(0:11) AI promises speed, accuracy, cost savings, risk reduction. (0:16) And in financial services, well, that promise is really strong for compliance, especially with regulations constantly changing. (0:22) So our mission today is to cut through that hype.

(0:25) We want to really understand the challenges, the nuances of making AI work effectively in this very demanding field.

(0:31) Yeah. (0:31) And it’s so important because financial compliance, it’s tough. (0:35) It’s always shifting.

(0:35) Regulations change, there’s constant pressure to cut costs, but stay compliant, perfectly compliant. (0:40) So the goal, using AI for efficiency, makes perfect sense. (0:43) It’s needed.

(0:44) But what we’re seeing is the way firms integrate AI often, well, it falls short. (0:50) It can create more problems than it solves sometimes.

(0:51) Okay, let’s unpack that. (0:52) So firms are eager, they want AI. (0:55) Our sources are saying AI is doing some useful things already, right?

(0:58) Automating document reviews, fraud detection reporting. (1:01) We’re seeing things like natural language processing, NLP, scanning docs for regulatory keywords or issues, even learning from past mistakes.

(1:10) That’s right. (1:11) NLP can flag risky language, deviations, things that need a human eye. (1:16) Some systems do get smarter over time.

(1:18) But there’s always a but, isn’t there? (1:20) Where’s the first big hurdle?

(1:22) It’s often something fundamental that gets overlooked, the workflow. (1:26) People get so focused on the shiny AI tech itself, they forget compliance isn’t just about finding errors. (1:32) It’s a process, a really complex one.

(1:34) It involves structured steps, different teams, approvals. (1:37) It’s a whole system.

(1:38) Ah, okay. (1:39) So the AI needs to fit into that system.

(1:41) Exactly. (1:41) If the AI isn’t deeply integrated into a strong, flexible workflow, it might look good in a demo, but it doesn’t actually help day-to-day. (1:49) You get these features that sort of sit on the side disrupting things instead of making them smoother.

(1:53) Right. (1:53) So you’re forcing your process to fit the AI, not the other way around.

(1:57) Precisely. (1:58) And that’s the key question. (1:59) Is the AI working with your team in process, or is it this separate thing you’re constantly trying to work around?

(2:06) If it’s the latter, you’re probably creating complexity, not efficiency.

(2:09) That makes sense. (2:10) Workflow first. (2:12) Okay, what’s next?

(2:13) I remember reading about data. (2:15) This idea that you crane the AI once and you’re set. (2:17) That sounds too easy.

(2:18) Yeah, that’s the next big trap. (2:20) The one-size-fits-none training model.

(2:23) Meaning?

(2:23) Well, many AI models are trained on data from just one specific company, which sounds great for that company maybe, but every firm is unique. (2:31) Different rules, different risks, different ways of operating, even different language sometimes.

(2:36) Okay.

(2:36) So you take a model trained for firm A and try to apply it to firm B. (2:40) It’s often going to make mistakes. (2:41) It might miss things or flag things incorrectly, false positives, false negatives.

(2:45) Because it doesn’t understand firm B’s specific context.

(2:48) Exactly. (2:49) And retraining isn’t simple. (2:51) It’s not a quick fix.

(2:52) Our sources say it often takes like hundreds or even thousands of documents, properly labeled documents.

(2:58) Wow. (2:59) That sounds intense.

(3:00) It is. (3:01) It’s hugely time-consuming, needs a lot of resources. (3:05) Most firms just don’t have the bandwidth for that kind of ongoing large-scale retraining effort.

(3:10) So these models aren’t very adaptable or scalable then?

(3:13) They can be quite brittle, yes. (3:15) It really highlights this need to move away from models that need tons of bespoke training data towards AI that’s inherently more adaptable. (3:28) Flexible intelligence, you could say.

(3:30) Okay. (3:30) So workflow integration, adaptable models. (3:33) What else can go wrong?

(3:35) Is there something else lurking?

(3:36) There is. (3:37) And it’s often hidden. (3:38) It’s the foundation some of these AI tools are built on.

(3:40) We’re talking about the underlying workflow engine itself.

(3:43) The engine under the AI.

(3:44) Yeah. (3:44) Many AI vendors, especially when they’re trying to get to market quickly, use third-party workflow engines. (3:49) Sometimes they’re white-labeled, so it looks like their own tech, but it’s actually a license from someone else.

(3:54) And the problem there is…

(3:56) Flexibility. (3:57) Or rather, the lack of it. (3:59) These engines often weren’t designed for the kind of complex, multi-layered approval processes you see in financial compliance.

(4:05) Think about it. (4:06) Compliance teams manage approvals across different departments, handle all sorts of submission types. (4:11) Marketing, internal comms, you name it.

(4:14) Right. (4:14) It gets complicated fast.

(4:15) It does. (4:16) And if that underlying engine can’t easily adapt to the firm’s specific structure and rules, you get delays, friction, frustration. (4:23) People start creating workarounds.

(4:24) Which introduces risk.

(4:25) Directly increases regulatory risk. (4:28) So the AI could be brilliant, theoretically, but if it’s running on a rigid, clunky foundation, its potential is seriously capped. (4:35) It’s like putting a race car engine in a tractor.

(4:37) It just won’t perform properly.

(4:39) Okay, this sounds like a minefield. (4:41) Neglected workflow, rigid training, inflexible engines. (4:45) It feels a bit discouraging.

(4:47) Is anyone actually finding a better way through this?

(4:49) Well, innovation is happening. (4:51) Our sources pointed towards approaches like Red Oak’s AI review module as an example of taking a different path, trying to avoid these specific traps.

(5:00) Red Oak, how are they doing things differently?

(5:02) Their core idea seems to be leveraging large language models, LLMs, the big foundational AI models like Chattopadhyay or Claude, combined with something called prompt engineering.

(5:14) Prompt engineering.

(5:15) Yeah. (5:15) So instead of needing to build custom models trained on thousands of your past documents, you use prompt engineering. (5:21) You basically write very specific, detailed instructions, the prompts, to guide the general intelligence of the LLM.

(5:29) You teach it your specific compliance rules, what to look for in marketing materials, for example. (5:34) It’s like giving a very smart assistant a detailed custom rule book for your company. (5:39) Much faster, much more agile.

(5:41) Interesting. (5:41) So how does that specific approach address the pitfalls we talked about? (5:45) Let’s start with workflow, the first hurdle.

(5:47) Right. (5:47) The key difference is that their AI is designed to work inside their flexible workflow engine from the start. (5:53) It’s not a bolt-on.

(5:54) It routes things, flags issues, guides the document to the right human reviewer, all within the flow. (6:00) It’s meant to be seamless, preventing those bottlenecks.

(6:03) Okay. (6:04) Integrated workflow. (6:05) Check.

(6:06) What about the data training nightmare, needing thousands of documents?

(6:10) That’s where the LLM and prompt engineering really makes a difference because you’re guiding a pre-existing powerful model with specific instructions. (6:17) You don’t need all those historical documents for training.

(6:20) Ah, so setup is faster, less complex.

(6:23) Exactly. (6:23) You define your rules, write your prompts, and the AI adapts much more quickly. (6:28) It makes it easier to implement and adjust across different clients.

(6:31) It directly tackles that one company model problem by being inherently more flexible.

(6:36) And that third issue, the inflexible underlying engine, being locked into something rigid.

(6:41) They seem to address that with flexibility too. (6:43) Firms can choose to use their own instance of an LLM, like their own secure version of ChatGPT or Claude, or they can use Red Oak’s version. (6:51) This gives firms more control, helps with the decority concerns potentially, and ensures it fits better into their existing tech setup rather than being this isolated system.

(7:00) It avoids relying on those potentially restrictive white-labeled engines.

(7:04) Got it. (7:05) So beyond the technical integration, what’s the practical day-to-day win for the compliance team or the person submitting the document?

(7:13) The big win is faster feedback, right at the start. (7:17) The AI can flag potentially problematic language, misleading words, guarantees, things like that, before a human reviewer even sees it.

(7:26) Oh, interesting. (7:26) So the submitter gets a heads up.

(7:28) Yeah. (7:29) They get a chance to fix obvious issues up front. (7:32) This speeds up the whole cycle massively.

(7:34) It means your expert human reviewers aren’t bogged down with simple repetitive checks. (7:38) They can focus their time on the really tricky, nuanced stuff that genuinely needs human judgment.

(7:44) Okay, so let’s pull this all together. (7:46) We’ve looked at the huge potential AI has in financial compliance, but we’ve also dug into some really common pitfalls. (7:53) Ignoring workflow, getting stuck with rigid training models that don’t scale, and hitting the limits of inflexible underlying tech.

(7:59) And then we saw how different approaches like Red Oak using LLMs and prompt engineering are trying to navigate around these by focusing on workload integration, adaptability, and cutting down that heavy training requirement. (8:10) It’s definitely a complex space.

(8:12) It absolutely is. (8:13) And it points to a bigger shift, I think. (8:15) The source mentioned Red Oak’s module is still in beta, launching more widely later in 2024.

(8:21) That just shows how fast this is all moving. (8:23) So the final thought for you, the listener, is really this. (8:27) When you’re evaluating AI solutions for compliance, you need to ask a critical question.

(8:32) Is this technology forcing me to bend my established complex processes to fit its limitations? (8:37) Or is this tool genuinely flexible enough to adapt and integrate into my specific way of working? (8:44) Because the smartest AI isn’t much good if it doesn’t actually fit your reality and help your team where they need it most.

(8:50) That’s definitely some crucial food for thought. (8:52) Are you looking for tools that force you to change, or tools designed to adapt to you? (8:56) Consider that as you navigate the AI landscape in financial services.

(9:00) Thanks for joining us on this deep dive. (9:01) We hope it’s given you a clearer picture of the promises and the pitfalls.