speaker_1): context and have it always always running. It's not really the data vendors or shouldn't be the data vendors' position to use up our clients' tokens, w willy-nilly, even if they are getting cheaper and cheaper and cheaper. There are still things that vendors can do even though the capability of
- 14:04
- Unidentified speaker (speaker_1): models is coming on all the time. Context is always always growing. There's there's a lot of literature on the subject, but
- 14:14
- Unidentified speaker (speaker_1): there are specific things that vendors control and
- 14:19
- Unidentified speaker (speaker_1): have the ability to influence. And we've we we touched on the data availability question, like is it available in the applications that our clients are using? But I think is it structured well? Um, you know, are relevant facts um identified chunked appropriately? Is the metadata attached? Um, are your schemas robust? Um,
- 14:39
- Unidentified speaker (speaker_1): or you know, do they have overlapping claims? Do they have overlapping chunks? Do you create collisions c collisions that way? Is the is the labeling and the the taxonomy is is is that robust? Um, you know, companies, topics of concepts, are you using stable, well-recognized identifiers, icins,
- 14:59
- Unidentified speaker (speaker_1): figgies, tickers? Are they consistently? Are you are your point-in-time tags robust? Um, because all of those things demonstrably have been proven to improve not only kind of the retrieval accuracy, um, but also the speed and also the efficiency. Um and I think you in in in our world now, we would hold ourselves up as a,
- 15:19
- Unidentified speaker (speaker_1): I guess, a a reasonably good provider in in in the space. That's increasingly becoming a buying decision that our clients are including because everybody is everybody's very excited about the capability and the capacity of models now to to automate certain parts of the investment research workflow. But obviously, as you know, people are
- 15:39
- Unidentified speaker (speaker_1): increasingly concerned with how much this costs.
- 15:42
- Unidentified speaker (speaker_1): And if you know as a vendor, you need to be able to say to your clients and to your prospects that we've taken these concrete concrete steps to um uh you know to make sure that using our MCP endpoint is as efficient as as r and as reliable as possible. And these are the concrete steps that we've taken. And So being able to show them that.
- 16:03
- Unidentified speaker (speaker_1): and being able to have that conversation is I think incredibly powerful. And part of the reason why we can have that conversation is because we also know what bad looks like. Because we we had bad. We just jammed all the PDFs into the the the context window. And you know the the evaluations that we've done And internally show that the kind of the
- 16:23
- Unidentified speaker (speaker_1): changes that we've made in our enrichment pipeline have led to those demand demonstrable improvements in the overall kind of retrieval of the content.
- 16:31
- Unidentified speaker (speaker_3): Yeah. Yeah,
- 16:32
- Unidentified speaker (UNKNOWN): that makes
- 16:33
- Unidentified speaker (speaker_3): sense. Um
- 16:34
- Unidentified speaker (speaker_0): how do you think about the the nuance of language? I mean, a lot of times when having an an actual conversation, it's the pause, it's the you know so socratic question asked back that doesn't really show up in the transcript, but when you're sort of conducting a call live, can be a tell. How
- 16:54
- Unidentified speaker (speaker_0): how do you think about the just the lossiness of the transcript structure relative to actually conducting the call live? And how do you think about putting that into metadata in the transcript library?
- 17:05
- Unidentified speaker (speaker_2): It's a good
- 17:06
- Unidentified speaker (speaker_1): it's a good question. Um and
- 17:07
- Unidentified speaker (UNKNOWN): I think I
- 17:08
- Unidentified speaker (speaker_2): think a
- 17:08
- Unidentified speaker (UNKNOWN): lot of
- 17:09
- Unidentified speaker (speaker_1): the um
- 17:10
- Unidentified speaker (UNKNOWN): a
- 17:10
- Unidentified speaker (speaker_1): lot of the value that's historically been in in in expo expert content has been precisely that that nuance. Um that kind of the the hedging, kind of the equivocation, the the umming and the ring. And that's that's where um I guess the more traditional NLP pipelines have have been ineffective
- 17:30
- Unidentified speaker (speaker_1): at
- 17:31
- Unidentified speaker (speaker_1): really understanding that context and and extracting the semantic meaning from what's been said. LLMs now are are much better. at
- 17:40
- Unidentified speaker (speaker_1): identifying hedging, comparing it to you know previous passages, immediately following passages, and and and identifying that, or at least flagging that as something that can be investigated by the how the human that's overseeing the the the whole process. And I think that's precisely the reason why we've seen a lot more interest in
- 18:00
- Unidentified speaker (speaker_1): this kind of content from a the the the the systematic community, because LLMs now,
- 18:07
- Unidentified speaker (speaker_1): you know, they
- 18:09
- Unidentified speaker (speaker_1): you know they can pull out that lossiness and make sense of that lossiness now in a much more effective way than they've previously been able to. And so, you know, taking these um these expert transcript libraries and actually looking at that as a as a as a factor in its own right, because you can process the whole thing at scale. You can control
- 18:29
- Unidentified speaker (speaker_1): for for more of that nuance than you've previously been able to do. So when you're relying on like
- 18:34
- Unidentified speaker (UNKNOWN): you
- 18:34
- Unidentified speaker (speaker_1): know like early like BERT stuff and and and the more conventional NLP pipelines means that
- 18:42
- Unidentified speaker (speaker_1): precisely the reason your content had value in the first place, because it was nuanced. Um, you know, that can now be you know extracted at at scale thanks to Thanks to LLMs.
- 18:51
- Unidentified speaker (UNKNOWN): Yeah, yeah, interesting.
- 18:53
- Unidentified speaker (speaker_1): Um
- 18:54
- Unidentified speaker (speaker_0): how how do you think about click through citation? That's been a I think a uh a struggle of the architecture of Claude and ChatGPT directly or although they're sort of working on that, whereas the vertical vendors have been you know from the start more tensional about a claim has
- 19:14
- Unidentified speaker (speaker_0): a sort of ability to click through citation whether it's the number of the transcript. What what do you think is the right amount of like inline citation, the right workflow back to a transcript, if someone's building an agent, what do you think is the right workflow to click through to the to the raw document?
- 19:32
- Unidentified speaker (speaker_2): I mean the w we we've always been a big
- 19:34
- Unidentified speaker (speaker_1): believer in inline citations. It's it's how the core of our web app has always functioned. You know, we've had the you know AI search natively within the third bridge application for for for for a long time. And you know, being able to demonstrate inline citations that ultimately can be traced through not only to like the underlying document, but actually
- 19:53
- Unidentified speaker (speaker_1): the the speaker associated with that document with their real-world kind of demonstrable expertise provides you know the users with a degree of comfort. when they're using the data that this is being pulled from reliable sources. Now, we we can obviously, you know, I don't know, your experience of using these tools and how often you
- 20:13
- Unidentified speaker (speaker_1): flick through the citations to the underlying source material. I think it happens less and less as people get more and more comfortable with the data sources that they've they've plugged in. They get more deliberate about you know how they write their their skills and how. they nominate which data sources they trust and want the the model to use versus which ones they want them to discount.
- 20:33
- Unidentified speaker (speaker_1): Uh which is why we've always been incredibly focused on um
- 20:37
- Unidentified speaker (UNKNOWN): on
- 20:38
- Unidentified speaker (speaker_1): being a pure play expert insights provider. We we we by design haven't kind of mingled any of our product with other other data, whether that's kind of like off the shelf or you know anything anything else that we you know we we may have we you know we may have acquired or licensed or anything like that. Um because you know people want to be able to reliably
- 20:58
- Unidentified speaker (speaker_1): use that MCP endpoint as the voice of the the the expert on the ground. And if it has a third bridge citation, um, you know they're
- 21:06
- Unidentified speaker (UNKNOWN): you know
- 21:06
- Unidentified speaker (speaker_1): they're they're taking comfort that it is it's it's from an individual because that's that's how we produce our content. You know we commission that content, it's produced through a conversation with an expert. And so it's it's you know it's it's sourced from somebody who's gone through vetting, ID and V, compliance approval. And and folks know that when they you know when they
- 21:26
- Unidentified speaker (speaker_1): invoke the the the third bridge MCP because that's that's what we we we kind of sell. And I think people are gonna get less and less concerned about clicking through the citation. We even see that now in user behavior, right? Citations are re the click through rate goes down and down and down the more people use the the kind of these tools. Um but I think the
- 21:46
- Unidentified speaker (UNKNOWN): more the
- 21:46
- Unidentified speaker (speaker_0): initial skep the initial skepticism, they're still hunting for hallucinations. They
- 21:51
- Unidentified speaker (UNKNOWN): want to see
- 21:52
- Unidentified speaker (speaker_0): the actual document and then they sort of realize to some to at least a large degree hallucinations a solved problem. They gain trust over time. They're just using it more nat more naturally.
- 22:03
- Unidentified speaker (UNKNOWN): Yeah,
- 22:04
- Unidentified speaker (speaker_1): 100%.
- 22:04
- Unidentified speaker (UNKNOWN): Yeah,
- 22:05
- Unidentified speaker (speaker_1): definitely. I mean, I don't I don't know if that that ties with with your your usage. Um
- 22:11
- Unidentified speaker (speaker_1): but but that's what we're seeing, yeah.
- 22:13
- Unidentified speaker (speaker_0): Yeah, yeah. No, I think I think it I think I think it I think it makes sense. What um
- 22:17
- Unidentified speaker (UNKNOWN): one
- 22:17
- Unidentified speaker (speaker_0): of the other issues I'm hearing from some clients is hey Brad, there's a there's a more and more MCPs coming out. We now have forty-five 45 MCPs. Um the system's sort of being bogged down by all of these MCP connectors. What do we do? Right. And some other clients are working on this more single pipe strategy,
- 22:37
- Unidentified speaker (speaker_0): ingesting all of the raw data and then piping that back into one into their AI terminal through one pipe. Can you just talk to us about what you've learned about that challenge across across clients? Like when clients have too many MCPs. what happens?
- 22:52
- Unidentified speaker (speaker_2): Yeah, I think I think it's it's
- 22:54
- Unidentified speaker (speaker_1): the same it's the same issue that that cards have been wrestling with when it was APIs, when it was data feeds. It's not a new problem. And I think to your point, the those that are you know resourced um
- 23:08
- Unidentified speaker (UNKNOWN): and
- 23:09
- Unidentified speaker (speaker_1): end up there they're they're still looking to pull the raw data into their into their own organizations. They want to apply they mean they want to apply their own embeddings. They want to apply their own taxonomy, their own schema, they want to consolidate all of these different data sources and um you know br bring them into their own RMS.
- 23:27
- Unidentified speaker (speaker_1): That's always been the case. That's that's still the case that you know firms that are resourced to do that prefer to do that. And I think we appreciate why. A lot of the
- 23:36
- Unidentified speaker (UNKNOWN): you
- 23:37
- Unidentified speaker (speaker_1): know the internal mapping and the knowledge graph at the you know club that's part of the proprietary research process, right? Applying that internal knowledge to um external data sources. That's that's unabated.
- 23:56
- Unidentified speaker (speaker_1): I think what I think what the MCP is is is enabling is the ability for people to spin those things up more mo much quickly and experiment with the the data sources that that they want and then ultimately over time bring those into you know an internal RMS. And I think what we've seen
- 24:14
- Unidentified speaker (UNKNOWN): on
- 24:14
- Unidentified speaker (speaker_1): the client side, which is which is really interesting, is is the ability for smaller clients to do more of this stuff themselves. So what what's what had historically been um the preserve of the very, very largest institutions with the most data scientists? You know, you're getting much smaller teams now who are looking to build stuff out
- 24:34
- Unidentified speaker (speaker_1): internally. And because they want to, you know, data governance in is is such a a big concern that they want to be running, you know, they want to be running these models in bedrock. They want to be running open weight models inside of their own perimeter. Um and you know, they want to also be able to bring in the the raw data and that they know that they're
- 24:54
- Unidentified speaker (speaker_1): licensing and process it that way.
- 24:57
- Unidentified speaker (UNKNOWN): And
- 24:57
- Unidentified speaker (speaker_1): so that that's the main way we've you know we've seen clients getting around that um that complexity that you know incorporating 40 different because people weren't really incorporating that many APIs or data feeds either. But I think I think the MCP um phenomenon has meant like you're you had this like explosion of of
- 25:18
- Unidentified speaker (speaker_1): endpoints
- 25:21
- Unidentified speaker (speaker_1): through Claude and through ChatGPT. But I think there will there will there will inevitably be some rationalization
- 25:28
- Unidentified speaker (speaker_2): there
- 25:29
- Unidentified speaker (speaker_1): where clients elect to bring in their most prized data assets and the ones that they think are the most valuable, apply all of their internal expertise to those data sets and then just continue to evaluate new data sets through MCP as they become available.
- 25:45
- Unidentified speaker (speaker_0): Yeah. Yeah, that makes sense. Um I wonder like if a if a client is evaluating Thurbridge in the
- 25:53
- Unidentified speaker (UNKNOWN): in
- 25:53
- Unidentified speaker (speaker_0): the demo, the sales process, what are the characteristics? Like what is that, what are they evaluating? Like what how how do I test and and judge a system
- 26:03
- Unidentified speaker (speaker_0): like Fairbridge MCP?
- 26:05
- Unidentified speaker (speaker_1): By typically comparing the responses they get to the same questions without
- 26:09
- Unidentified speaker (UNKNOWN): it.
- 26:10
- Unidentified speaker (speaker_1): And that's the easiest the easiest evaluation is um, you know, I have all of these research questions. This is my this is my basket of names, this is my watch list. You know, identify the the the biggest headwinds or tailwinds to to to to my names that have been talked about in the last two weeks. or whatever going into earnings. Um I'll run that
- 26:30
- Unidentified speaker (speaker_1): across the the existing assets that I have at you know at at my disposal. I'll run that against the open web and then I'll compare those responses to the richness of the response that I do or don't get when I'm connected to the Third Bridge MCP or or any other MCP candidate. Candidly. I mean that's how that's how we're seeing appliance evaluate that. Um And I think I think it yeah it
- 26:50
- Unidentified speaker (speaker_1): it you know it's the same thing they've been using to evaluate, you know, data vendors
- 26:56
- Unidentified speaker (speaker_1): since forever. It's the the freshness, the accuracy, the quality of the data. Um and actually I think MCP gives you uh an even more effective ability to evaluate data vendors because you know the the other historically you had to throw the the FTP over the wall, uh
- 27:16
- Unidentified speaker (speaker_1): say this is everything we've got, and then they would have to go and pass through it all and and and you know ingest it themselves, harmonize it just purely just to to see if it was valuable and worked with their existing pipelines and and added value to the to the end users. Now it's a lot easier to to evaluate precisely for that reason. You can ask, you know, basically you can ask Claude or um
- 27:36
- Unidentified speaker (speaker_1): or Chat GPT, hey, what in here is new? Does that I don't already know about that you don't already have access
- 27:41
- Unidentified speaker (UNKNOWN): to?
- 27:41
- Unidentified speaker (speaker_1): Um
- 27:42
- Unidentified speaker (UNKNOWN): so
- 27:43
- Unidentified speaker (speaker_1): actually it's much much much easier to evaluate than it has been.
- 27:47
- Unidentified speaker (speaker_0): More more possible to start backtesting strategies, even more you know, systematic qualitative strategies, you know, something like a management quality score that I could then pull in historical transcripts and see if. there's correlation to stock price movements over time. Have you
- 28:07
- Unidentified speaker (speaker_0): seen you using the MCP to to back test strategies more systematically?
- 28:13
- Unidentified speaker (UNKNOWN): In
- 28:13
- Unidentified speaker (speaker_2): into the in into the MCP to
- 28:15
- Unidentified speaker (speaker_1): to to kind of test the concept, sure. Um but then you'll always for for for a real, you know, production back test, um
- 28:23
- Unidentified speaker (UNKNOWN): there
- 28:24
- Unidentified speaker (speaker_1): is still a preference for the for the flat file. Um people still want to have the whole thing at their disposal.
- 28:29
- Unidentified speaker (UNKNOWN): Um but
- 28:31
- Unidentified speaker (speaker_1): but yeah, I for for spinning up POCs, we've seen like the MCP is a very effective way to do that. And we're providing, you know, access to systematics and fundamentals through through the MCP. But, you know, my my suspicion is that, you know, the the systematic community will always have a preference to
- 28:51
- Unidentified speaker (speaker_1): accessing the raw data um and being able to you know apply because again it it it it kind of depends on you know how you've done your own vectorization and embeddings um the decisions that that we as a vendor take to then expose our our content through the MCP, those would be things that the the the quants in
- 29:11
- Unidentified speaker (speaker_1): particular would have to account for in terms of constructing that back test. Um to make sure you know they're they're working with the same model, same embeddings, and that's yeah, that's that's you know very often not not the case. And so I think for those more complicated, involved back tests, you're still gonna, you know, the our clients still have a preference to
- 29:30
- Unidentified speaker (speaker_1): towards the the bulk feed. Um where we're seeing the you know, where we're seeing the MCP kind of really take off is in um those those pr potentially those. slightly more resource constrained firms who, as as, as we mentioned earlier, want to get to production, you know, much more, much more quickly. Um and maybe have a have a have a
- 29:50
- Unidentified speaker (speaker_1): greater tolerance for a I guess a lower fidelity of of of signal. For those for those use cases, you know, for for the fundamental, the quantumental quantum use case, yeah, we're s we're seeing we're seeing a lot of people incorporating the MCP for
- 30:03
- Unidentified speaker (UNKNOWN): that.
- 30:04
- Unidentified speaker (speaker_0): Yeah. Yeah, that makes sense. Um
- 30:08
- Unidentified speaker (UNKNOWN): what
- 30:08
- Unidentified speaker (speaker_0): um what what has been your experience with with voice models and AI expert, AI driven expert network calls thus far?
- 30:17
- Unidentified speaker (speaker_1): Um It's it's uh it's funny. If you'd asked me even a few years ago whether I thought they would they would take off, I would have been I I'm would have been skeptical. Again it was it was partly the the same view I had of of clients you know not wanting to publish their own calls. Um it was not my in my bingo card that experts would be you know
- 30:37
- Unidentified speaker (speaker_1): really, really stoked to um to to to speak to to agents um and not kind of interact with humans because one of It's the a you know one of the the main things that that experts um
- 30:49
- Unidentified speaker (UNKNOWN): you know
- 30:50
- Unidentified speaker (speaker_1): continue to give. us feedback on is that they value the quality of that human interaction. Um they like preparing their their ideas with with that of you know a a another another human. Um I do think that agent led interviews do have a place. Um I think they they are very,
- 31:10
- Unidentified speaker (speaker_1): very effective tools for scaling um obviously a high volume of responses on a more consistent um set of questions. I think there's no it's not really a surprise that you're seeing vendors in in our space using them for for channel check type products, um,
- 31:30
- Unidentified speaker (speaker_1): where you are looking to generate a very high end on a consistent, a consistent number of of questions or a consistent set of questions. And then you want to sort of plot those those changes through time and and provide um you know insight and and and signal to clients clients that way. What I don't what I you know what I am still
- 31:49
- Unidentified speaker (speaker_1): skeptical around is the ability for um
- 31:53
- Unidentified speaker (UNKNOWN): you know
- 31:54
- Unidentified speaker (speaker_1): agentic calls to to to fully displace those very, very high high-touch, touch, very involved engagements that clients will have with experts or that are our analysts will have with experts
- 32:06
- Unidentified speaker (speaker_1): that are of a you know a certain a certain type. And I think you know depending on the the kind of the the the the seniority of of the expert, I think you know some of those types of profiles will never be willing or will never have to do agentic calls. Um, They will always be able to speak directly with analysts.
- 32:27
- Unidentified speaker (speaker_1): and they will always be able to set their preferences um to be able to speak with analysts. But I do think there's a very you know large portion of the expert community that are willing to do calls with agents and uh and and uh and are happy with that process. And so that
- 32:43
- Unidentified speaker (speaker_1): that whole capability has has massively increased the industry's capacity to um
- 32:50
- Unidentified speaker (UNKNOWN): to
- 32:50
- Unidentified speaker (speaker_1): conduct calls and to create content. And my I guess my my overall concern there is then it it shifts the bottleneck on to the supply side and like the availability of experts. And their their willingness and ability to engage with everybody who's spun up a you know
- 33:10
- Unidentified speaker (speaker_1): a voice agent and has you know reached out to their role Rolodex or whatever Rolodex they've they've they've bought and the contacts they've pulled from from from from LinkedIn. And so I think that I think that the industry is going to have to contend with uh you know, not only how do you you know obviously maintain relationships with your with your clients
- 33:30
- Unidentified speaker (speaker_1): and continue to provide value there, but actually how do you improve the lives of the contributing experts on your platform as well. Um because I think what we're already starting to see is um you know experts rationalizing the firms that they work with because it's not sustainable to be responding to
- 33:49
- Unidentified speaker (speaker_1): everybody that that reaches out to you through LinkedIn or through email um or through this, that and the other channel. And so I think I think while it's Well it's it's it's very good, it's a good thing for for us and our and our firm in terms of our capacity to um you know find experts, match them with with clients and have that interaction taken care of asynchronously. So you know the
- 34:09
- Unidentified speaker (speaker_1): client can be doing something else, ten of these calls can be running in parallel, they get the results populated into their workspace, you know, as they come through. Um that's great, but you know, that there's only one expert. So like yeah, yeah. And like like you y you can't you can't have every single analyst who who's running due diligence on a deal at that
- 34:29
- Unidentified speaker (speaker_1): same time.
- 34:34
- Unidentified speaker (speaker_1): And so I think the the the firms that are going to do um to do to do well are gonna be those that have a you know a a laser focus on on the s on the supply side. Uh
- 34:44
- Unidentified speaker (speaker_0): yeah.
- 34:45
- Unidentified speaker (UNKNOWN): And
- 34:45
- Unidentified speaker (speaker_1): making sure that
- 34:46
- Unidentified speaker (UNKNOWN): I wonder
- 34:46
- Unidentified speaker (speaker_0): and this is just anecdotal. I mean I've sort of I've been approached from time to time to do GLG stuff. style calls and I've just never done them because, you know, the setup and then scheduling.
- 34:57
- Unidentified speaker (speaker_0): But the idea that I could go into a link at any time and talk to a voice agent for twenty or thirty minutes, ask answer a few questions, and I could do this late at night or when I'm in the car, just with more flexibility around my schedule, sort of makes me wonder like, do clients like that flexibility? flex like that flexibility to work around their own?
- 35:17
- Unidentified speaker (speaker_1): Yeah, a hundred percent. And experts like that that flexibility and and clients like that that flexibility. Um, but the most important thing I think for for for everybody involved there is is is choice. And you give them the ability to to use that model if they want to. Um, but obviously if they have and again you'll find these very
- 35:37
- Unidentified speaker (speaker_1): um you know difficult to sort. source experts who have a preference for in-person or you know live engagements with clients. And so you need to be able to offer you know all of those kind of different eventualities. Um but I do, you know, we see it in our own business.
- 35:53
- Unidentified speaker (UNKNOWN): You
- 35:53
- Unidentified speaker (speaker_1): know, voice agents um
- 35:56
- Unidentified speaker (speaker_1): aren't like a uh they're not a curiosity anymore. Like they they are in production across a number of our peers. ourselves included, um, and they're established as a as a component of of many of our clients process. So they are, you know, they are being used. Um, but I think they, you know, they they will never kind of displace the
- 36:16
- Unidentified speaker (speaker_1): the the totality of the volume of of peer to peer interactions that the expert network industry facilitates.
- 36:23
- Unidentified speaker (speaker_0): I guess there's probably a generational thing, and maybe industry. um industry delineation between willingness to like your sixty two year old copper mining executive, maybe.
- 36:33
- Unidentified speaker (UNKNOWN): Yeah, like the ag
- 36:34
- Unidentified speaker (speaker_1): the ag equipment distributor in the Midwest.
- 36:40
- Unidentified speaker (speaker_1): Exactly right. And so you need to be able to identify those individuals and then map their preferences and their willingness to do this. Um
- 36:48
- Unidentified speaker (speaker_0): The voice ages I like I think like many people like when you get into this like you know agentic voice tree for the last five years, like it makes you want to throw something in your computer screen. But recently my my son and his friends will prank call dominoes, and
- 37:02
- Unidentified speaker (UNKNOWN): I
- 37:02
- Unidentified speaker (speaker_0): was listening, it's like it's just an agent, and they sound more and more natural, like some of the pauses and language ticks text they're putting into these voice models, like it it makes them a little bit more palatable to speak. to lately, just by my anecdotal experience.
- 37:18
- Unidentified speaker (speaker_2): Yeah. No, we again the the the early
- 37:22
- Unidentified speaker (UNKNOWN): POCs
- 37:22
- Unidentified speaker (speaker_1): that we were doing in the voice agent space, they're like they're unrecognizable um kind of compared to where we are where we are now. Um and yeah the process is the the the progress you know has been has has been meaningful. Um but again I I what is again what is really important is the training that sits behind those voice agents and whether they've
- 37:42
- Unidentified speaker (speaker_1): been developed on with a corpus of um interviews
- 37:49
- Unidentified speaker (speaker_1): or or not, um whether that's deep and and how deep deep that is. And and one of the you know one of the advantages that that you know Thirdbridge Third bridge and other providers have over you know newcomers is an established library of interview transcripts, um, to which we obviously own all of the IP that we
- 38:09
- Unidentified speaker (speaker_1): can we can use to to develop those models, to improve those models, to fine-tune those models. And we also have a deep bench of sector analysts that sit on the Third Bridge payroll who can also, So you know, produce and evaluate the responses against a you know a golden set of responses that you would expect to
- 38:28
- Unidentified speaker (speaker_1): receive from those questions and
- 38:30
- Unidentified speaker (speaker_1): what you would then, as the voice agent, expect to follow up with.
- 38:45
- Unidentified speaker (speaker_1): Are you know dimensions of the conversation where where Thirdbridge has domain expertise?
- 38:53
- Unidentified speaker (speaker_1): We've been conducting interviews for
- 38:56
- Unidentified speaker (speaker_1): over a decade in a very systematic fashion.
- 39:01
- Unidentified speaker (speaker_1): We have a playbook and we think we're one of the best when it comes to that. And so what we're able to do is
- 39:08
- Unidentified speaker (speaker_1): turn those analysts to to towards evals so that we can continuously improve that that product as well.
- 39:15
- Unidentified speaker (speaker_0): Yeah, that makes a lot of sense. Um
- 39:17
- Unidentified speaker (UNKNOWN): what do
- 39:18
- Unidentified speaker (speaker_0): what do you think the sort of frontier uh fundamental investing looks like today with respect to expert network calls and agents and MCPs more broadly? Like what are you seeing and hearing emerge from your clients uh today and where do you think that goes over the next six to twelve months?
- 39:36
- Unidentified speaker (speaker_2): I think um
- 39:37
- Unidentified speaker (UNKNOWN): I think
- 39:38
- Unidentified speaker (speaker_2): what's really interesting is
- 39:41
- Unidentified speaker (speaker_1): the new personas that we're finding kind of engaging with expert content. Um obviously the you know the the our main constituents are the analysts. Um they always happen they they you know they do the calls, they read the research. Um but this is also kind of through a sort of a partnerships lens. But we're also um you know
- 40:01
- Unidentified speaker (speaker_1): finding a lot more interest in like the the PM and like the risk persona where you know the the the the semantic retrieval that's now available through LLMs and the the you know the ability to do that at scale means that that's also a very useful input on the you know the portfolio
- 40:21
- Unidentified speaker (speaker_1): side on the risk side um because now you you know if you've if you've if you've if you've done your homework and if you've if you've structured the the content in the right way, if you've tagged entities to to claims, it's it's easy enough for those to then be pulled into, you know, into the portfolio compared to to my positions. And then I can just
- 40:41
- Unidentified speaker (speaker_1): escalate all of those red flags, green flags around expert X says we're having inventory pull pull-through through challenges in this region and you know this this distributor is distributing,
- 40:53
- Unidentified speaker (UNKNOWN): you
- 40:54
- Unidentified speaker (speaker_1): know, X products. You can then start to, you know, map different interdependencies as well that might be relevant at the the PM level. And you don't you don't necessarily need the the analysts to take all of those, distill those, and then give them,
- 41:10
- Unidentified speaker (UNKNOWN): you
- 41:10
- Unidentified speaker (speaker_1): know, to the portfolio manager in the same way. Um so with we're s and I think that's you know, that's precisely why you're able to see a lot of these um
- 41:19
- Unidentified speaker (UNKNOWN): you
- 41:19
- Unidentified speaker (speaker_1): know much leaner hedge funds now as well. People more people going out on their own. You know, they bring in a they bring in a uh you know a a portfolio management software, maybe pu a pure play provider. There are I think there are a number of really interesting players in that space right now. And they're able to map the signals. coming up
- 41:39
- Unidentified speaker (speaker_1): from alt data, from expert insights, you know, directly to those holdings, and then get programmatic alerts when somebody says something negative about something that I carry particular exposure to.
- 41:57
- Unidentified speaker (speaker_1): With yeah, the the the PM now, as opposed to maybe like market data and analysts who were previously like the the the mainstay of your business. Um that's not to say that they they they aren't still, but it's it's really interesting now the the level of engagement that you know the more you know the more senior folks within the
- 42:17
- Unidentified speaker (speaker_1): fund are applying to these types of content sets because you know ultimately they're making the I guess the judgment call at the end of the day. Uh and they want to have more of an understanding of where. that that content is coming from as well.
- 42:29
- Unidentified speaker (speaker_0): And historically there's been this giant game of telephone where you have the junior analyst doing the research and then feeding that back up the the chain. And you know, m uh most often those junior analysts are doing a good job with high fidelity of sort of translating those insights, sites but but not always, right? The ability to have sis
- 42:49
- Unidentified speaker (speaker_0): systematize this is a really
- 42:51
- Unidentified speaker (speaker_3): interesting
- 42:51
- Unidentified speaker (speaker_0): really interesting um idea interesting uh response
- 42:56
- Unidentified speaker (UNKNOWN): what
- 42:56
- Unidentified speaker (speaker_0): what what what what comes next i mean there's sort of the you draw the you sort of draw the line up to up and to the right and you can start to see these systems just get better and better right what what do you think the future of fundamental investing looks
- 43:09
- Unidentified speaker (UNKNOWN): like
- 43:10
- Unidentified speaker (speaker_3): um I
- 43:11
- Unidentified speaker (UNKNOWN): I
- 43:12
- Unidentified speaker (speaker_3): think there's a there's a there's a few schools of of thought.
- 43:15
- Unidentified speaker (speaker_1): Um I think everybody, you know, I and I think a few other guests on your podcast have said the same thing that you you will get this kind of convergence in the middle between, you know, traditionally discretionary fundamental investment research and and and capital allocation and and quant research and allocation.
- 43:35
- Unidentified speaker (speaker_1): They will they will meet in the middle because
- 43:38
- Unidentified speaker (UNKNOWN): I
- 43:38
- Unidentified speaker (speaker_1): I do believe that that AI has made a lot of these kind of mainstay discretionary fundamental resources much more useful at scale. Um so so much so that they can be used as, you know, a a a factor in systematic decision making. Um but I do think a lot of those like
- 43:58
- Unidentified speaker (speaker_1): market factors and and data sources that the the quants use
- 44:03
- Unidentified speaker (speaker_1): fundamental guys are becoming more aware of those as well. So like pricing signals, like volume, like who's in the market. And you can argue, you know, that may change with dark pools and this, that, and the other. Um, but the, you know, the ability to um to to analyze quantitative data is is being it's it's it's being commoditized. And I and I do think
- 44:23
- Unidentified speaker (speaker_1): that the differentiation will come
- 44:25
- Unidentified speaker (UNKNOWN): in
- 44:25
- Unidentified speaker (speaker_1): the form of who can generate more profound differentiated fundamental insights on the specific company in question or group of companies in question. And I think there are there are many interesting ways that we're seeing clients exploring those now. Whereas the fundamental guy and and you know, they have always
- 44:45
- Unidentified speaker (speaker_1): taken like a uh, or many of them have always had a more holistic view of their coverage. So not only the company. companies they focus on, but also the peers, the competitors, the suppliers. But that is a very difficult task for an individual to be across, particularly at scale, depending on how many names are in their coverage. Um, but I think what what what we're seeing
- 45:05
- Unidentified speaker (speaker_1): now is people taking more of a
- 45:08
- Unidentified speaker (UNKNOWN): like,
- 45:08
- Unidentified speaker (speaker_1): how do I model this entire ecosystem? And how do I systematically flag something over here in this happening? that's happening in the the branches? And how do I link that back to something that's in the core, the core of my my my thesis? And and how many different um you know orders of of separation does that piece of information go through? And how do I how do I how do I
- 45:28
- Unidentified speaker (speaker_1): map that up? Um and I think that's fundamental investing at its core is like this is my asset and this is how it relates to the other assets, you know, in its in its in its world. And I think the the the really interesting thing about expert insights is kind of the way that the content is generated from not only company formers, but
- 45:48
- Unidentified speaker (speaker_1): extensively from suppliers, competitors, and customers, is you actually start to be able to create some really, really interesting linkages between contents, content and companies. You know, were previously very, very difficult to do so, particularly on kind of private markets. So if somebody said something negative over here, there was not always a very easy way to map that
- 46:08
- Unidentified speaker (speaker_1): more systematically to um uh to the companies they care about and the companies I follow. Um so we've been doing a lot of really interesting work on the knowledge graph side as well, in terms of how we're mapping companies to each other through you know the the claims that we kind of see in the in the in the content. Um and I think that's that's where fundamental research is
- 46:27
- Unidentified speaker (speaker_1): is going, more to like mapping entire ecosystems and modeling entire ecosystems, as opposed to modeling individual companies now, which I think is um
- 46:39
- Unidentified speaker (speaker_1): I mean you've seen you obviously we've seen the progress that Claude has made and ChatGPT has made in terms of like the Excel modeling, like the three statement models, like it's it's excellent at doing those now. Um and I think what what the domain of of judgment is going to move now more towards well, how how does how does this model interact with that model, interact with that model?
- 46:59
- Unidentified speaker (speaker_1): And can I generate any differentiated insights that way? Um so I think yeah, it's it's more it's it's more of the same, I guess, but just at a much much bigger scale.
- 47:08
- Unidentified speaker (speaker_3): Yeah,
- 47:09
- Unidentified speaker (speaker_0): no, totally agree.
- 47:10
- Unidentified speaker (speaker_3): Um Uh
- 47:11
- Unidentified speaker (speaker_0): I guess I agree with that framing. I mean, back in 2008, you could go do 20 GLG calls and that was variant perception. The guy did these checks, and now you're competing with market participants who get a hundred fresh expert network calls piped into their agentic UI every morning. Um and so how do you
- 47:31
- Unidentified speaker (speaker_0): find something that how how do you find differentiation when news is c news and calls and data is become commoditized? has become beta almost by definition. Uh so yeah, connecting.
- 47:43
- Unidentified speaker (UNKNOWN): And
- 47:43
- Unidentified speaker (speaker_1): I d I don't know I don't know if expert libraries become beta over time. Yeah. Perhaps perhaps they perhaps they do. And I I I think as a business we wouldn't necessarily complain about that being the case. But what I think that would then do is is, you know, create more of an emphasis on bespoke custom proprietary research, which
- 48:03
- Unidentified speaker (speaker_1): clients will will still want to do. I think the content library has validated expert insights as a meaningful input to the research process. And the libraries that have the quality, the breadth, the freshness will will do well. But what I also anticipate is this
- 48:21
- Unidentified speaker (speaker_1): You creating know, more demand for truly differentiated proprietary expert calls as well. So, you know, firms doing more of their own private calls with more senior experts, more expensive experts, um to to bring more of that research back in-house and have more proprietary
- 48:41
- Unidentified speaker (speaker_1): content as well. Um I think those two things will exist in parallel for a long time, where um, you know, big broad libraries of expert content um will provide that. immediate low latency response to a research question that a client has.
- 49:09
- Unidentified speaker (speaker_1): um to to to fill in those fill in those gaps and fill in those banks. Um
- 49:13
- Unidentified speaker (UNKNOWN): and
- 49:14
- Unidentified speaker (speaker_1): so I don't I don't see it reducing the demand for for proprietary one to one research either. But obviously it will it it will reduce the demand for mediocre proprietary one to one research. And I think there will you know there'll be many calls that now no longer need to happen because there are X number of you know scaled up content libraries on the market now. And
- 49:34
- Unidentified speaker (speaker_1): that does most that gets you most of the way there. Um
- 49:38
- Unidentified speaker (speaker_1): Yeah
- 49:38
- Unidentified speaker (UNKNOWN): and I'd
- 49:38
- Unidentified speaker (speaker_3): probably
- 49:39
- Unidentified speaker (speaker_0): sort of add like, you know, I think mosaic theory continues to hold where in sort of the three dimensionality of like what's in the stock, what do I believe, what's the sort of gap in that expectation holds. I mean, the vast majority of expert network calls or transcripts don't have a one to one alpha signal. Like almost never did That I do twenty 20 calls and
- 49:59
- Unidentified speaker (speaker_0): then it's sort of a clear, you know, uh a clear line to an alpha signal. It's a it's a process to deepen comprehension about companies, industries. And over time, the fundamental investor is in modeling, doing costs, document management, looking at data, and sort of a you know probabilistic view of the
- 50:19
- Unidentified speaker (speaker_0): future emerges through that process, and you're continuously comparing that to market implied expectations. That's sort of the whole basic at
- 50:26
- Unidentified speaker (UNKNOWN): least
- 50:26
- Unidentified speaker (speaker_0): public. in markets, the whole name of the game. So yeah, I don't necessarily know that like quants systematically over overlay that, although there's certain certain channel checks that provide us that can almost become a become a data set. Um
- 50:40
- Unidentified speaker (speaker_0): but um
- 50:41
- Unidentified speaker (UNKNOWN): yeah,
- 50:41
- Unidentified speaker (speaker_0): no it's definitely it's definitely rapidly shifting. You know, definitely rapidly shifting how we spend our time, right?
- 50:47
- Unidentified speaker (speaker_1): It is. Um it is. And I think one of you know
- 50:51
- Unidentified speaker (speaker_1): one of the the interesting questions is well, if you're if you're freeing up the analysts' time um to not conduct as many calls, like what are they what are they choosing to to kind of fill that time with? Um and I think what we're seeing is people using that to expand expand coverage. Um
- 51:08
- Unidentified speaker (speaker_0): awesome, great Dan. Well this was really interesting uh and exciting stuff. What's the best way for people to get in in touch with you if they want to learn more? about Third Bridge?
- 51:16
- Unidentified speaker (speaker_1): Um,
- 51:17
- Unidentified speaker (UNKNOWN): feel
- 51:17
- Unidentified speaker (speaker_1): free to reach out on LinkedIn um or you know info at thirdbridge.com. Um somebody will get back to you. Um yeah, happy to take you through all of the you know the offering, the data, data sets. Uh yeah, those are probably the best ways.
- 51:30
- Unidentified speaker (speaker_0): Sounds good. Well, thanks a lot for the time today.
- 51:33
- Unidentified speaker (speaker_1): Thanks, Brett.
- 51:34
- Unidentified speaker (speaker_1): Bye.