Fireside Chat with Teresa Heitsenrether

In a rare behind-the-scenes look at how GenAI is reshaping daily work at JPMorganChase, Teresa Heitsenrether described a transformation that now touches every corner of the firm. What began as an experiment to put AI tools in people’s hands has become a company-wide ecosystem—connected to internal data, earnings, and news feeds—powering everything from lawyers negotiating contracts to call centre teams resolving issues faster.

Key Takeaways:

1. Heitsenrether urged leaders to rethink what business looks like when the marginal cost of thousands more digital employees approaches zero—where productivity soars and roles evolve - and managers must learn to lead both human and digital colleagues.
2. More than 150,000 employees use the firm’s LLM suite each day. To prevent a plateau in adoption, the focus is on giving people access to the right tools, training, and mindset to make AI work for their everyday business challenges.
3. Whilst model partnerships from Anthropic and OpenAI underpin LLM Suites stack, JPMorgan’s centralized controls and data protection frameworks are a critical part of making innovation safe and scalable.

Pushing the limits: how can financial services firms develop specific, proprietary innovations?

Financial services companies are under increasing pressure to transform their businesses, not just by bolting AI onto existing processes, but to rethink their models entirely. The session explored how finding the right AI partners (both internally and externally), as well as meticulous measurement and ongoing evaluation can make or break a firm’s journey towards ROI.

Key Takeaways:

1. Adoption needs to be a CEO-level mandate, a point both Martin Kon (Cohere) and Bobby Grubert (RBC) emphasised, flagging that transformations only happen when there’s a direct line between the technology and the executive team.
2. The future labor pool in a bank is a mix of agents and humans, and the key to growth is building domain-specific knowledge into both. Banks that don’t think about how processes might change with this new talent pool risk falling behind their competitors.
3. As other parts of the world are speeding ahead, warns Drew Cukor (TWG AI), financial services companies need to set the culture, build the compliance team and set up robust measurement systems so they can stop just studying use cases and build AI systems that can transform the firm top-to-bottom.

Unlocking value through oversight: how can financial services firms supercharge AI through de-risking and governance?

A sharp discussion on the tension between innovation and oversight reframed “AI risk” as familiar terrain within the banking sector—just scaled, accelerated, and less visible. Panelists argued that the goal isn’t to invent new governance, but to adapt existing frameworks to systems that learn, reason, and act. The real unlock comes from making oversight itself intelligent and continuous.

Key Takeaways:

1. Ian Glasner (HSBC) emphasised that AI risk ≠ new risk. Banks should anchor governance in established taxonomies (operational, reputational, cyber), but also enhance observability and traceability to handle autonomous or agentic models.
2. Al Slamecka (Cisco) emphasized the need to invest in “Agentic Ops.” Once models start executing tasks, not just advising, accountability and audit trails must evolve—as human oversight can’t be bolted on after the fact.
3. Oversight done right becomes a growth function. Embedding risk intelligence early lets financial services scale AI safely into KYC, lending, and customer service use cases.

Fireside Chat with Marco Argenti (Goldman Sachs)

Even amid AI’s rapid evolution—from reasoning models to autonomous agents—Marco Argenti emphasized that the fundamentals haven’t changed: data and people remain “fixed stars” in a time of change.

Key Takeaways:

1. Firms must build a strong “data quality story,” where information is accurate, deduplicated, and machine-readable. Goldman’s Legend data platform adds a semantic layer atop raw data, improving grounding and reducing hallucinations—vital as the firm moves into an “age of agents.”
2. AI performs best where feedback is clear, but the next challenge is encoding “tribal knowledge”—the accumulated intuition behind human judgment—to extend AI into complex domains like auditing and investing.
3. With ~20% developer efficiency gains already realized, Argenti noted that future differentiation will depend less on cost and speed, and more on how effectively firms scale collaboration between human expertise and intelligent systems.

Fireside Chat with Prem Natarajan (Capital One)

A detailed and nuanced map for how Capital One has built a “tech company that does banking.” From internal tools to the industry’s first customer-facing agentic chatbot, Prem Natarajan described the firm’s continuing focus on building reusable platform capabilities, such that every line of business can deploy AI faster.

Key Takeaways:

1. By fine-tuning open-weight models like Llama and Nematron with proprietary data, Capital One achieves faster, lower-latency performance at a fraction of the inference cost.
2. Talent is destiny: Capital One’s modern tech stack and research partnerships attract top AI talent motivated by impact—as Natarajan told us, the ultimate ROI is a happy customer.
3. These ingredients got them to Chat Concierge—an industry-first, customer-facing agentic chatbot—that demonstrates the importance of combining reasoning and specialization to break complex tasks into atomic ones, delivering a 25–30 point performance lift in production use cases.

Looking beyond the use case: how can financial services leaders deliver true AI transformation?

Years of AI pilots and proofs of concept have been replaced by a need for scale. The session explores how financial services companies are reconciling the breakneck pace of innovation against delivering tangible results—and thereby preparing for a world of AI ubiquity.

Key Takeaways:

1. Rigidity is a one-way ticket to irrelevance: Building an AI strategy requires adaptability. Sathish Muthukrishnan (Ally Financial) reminded us that new, better models are released every day, and unless firms are built to take advantage of the strengths of each, they risk falling further behind.
2. Talent is still the differentiator. Jodie Wallis (Manulife) stressed the scaling use cases require buy-in at all levels of a business, starting with the C-suite. The most effective tools touch multiple processes. For them to work well, you need people who can provide context to the data that feed them—and you need champions that model their practicality and efficacy.
3. Find the patterns: To move beyond the pilot stage, look to the processes that have similarities across disparate parts of the firm, as well as the types of tools that can be repurposed across multiple lines of business.

Fireside Chat with Jason Droege and Greg Ulrich

This session explored the realities of moving from pilots to production at scale. Greg Ulrich (Mastercard) called out the past six months as a turning point—graduating from proofs-of-concept to daily-use tools across business units. Meanwhile Jason Droege (Scale AI) cautioned that success depends less on hype and more on picking the right problems to solve.

Key Takeaways:

1. Greg Ulrich shared how Mastercard shifted AI from small tests to a shared platform used by thousands for research, knowledge, and content creation. Specific to the risk team, Mastercard added a merchant-graph signal to fraud models across regions without adding latency.
2. ROI is slow, not sudden. Major process changes take time (typically 9-12 months) from idea to measurable impact. Consequently, Jason Droege cautioned that steady, compounding gains are more credible than claims of overnight transformation.
3. Whilst AI gets likened to a magic wand, implementation is not. Defining potential problems for AI to solve need careful assessment, review, and prioritization. To that end, Scale AI’s teams target tasks humans handle poorly, and then refine outputs through feedback loops. In the context of agentic AI, this scoping makes the critical difference between producing a promising demo versus an application with durable value.

Fireside Chat with David Griffiths

Citi’s AI rollout has been as much about people as technology. Across 83 countries, more than 4,200 “AI accelerators”—or superusers / internal evangelists—accelerate adoption via sharing short demos, success stories, and word-of-mouth enthusiasm. A single, global platform has replaced a patchwork of tools, making AI easier to govern and harder to ignore.

Key Takeaways:

1. Grassroots engagement has turned AI from a top-down initiative into everyday behavior. Internal tools like Citi Assist are now part of daily workflows, while developer-focused AI systems deliver some of the clearest returns: “people literally stop you in the hallway to show what they’ve built,” said David Griffiths.
2. Agentic systems need solid underpinnings. Citi continues to invest in private cloud infrastructure to scale agents efficiently, favoring open-source frameworks to stay flexible and avoid lock-in.
3. As agents take on more complex, customer-facing roles, Citi is doubling down on reliability.

Delivering on agentic potential: how can financial services firms develop agents to add real value?

Lively and pragmatic, this session tackled the buzzword head-on. “Are we drowning ourselves in taxonomies?” asked moderator Edward J. Achtner (BMO). In response, panelists unpacked what agentic really means—moving beyond vague definitions by outlining several practical steps to apply agents to complex and interconnected processes and workstreams.

Key Takeaways:

1. Panelists agreed that arguing over terminology misses the point. What matters is designing agentic systems that deliver real business outcomes, not just fit a theoretical definition.
2. Effective agents begin with small, clearly scoped tasks that can improve through feedback and experience. Manuela Veloso (JPMorganChase) urged teams to “just get started” by solving for the minimum condition—and then apply an iterative framework that gives agents skills, memory and the ability to learn from one another (in that order).
3. Success depends as much on people and infrastructure as on models. Building the right feedback loops, data access, and governance frameworks will ultimately determine where and when agentic systems stay in the lab—or move into production.

Fireside Chat with Hari Gopalkrishnan

There aren’t many processes at Bank of America that are not being reimagined through the lens of AI, according to CTO Hari Gopalkrishnan. The fireside chat explores how the bank develops new AI projects, changes culture to match the times, and tracks the gains delivered by expanding access to the technology.

Key Takeaways:

1. Fix your processes, then rethink them. Applying AI to broken processes is only going to accelerate the speed at which you deliver bad results. True ROI comes from end-to-end redesigns of how different parts of the bank work.
2. Make everyone an inventor. From executive-level training to bottom-up initiatives for idea generation, let people who understand their domains co-create AI solutions that make a difference in terms of productivity and efficiency.
3. Focus on what you measure. With coding agents, test how much faster engineers can complete a task. With call center automation, measure your call volumes. The bank is focusing on setting up robust systems to measure value in all the ways it uses the technology.

Building with agents in action: how are firms democratizing and maximizing the value of agents in deployment?

This discussion focused on how organizations can scale agentic AI without losing coherence. Panelists agreed that the real challenge isn’t building more agents, it's about building them better through unified, observable, and integrated processes rather than scattered experiments.

Key Takeaways:

1. Kunal Madhok (Wells Fargo) warned that without a unified foundation, companies risk “50-60 disconnected solutions” instead of genuine transformation. Shared platforms with observability and component reuse are key to scaling responsibly.
2. David Walker (Westpac) reminded the audience that AI must meet the same standard that banks run on - trust. Building confidence starts with leadership sponsorship, clear purpose-based access, and psychological safety for teams to experiment and pitch ideas.
3. Productivity is easy to track—but experience, both customer and employee, is the deeper metric. To that end, Westpac’s team surveys internal users and benchmarks code-generation pilots to quantify adoption and satisfaction together.

Evolving the playbook: what have we learned so far about deploying AI at scale?

A cross-bank panel examined what it really takes to scale Agentic AI beyond flashy pilots—digging into platform strategy, trust and quality foundations, ROI discipline, and the realities of organizational change.

Key Takeaways:

1. Traditional role-based access doesn’t work for agents, requiring banks to establish trust through purpose-based controls. Dan Jermyn (CommBank) stressed the need for purpose-based entitlements, testing sandboxes, and deep observability before deployment.
2. ROI = Adoption Rate × Process Dedesign. Real value comes from rethinking workflows, versus counting minutes saved. Controlled experiments and adoption metrics, not just cost cuts, signal true impact.
3. AI Maturity = Assist → Augment → Act. Scaling top-performer expertise, isolating critical points of failure, and safely advancing toward semi-autonomous actions prove critical to deploying AI at scale. This checklist requires a change in mindset from stakeholders—treating AI as a long-term business transformation, not simply a technology upgrade.

Transforming the industry: how is AI revolutionizing knowledge in financial services?

This discussion explored how financial institutions can turn streams of raw data into usable knowledge. Tucker Balch (Emory) opened by asking whether data and knowledge are truly distinct—a question that set the tone for the ensuing debate…

Key Takeaways:

1. Panelists agreed that data only becomes useful when brought to life with meaning. “Linguists at one asset manager need data animated by context,” noted Jonathan Pelosi (Anthropic), describing how the platform fine-tunes Claude with reinforcement learning on financial tasks to improve factual grounding.
2. Igor Tulchinsky (WorldQuant) emphasized the shift towards AI-driven simulations and the importance of language in financial modeling.
3. Antonio Bravo (BBVA) detailed their efforts in mainstream AI adoption, digital transformation, and capturing internal knowledge to improve client relationships and operational efficiency.