An insightful conversation with Arjun Sirrah, Founder & CEO of Titan on banking-native AI, Safe at Speed innovation, AI adoption challenges, explainability, and regulatory-ready transformation in banking.
Arjun, to begin with, could you share your background, the team you’ve built at Titan, and how Titan came to life? What was the original problem you set out to solve, and how has your vision for the company evolved since then?
I started my career at Goldman Sachs in the Securities Division, right after the financial crisis. I loved my time there, but for anyone working with machine learning in the early AWS era, the idea of starting a company became increasingly compelling. I started my first company in 2013, then sold it a year later. I then joined as the founding CTO of a company that became Laurel Road. Built inside a community bank in Connecticut, Laurel Road became a digital bank doing a billion dollars of lending a year, entirely on home-built software, and was eventually sold to Key Bank.
The interesting thing about this journey is that since most of the development was done inside a community bank, the team quickly learned what it meant to operate in the banking environment. We spent time with the FDIC, going through regulatory exams, and working directly with bank operators. That actually created the foundation for what we’re building at Titan; AI agents and models that think, reason, and operate the way bankers do.
It’s no secret financial institutions today are working to implement AI. However, there are three main obstacles they all face in adopting AI – security, explainability, and domain expertise. Large language models (LLM) are impressive generalists, but that’s very different from having real depth in banking. Titan is built to solve all three.
Our core insight was that if we built smaller models trained exclusively on banking, rather than routing everything through a general LLM, we could get those models to reason the way bankers do.
Our team is well versed in technology and banking as five of our first eight employees are former or current CTOs, each having operated inside a bank. There’s a real difference between reading about a second-line review and actually living through one. That lived experience runs throughout the company, and it shows up in all we do.
Most discussions around AI in banking focus on data quality. You’ve argued the bigger issue is process fragmentation. What do you see as the most misunderstood barrier to successful AI adoption in financial institutions today?
Most bankers assume success with their AI solutions comes down to picking the right model or vendor, and that the biggest challenge will be the state of their data. Data is certainly an important challenge, but the banks getting the most effective outcomes learned quickly that AI doesn’t only struggle with unreliable data, it struggles when processes are fragmented and inconsistent because AI can’t create clarity from operational ambiguity.
Most banking data lives across dozens of systems, making it unstructured, disparate and complex. The biggest challenge for institutions is to understand the process in how their data is utilized, by who, at what stage and for what purpose, and why process and data readiness are the real prerequisites for AI success. Finally, banks must be willing to refactor their processes to be AI-native – as opposed to deploying AI in outdated and fragmented processes.
Many banks are still stuck in pilot mode with AI. What typically breaks when they try to move from experimentation to real operational deployment at scale?
Most AI platforms today share the same fundamental flaw: they can generate answers, but they can’t truly explain them. You get an output, but not a defensible record of how that output was produced. In banking the standard isn’t whether an answer is helpful, but whether it can be justified, reconstructed, and defended under scrutiny. This is where most AI adoption breaks down and pilots falter. Banks operate under strict model risk management expectations, including frameworks, where every model-driven decision must be transparent, traceable, and auditable. As AI becomes subject to these same standards, institutions need to do more than generate outputs. They need to be able to explain them in a way that holds up to compliance review, internal audit, and regulatory examination. Without that, AI introduces risk rather than reducing it.
You’ve spoken about “Safe at Speed” as a new playbook for regulatory-ready AI. What does “safe innovation” actually look like in practice for community financial institutions without slowing down transformation?
As community financial institutions move from evaluating AI to actual deployment, a major concern is how to innovate safely without slowing down and without triggering regulatory concerns. Most institutions don’t have the luxury of large AI teams or massive tech budgets. It’s also why one-size-fits-all AI models often fall short. Regardless of asset size, banking is full of both complexity and nuance; policies, procedures, exceptions, regulations, and judgment calls that aren’t fully captured or always contemplated in the generic datasets that today’s LLMs leverage.
When AI is designed with banking-specific ontology, trained by bankers, regulators and financial lawyers, and grounded in real operational workflows, it can scale what institutions’ best operators already do. Tasks like complex search and retrieval, and stare and compare data verification across systems and policies, or first-pass reviews in areas like deposit account opening or risk operations are ideal starting points. This approach amplifies expertise, with the AI learning from institutions’ strongest performers and applying that knowledge consistently, allowing teams to focus on higher-value work.
Instead of fearing regulators, you suggest training AI systems to think like them. How does that philosophy translate into how Titan is designed and deployed in real banking environments?
Many bankers feel that AI represents an entirely new regulatory framework just waiting to catch them off guard. But regulators are approaching AI the same way they approach any new technology: by seeing how existing rules and guidance may apply, and in the case of AI as we saw recently with their guidance on SR26-2, being willing to change when old rules do not apply.
Security, model risk management, third-party risk, data governance, and human-in-the-loop oversight aren’t anything new in banking technology. It’s only the technology itself that’s the new part. The institutions that are having a hard time are those that stood up AI initiatives in isolation — separate from risk, compliance, and the operational rigor that governs every other critical system. The reality is they should be planning these systems so that they naturally map to how they operate and, just as importantly, to how their examiners already think.
Institutions that engage with their examiners early on in their AI implementation, document all their choices and reasoning, and approach their strategies with an eye towards auditability and supervision from the start will often find regulators to be collaborative partners rather than adversaries. When examiners ask to see an AI framework, the goal is to demonstrate it’s already being managed with the same rigor as any other critical banking system, like credit models, core systems or BSA operations.
With over 60% of bank employees already using public AI tools, what are the most critical risks you see emerging, especially around data leakage, compliance gaps, and lack of auditability?
This is one of the main reasons Titan was born. The truth is the majority of banking employees are already using generic AI tools and LLMs. What’s more, they’re often entering sensitive data into them with no clear guardrails nor auditability.
Severe data leakage and confidentiality breaches are among the most immediate risks. More than 20% of files uploaded to public AI tools are estimated to contain confidential information, contributing to active data security incidents. Bypassing security controls results in leaving sensitive banking data in unregulated public environments, creating significant privacy and regulatory exposure. These gaps also widen as unclear ownership of AI-generated decisions makes it harder for institutions to demonstrate accountability for examiners. This lack of auditability and governance, which are critical requirements in highly regulated environments, further increases risk for institutions.
Titan is described as a banking-native AI platform. What does “bank-native” mean in technical and operational terms, and how is it fundamentally different from adapting general-purpose AI tools?
Bank-native AI is the difference between a model that provides a response and one that truly understands the questions from a banking perspective. Generic, general-purpose AI can draft emails or summarize text, but it doesn’t know the nuances of navigating a consumer lending decision, why a first-day letter sends teams scrambling, or why your second-line may be requesting changes to your deposit account opening process. It also fails to produce the documentation needed to stand up to an audit or regulatory exam.
Titan is the first banking-native AI platform purpose-built for the operational, regulatory, and productivity realities of financial institutions. The distinction starts at the model level. General-purpose AI is optimized across thousands of domains — fantastic broad capabilities at the cost of banking precision. Standard fine-tuning adjusts how a model responds but doesn’t change what it actually knows about regulatory relationships or the institutional logic of banking compliance. We take a structurally different path: our models are trained on Titan’s Banking Ontology and from the ground up on enforcement actions, examination findings, and supervisory correspondence — the documents that define how regulators actually interpret rules, not just what the rules say. Every output is anchored to a specific regulatory citation, and regime-tagging ensures the model identifies the governing regulatory framework before generating an answer, which is what eliminates the hallucinated citations that plague general-purpose tools in high-stakes banking contexts.
We combine three integrated layers into a single banking-native platform. This includes a secure “front door” that replaces shadow AI with governed access, explainability through full audit trails, PII controls and data-loss prevention; our own banking models, purpose-built and trained with former regulators, bank operators, and financial lawyers to produce policy- and examiner-ready outputs. These outperform leading general-purpose LLMs by 30–80% on banking tasks with materially lower hallucination rates. Finally, our banking agents are configurable human-in-the-loop colleagues that automate repeatable workflows across risk, compliance, underwriting, and operations, keeping every decision traceable and explainable.
You emphasize that every agent decision in Titan is readable and explainable. Why is explainability so essential in banking AI, and how do you balance it with model performance and speed?
Regulators aren’t going to accept a “black box” when it comes to any AI decisioning. One of the biggest misconceptions is that institutions can move fast now and “bolt on” explainability once regulators start asking questions. The reality is not only is this untrue, but it is also actually the opposite. Bottom-line – if an AI system can’t transparently explain how it reached a decision, it can’t be trusted. And if it can’t be trusted, it will never scale effectively across an institution.
The majority of banks and credit unions already know this from experience. They understand examiners don’t just want your answers; they want to see your work, too. When AI systems are built to show their work, to document their data sources, reasoning steps, and outcomes, they become deployable across far more use cases, lines of business, and users, and with much greater confidence.
Finally, what advice would you give to banking executives and transformation leaders who understand the urgency of AI but are still hesitant about operationalizing it at scale?
I have empathy for where these leaders sit – banking is one of the most regulated industries on the planet. Hesitation isn’t irrational. But when hesitation becomes paralysis, the cost of not moving becomes just as real as the cost of moving wrong.
Stop waiting for a perfect AI strategy before you start. The institutions winning right now began with one high-value, well-scoped problem, got it working, measured it rigorously, and then scaled. Clarity comes from doing, not over-engineering.
Governance is your accelerator, not your brake. Get risk and compliance in the room from day one as co-owners of AI – not gatekeepers – and you’ll have an internal pathway instead of a bottleneck.
And bring your people along. A lot of the hesitation I see isn’t really about the technology – it’s about change management. Re-skill your ops teams. Give your relationship managers AI assistants.
The urgency you feel is correct. Act on it – just act with intention.
