2023 Conference on the Economics of Financial Technology · University of Edinburgh Business School · 21 June 2023. Presentation image from the conference recording.
Financial conferences produce hours of discussion about regulation, banking, market structure and emerging technology. The challenge is not generating a shorter summary. It is finding the relevant insight, identifying who said it, preserving its context and making the underlying evidence available for scrutiny.
In this 10:47 presentation, Geoffrey Baker demonstrates how multimodal AI can turn recorded financial discussions into searchable, classified and source-linked research. The accompanying 2023 presentations also address a question executives still face: how do you obtain value from generative AI without surrendering accuracy, confidentiality and human judgment?
WATCH THE PRESENTATION · 10:47
The demonstration drew on 35 video sessions from diverse financial-services, banking and fintech conferences. The team curated subject matter topics such as behavioral finance, synthetic data and distributed ledgers. They also supervised classifications for identified needs and decisions. Conference agendas informed the starting taxonomy; AI-assisted processing and human review followed.
Rather than treat a recording as a single document, the workflow organized the corpus by session, speaker, topic and transcript passage. This made it possible to examine how an issue appeared across conversations, who raised it, and whether the surrounding remarks supported an extracted claim.
The presentation showed an indexed recording library with speakers, durations, summaries and extracted insights. One illustrative panel brought together Bernanke, Diamond and Dybvig in a discussion of banking, financial crises and their research. Researchers could move between a synthesized account of the panel and the underlying transcript.
Examples grouped observations as pain points, identified needs, issues, objections, buyer signals and decisions, alongside domain topics such as decentralized finance, open banking and sustainable finance. Speaker-level views and cross-session comparisons turned free-form remarks into research questions and candidate patterns.
The important distinction is between a model's proposed classification and a finding accepted by an expert. The presentation displayed extracted, contextual quotations and a review interface; the presence of a label did not make it verified truth.
The AI-Ally demonstration showed generative questioning alongside a transcript and recording. A researcher could ask for a concise explanation of a speaker's remarks, inspect the relevant passage and continue investigating. In the deck, a regulatory discussion involving China's big-tech sector illustrated this interaction.
The broader use case is knowledge preservation: conferences and expert conversations can be revisited and compared rather than disappearing into video archives. Private, proprietary datasets and collaborative learning were presented as possibilities, not demonstrated financial returns.
The companion conference presentation by Taylor Spears, Lecturer in Financial Technology and Organizations at the University of Edinburgh Business School, supplied an important counterweight to the demonstration. Spears used the idea of legibility: technology makes complex information easier to classify and analyze, but the resulting categories can also erase context and local knowledge. Generative AI may expand researchers' access to audio, multilingual text and unstructured records, while creating new problems in verifying outputs.
Spears identified practical challenges for financial researchers: outputs vary with prompts; convincing model-generated statements can be false; and peer review becomes difficult if empirical work depends on poorly understood AI transformations. For regulated institutions, this connects with established model risk practices: validation, documentation, independent challenge and controls over models used in decisions.
The discussion went beyond individual errors. Large models can be opaque, shared datasets may induce correlated behavior, and dependence on a concentrated set of cloud providers may create operational and systemic exposures. Spears' slides used portfolio insurance and the 1987 market crash as a historical warning about widespread reliance on similar decision models. This is not evidence that generative AI caused that event.
Spears argued that standardization can make a field more legible while omitting facts that matter locally. The same trade-off is present when an AI pipeline assigns every conversation to a controlled taxonomy. A topic count is useful for locating possible evidence; it is not a substitute for reading the passage in which it appeared.
The model-risk discussion pointed to US Federal Reserve SR 11-7 and comparable European and UK regimes. Large generative models complicate familiar practices such as independent challenge, documenting limitations and producing feasible comparison or challenger models. Leaders need to decide what level of testing and human sign-off is proportionate to the consequence of a wrong answer.
In the financial-stability discussion, Spears raised nonlinearity, opaque data dependencies, model herding and reliance on concentrated cloud providers. The practical concern for a sponsor is not only whether one model makes a wrong prediction, but whether shared technology and data can cause institutions to respond similarly or fail together.
Companion conference perspective: Spears' presentation provides research and risk context for the video demonstration by Geoffrey Baker.
The presentation was an early demonstration of financial research supported by AI, with evidence linked to its sources. It was not evidence of a fully autonomous enterprise agent or independently measured financial returns. A credible operating mandate should answer four questions: What decision or workflow improves? What records and evidence must be retained? Where does a qualified human validate an answer? How will quality, effort and risk be measured before additional investment?
The experiment's 35 video sessions describe the research sample, not analyst hours saved. The presentation illustrated a working research approach; it did not establish a production-wide accuracy score, verified cost reduction or financial return.
From demonstration to accountable value: An AI initiative becomes an operating result only when the organization can connect a defensible answer to the workflow, the human decision and the economics that justified the investment.
Strategy & Value
October 14, 2026
A sales playbook is a guide that helps sales teams close more deals and increase win rates.
October 14, 2026
A successful sales strategy prioritizes providing customer value and building trust over focusing on price alone.
When important decisions need operating experience, get principal-led support.
Principal-led advisory and operating leadership for executive teams facing consequential AI, governance, transformation, commercialization, and recovery mandates.
advisory@inteligems.io
