# Five AI Questions We're Hearing from Financial Services Leaders > Source: > Published: 2026-09-09 18:09:29+00:00 The AI, governance, data, and cost questions worth asking at Sibos in Miami. by [Junta Nakai](/blog/author/juntanakai), [Erin Butler](/blog/author/erin-butler), [Roshni Joshi](/blog/author/roshni-joshi), [Antoine Amend](/blog/author/antoine-amend), [Jennifer Miller](/blog/author/jennifer-miller), [Andrea DeSosa](/blog/author/andrea-desosa), [Rajaram Suresh](/blog/author/rajaram-suresh), [Kim Hatton](/blog/author/kim-hatton), [Naeem Rehman](/blog/author/naeem-rehman), [Spencer Cook](/blog/author/spencer-cook), [Ed Bullen](/blog/author/ed-bullard) and [Felix Mutzl](/blog/author/felix-mutzl) Last year at Sibos Frankfurt, the question was whether AI works. This year: **can your AI earn your trust?** The data backs up that shift. In Grant Thornton's 2026 Banking Insights AI Impact Survey, half of banking executives said governance and compliance were already limiting AI performance, yet only 18% said they were confident they could pass an independent audit of their AI controls. Our own research with Economist Enterprise, the 2026 report [Making AI Deliver](https://www.databricks.com/resources/analyst-research/making-ai-deliver), found a similar gap for agentic AI specifically: fewer than half of organizations formally mandate a governance framework for autonomous systems. These are the five questions we hear most from chief financial officers (CFOs), compliance officers, treasury leads and relationship managers. They're worth asking whether or not you're headed to Sibos this year. "What's our real liquidity position right now, across every entity and currency?" That's the question a treasury officer asks every day. It's the one AI should finally answer without a three-day spreadsheet sprint. Getting a model to summarize a report is straightforward. The harder part is putting governed data and AI into the decisions finance teams make every day across the balance sheet, liquidity, spend, and operations. The goal is to help teams spot changes, investigate what is driving them, and do more than automate the reporting cycle. At the Databricks booth (**#DISL25**) at Sibos Miami, we'll present a talk, [Trapped Capital, Released by AI: Reconciliation Your Regulator Can Audit](https://www.databricks.com/lp/event/sibos-miami-2026). It finds the GL-to-Risk data breaks that silently inflate risk-weighted assets, investigates them with cited evidence, and releases trapped regulatory capital, with a human maker-checker in control and an audit trail an examiner can actually follow. We'll also dig into what comes after the money moves. Tokenized settlement rails solve part of the problem, but reconciliation, financial-crime controls, and intraday liquidity management still have to happen off-chain. Our lightning talk, [Tokenization Moved the Money. It Didn't Move the Hard Part](https://www.databricks.com/lp/event/sibos-miami-2026), walks through the operating model for keeping token positions, deposit sub-ledgers, and funding accounts continuously aligned. Plus you will get a live look at a tokenized-deposit reconciliation control plane. "Is this alert real risk, or noise?" That's the question an AML analyst asks dozens of times a day, and it's the one that separates AI that actually helps from AI that just adds another system nobody trusts. AI can investigate cases, connect context and support analyst decisions. But every recommendation needs a traceable source: what data fed it, what it recommended, whether a person reviewed the output and whether the institution can reconstruct the investigation for an audit or regulatory review. Governance can't live only in a policy document. It has to be built into the data, access controls and workflow itself. We'll have a live version of this at our booth: an agent working real AML alerts on the bank's own lakehouse, seeing only what the analyst is cleared to see. It drafts the Suspicious Activity Report (SAR) and leaves the filing to a human, with every step logged against the know-your-customer (KYC)/AML evidence and the model that ran. It's the same question that we'll cover in our Discover Stage session, [How financial institutions are governing AI in financial crime operations](https://www.databricks.com/lp/event/sibos-miami-2026), on Thursday, October 1, 11:00 AM, proving what the AI did, not just what it can do. "What does this client actually need from us right now?" Every relationship manager asks some version of that before a call, and the honest answer usually depends on data scattered across three systems and a research folder they didn't have time to read.[RBC Brewin Dolphin](https://www.databricks.com/customers/rbc-brewin-dolphin) uses generative AI to automate preparation for client-review meeting packs, saving an estimated 4,700 hours annually. Other firms are taking this further. At one global asset manager, multi-agent workflows generate investment commentary across hundreds of funds, extending advisor coverage while compliance stays in the loop. At another, agents blend proprietary fund scoring with internal documents to produce audit-ready portfolio analyses in seconds, surfacing opportunities a team wouldn't have had time to find on their own. The goal isn't removing people from high-value decisions. It's reducing the work around those decisions so more time goes to judgment and the relationship itself. "Why did we just decline a good transaction?" That's what a head of transaction banking asks when risk decisions run on data that's minutes, not milliseconds, old. Most institutions already have the data. It's just spread across systems, delayed by batch processes or hard to use in a governed way. [Coinbase](https://www.databricks.com/customers/coinbase/lakeflow) runs real-time fraud detection on Databricks with sub-100ms P99 latency, a useful benchmark for what "real-time" actually needs to mean. AI economics look manageable in a pilot and very different in production. On a recent Bloomberg CFO Briefing, one CFO called AI token expense a "trivial number" for the first half of 2026, while noting that not every task needs the most expensive model. That won't stay trivial for long, at that bank or anywhere else. Cost discipline needs to start before spend becomes hard to track, not after. For a closer look at how engineering teams are already solving this problem, see Databricks' blog post, [Managing AI Coding Costs at Scale](https://www.databricks.com/blog/managing-ai-coding-costs-scale), which breaks down the levers, from model routing to spend visibility, that keep AI costs from outrunning the value they create. If these are the conversations your team is already having, we'd rather have them in person. Databricks will be in Miami for Sibos 2026, September 28–October 1. Email [sibos@databricks.com](sibos@databricks.com) or visit our [event page](https://www.databricks.com/lp/event/sibos-miami-2026) to see the full agenda and to schedule a meeting. Subscribe to our blog and get the latest posts delivered to your inbox.