Trading Places in Financial Services: How Accountability Changes Seats in Agentic Systems
Table of Contents
Data + AI leaders across financial institutions today are currently faced with an increasingly difficult governance question: when an AI system makes a decision that goes wrong, who’s accountable?
The business built the process. Technology deployed the model. Risk set the guardrails. And yet, none of these functions alone feels accountable for AI failures.
At the CDO Magazine Financial Forum in NYC, Diego de Aragão, CFA, SVP, Balance Sheet Executive Decision Support at Citi, Andrew Reiskind, Chief Data Officer at Mastercard, Sami Huovilainen, Managing Director, Head of Next Generation Analytics, US Personal Banking at Citi, and Charles Morris, SVP, Agentic Enterprise Strategy Leader at Truist came together to discuss this question – and more.
Moderator Will Robins, VP Product at Monte Carlo, opened the session by asking the audience to raise their hands if they felt their organization had figured out its chain of command for AI accountability.
Almost no hands went up.
The panelists, a group that between them oversees everything from Citi’s $2.6 trillion balance sheet to Mastercard’s enterprise data governance to Truist’s emerging agentic strategy, spent the next hour trying to map the territory.
Table of Contents
Structuring for Speed Without Losing Control
The old playbook — stable functional teams, gradual iteration, use-case-by-use-case ROI sign-off — is straining under the weight of agentic AI. Charles Morris argued for moving away from use-case backlogs toward “compounding capabilities”: building systems whose combined value exceeds the sum of their parts.
“It’s not enough for a use case to justify its own NPV,” he said. “Does it contribute to a set of self-reinforcing capabilities where the sum of the whole is greater than the sum of the parts?”
Diego de Aragão described Citi’s approach through three lenses: making solutions repeatable across the organization, targeting value that doesn’t require massive architectural overhauls, and ensuring accountability roles are defined before deployment — not after something breaks.
Andrew Reiskind flagged a governance shift catching many institutions off guard. Product teams that once handed models off to data teams and walked away are now deeply invested in which LLM version is being used and why. “You have a whole new group of stakeholders who’ve never been engaged in governance before,” he said — requiring a major expansion of education efforts just to keep pace.
Who Gets the Call at 3am?
The answer the panel largely agreed on: the business sponsor.
Sami Huovilainen, whose team is deploying LLMs across Citi’s 300 million annual customer service calls, stressed that the more important question is what happens before launch — establishing explicit risk thresholds, running millions of adversarial test scenarios, and ensuring a kill switch is always in place.
“There will be mistakes, one hundred percent. But before you launch, you’ve already made it clear to your seniors that this is what’s bound to happen and the benefits outweigh the risks.”
Reiskind noted that back-end incident response processes haven’t changed much — AI is just another failure mode alongside data outages. What has changed is sensitivity: a chatbot publicly misbehaving carries reputational stakes that a quiet system outage does not.
Morris introduced the “Truist digital teammate” framing: treat an AI agent like an employee — it has a boss, and that boss is accountable. But he pushed on the harder challenge: how do organizations continuously improve as agents become standard, not just a pilot? “What happens when this is business as usual?” Most organizations, he acknowledged, don’t yet have an answer.
Measuring What You Can’t Easily Count
Reiskind was blunt: in financial services, vibes don’t cut it. But he acknowledged that some value is genuinely hard to quantify — especially when AI is discovering inefficiencies rather than fixing known ones. He described a former governance requirement where ROI had to be embedded in the AI approval process itself: if a system wasn’t generating value, it didn’t meet governance standards.
Huovilainen offered the cleaner example: for customer service automation, the value chain is clear — containment rates, handle time, call quality. The team committed to specific metrics at investment time and returns monthly to prove delivery. “Every month after launch, you need to come back and show you’re delivering what you committed — otherwise you’ll never get to solve the second problem.”
Morris’s spicier take: the organization most in need of evolution isn’t data or technology — it’s finance. When AI pulls forward capacity or improves outcomes through leading indicators rather than direct cost cuts, most financial models can’t capture it.
“We’re going to leave opportunities on the table if we don’t figure that problem out.”
The Real Definition of AI Maturity
De Aragão closed with the sharpest take of the session: real AI maturity isn’t a Chief AI Officer or a Center of Excellence. It’s when every senior leader across the firm understands where AI is operating within their remit, what decisions it’s making, and what outcomes it’s driving — and that becomes normalized, not remarkable.
Accountability in the AI era can’t be owned by a governance team on behalf of the enterprise. It has to be distributed, internalized, and lived by whoever is responsible for the result. Which, increasingly, means everyone. Tools like AI observability provide the visibility teams need to operationalize this ownership.
Five Takeaways
- Define risk appetite before launch, not after failure — and get second-line sign-off before going live.
- The business sponsor answers first when something goes wrong, not the technology team.
- Governance now requires education at scale as product teams engage directly with AI systems.
- Compounding capability models are replacing use-case backlogs as the right frame for prioritization.
- Finance organizations must evolve their measurement models to capture value that doesn’t fit neatly into cost reduction or revenue increase.
Our promise: we will show you the product.
