AI Agents for Compliance and Customer Support

The newer models (e.g. GPT o1, o3, Claude 3.5 Sonnect, etc) have really wowed everyone with their reasoning capabilities. Their ability to solve math/coding problems has been amazing. This got us reflecting on the different types of intelligence. The majority of recent research has gone into 'reasoning' intelligence. While helpful at a high level, this doesn't make that big of an impact on customer-facing AI agents. Our agents don't need to be able to write complex algorithms. Instead, we're much more interested in the 'instruction-following' type of intelligence. In front of millions of individual customers, the real challenge is making the agent follow instructions perfectly every time. At Decagon, our agents handle tasks traditionally done by customer support reps—answering tickets, helping customers troubleshoot, or handling basic admin tasks. To achieve the best results, we must ensure our agents provide precise, reliable, and customizable answers, just like our customers expect from their best human CX agents. There's no room for errors when a customer relationship hangs in the balance. Therefore, a lot of our research work is geared toward optimizing the ability of our agents to follow instructions, which can often be complex and multi-layered. Why? → They guarantee accuracy  Customers in regulated industries have restricted information the AI can never say. Tight control over rules and outputs is non-negotiable. → They allow customization Every customer has a unique tone, brand, and requirements. Great AI agents adapt seamlessly to reflect the customer’s voice. → They make our agents trustworthy Without perfect instruction-following, mass adoption is impossible. Companies, specifically those in sensitive categories, need confidence that their AI agent will behave exactly as intended. As models continue to improve, the ability to follow instructions will only become superior, amplifying the AI agents’ capabilities—whether that’s adhering to brand guidelines or ensuring compliance in sensitive industries. At Decagon, we’ve seen firsthand how instruction-following drives the best outcomes for our customers. So, throughout 2025 we’ll continue to be laser-focused on this form of AI to deliver the most advanced CX product for our customers. If you have any questions about how our agent works, shoot me a DM. Happy to explain further.

Today we’re presenting the findings from our clients using AI agents, in production for 3 months. We can cut customer wait times stuck in a KYC/sanctions queue from 20 days to 2 minutes. This is a huge unlock for anyone onboarding customers. “Compliance Officer” is the 5th fastest growing occupation in the United States! Major banks average 307 employees just for KYC alone, yet can't hire more compliance officers fast enough. More than headcount, this costs customers and revenue. We deployed AI agents in production environments at multiple financial institutions for 3+ months and show AI Agents can meaningfully improve KPIs: - For one FI, the daily backlog was 14 hours and they couldn't keep up with it.  - So the backlog kept growing  - As did the average customer wait time stuck in a queue, to 20 days. Using Agentic AI, we were able to  - Automate majority (95%) of the cases and  - bring down daily backlog to 41 minutes (from 14 hours).  - Most importantly, avg customer wait time went down drastically to 2 minutes. Perhaps the most counterintuitive finding. Agentic AI when trained and deployed according to our framework, can be more accurate than humans. We found AI agents follow operating procedures in 100% of cases vs <95% for humans. Humans never follow SOP to the minute details and with rote work, they are more error prone. FI's rightly worry, what about hallucination? What about data privacy? Will the regulator allow it These live, production data points are all within existing regulatory frameworks (SR 11-7 compliant). Our Agentic Oversight Framework maintains complete human accountability while delivering: - Alignment to Standard Operating Procedures (SoPs) - A full audit trail of every data element accessed - A full, explained decision rationale, reviewed before every case is progressed - Continuous learning from expert reviewers - Automated drift detection and safeguards The white paper is a playbook for how financial institutions can safely implement agentic AI while fully complying with regulatory requirements. Real results. Real institutions. Real transformation. You might ask: what is AI about all of this and how's it different from ML and rules based systems. In short, rules systems are rigid but Agentic AI can adapt. All those details in the white paper:

Hot take: Agents that can do "everything" are useless. We need to be focusing on 𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗶𝘇𝗲𝗱 𝗮𝗴𝗲𝗻𝘁𝘀 that enterprises actually need, and can deploy today. AI agents might be the next step beyond basic generative AI, but without RAG, they're just expensive guessing machines. 𝗥𝗔𝗚 𝗴𝗿𝗼𝘂𝗻𝗱𝘀 𝗮𝗴𝗲𝗻𝘁𝘀 𝗶𝗻 𝗿𝗲𝗮𝗹-𝘄𝗼𝗿𝗹𝗱 𝗱𝗮𝘁𝗮, allowing them to safely interact with internal documents, retrieve key information, and cite actual regulations and policies. Our new ebook with StackAI and Weaviate breaks down three production use cases that we've seen enterprises get actual value out of: 1️⃣ 𝗥𝗙𝗣 𝗗𝗿𝗮𝗳𝘁𝗶𝗻𝗴 𝗔𝗴𝗲𝗻𝘁 Enterprises waste enormous time reconstructing institutional knowledge during proposal cycles. This agent retrieves relevant historical proposals, reference architectures, and past win narratives from OneDrive, then assembles a first-pass draft aligned to past successes. It generates technical proposals, business proposals, formats everything into a cohesive document, saves it to SharePoint, and notifies reviewers via Outlook. 2️⃣ 𝗦𝗮𝗳𝗲𝘁𝘆 𝗮𝗻𝗱 𝗖𝗼𝗺𝗽𝗹𝗶𝗮𝗻𝗰𝗲 𝗔𝗴𝗲𝗻𝘁 Traditional compliance chatbots hallucinate, which is unacceptable in regulated environments. This agent grounds every response in approved regulatory documents, internal policies, and audit frameworks. It reviews engineering specs against internal guidelines, produces compliance assessments, generates executive summaries for non-technical stakeholders, and delivers formatted reports. 3️⃣ 𝗧𝗿𝗶𝗮𝗴𝗲 𝗮𝗻𝗱 𝗥𝗲𝘀𝗽𝗼𝗻𝘀𝗲 𝗔𝗴𝗲𝗻𝘁 Claims workflows depend on precise interpretation of coverage rules and procedures. This agent extracts claim details from incoming emails, classifies urgency into four tiers, routes accordingly, and generates appropriate responses - all grounded in official policy text and operational guidelines. These workflows are operational systems handling actual business use cases with proper permissions, auditability, and human-in-the-loop review where needed. 2026 is the year we move from "look what AI can do" to "here's how AI reliably does this every day." 𝗦𝘁𝗮𝗰𝗸𝗔𝗜'𝘀 𝗻𝗼-𝗰𝗼𝗱𝗲 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗰𝗼𝗺𝗯𝗶𝗻𝗲𝗱 𝘄𝗶𝘁𝗵 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲'𝘀 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗹𝗮𝘆𝗲𝗿 𝗺𝗮𝗸𝗲𝘀 𝘁𝗵𝗶𝘀 𝗮𝗰𝗰𝗲𝘀𝘀𝗶𝗯𝗹𝗲 𝘁𝗼 𝗮𝗻𝘆𝗼𝗻𝗲 - you can build these workflows visually, connecting to the tools you already use (SharePoint, Outlook, Gmail, etc.) without creating custom code. Download the full ebook here: https://lnkd.in/dFpmSgX6

Just wrapped up a fascinating deep dive with the PwC team today, and I'm deeply pleased with what I saw in agent OS and how well PwC and AWS can work together to help customers. You know that moment when you see a solution that just clicks? That's what happened today. PwC has cracked something that's been driving enterprises crazy: how do you get AI agents from different vendors to actually work together? Here's what blew me away: watching their Agent OS work our AI stack. You've got agents leveraging S3 for persistent memory, OpenSearch powering knowledge retrieval, and Bedrock models like Claude and Nova doing the reasoning. PwC's orchestration layer ties it all together with MCP servers, making every AWS service callable from agent workflows. But here's where it got interesting – we talked about agents that spawn other agents dynamically. A customer support agent realizes it needs compliance expertise, spins up a compliance agent on AgentCore, pulls context from S3, and hands off seamlessly. We started sketching scenarios and I kept thinking "this is how enterprises will actually use AI." What makes this partnership promising: PwC brings that enterprise AI command center connecting agents 10x faster, plus decades of understanding how industries work. They know healthcare agents need HIPAA handled differently than financial agents need SOX compliance. They've been helping enterprises transform core processes long before AI. We bring infrastructure that scales securely. Together, we're giving customers a path to bet their critical processes on AI, with governance frameworks that turn "cool demo" into "this runs my business." The use cases this will unlock for customers adopting agents are genuinely exciting. Truly transformative. This feels like enterprise AI growing up. #AI #EnterpriseAI #AWSBedrock #AgentCore #AIAgents #PwC Matt Wood Jacob Wilson Scott Cook

AI Agents in Banking: The Truth No One Talks About Over the last 12 months, I’ve built and deployed 50+ custom AI agents across tier-1 banks and financial institutions. And here’s the truth—what actually works in banking is very different from what most people are selling online. Forget the flashy “$50K/month with no-code AI agents” headlines. In reality, banks aren’t buying dreams. They’re investing in precision, reliability, and measurable ROI—with strict compliance guardrails in place. The most successful AI agents I’ve built don’t try to do everything. Instead, they focus on solving one high-impact problem exceptionally well, such as: 🔹 KYC automation – Extracting and verifying documents, cutting manual review time by 60% 🔹 Fraud detection – Real-time transaction monitoring that reduces false positives by 40% 🔹 Customer service AI – Handling up to 70% of routine inquiries, boosting CSAT and reducing ops cost These agents aren’t built for show. They’re built for scale. They integrate cleanly with legacy systems, follow strict audit trails, and pass scrutiny from compliance and legal teams. Most importantly, they drive outcomes that matter—time saved, risk reduced, and customer satisfaction improved. In the world of banking, flashy doesn’t cut it. Real innovation is quiet, consistent, and measurable. If you’re working on AI for financial services, focus less on what’s trending—and more on what truly moves the needle. #AI #BankingInnovation #AIagents #Fintech #Compliance #RiskManagement #KYC #Automation #RegTech #FraudDetection #CustomerExperience

We just hit 80% autonomous resolution at 99.8% accuracy. But here's what nobody tells you about building AI support that actually works: The architecture matters more than the model. Most AI support tools are built like this: → Customer asks question → System searches your docs → AI writes an answer from what it found Sounds logical. Works for FAQs. Breaks completely when real money, real risk, or real regulations are involved. We learned this the hard way at Fini. We started like everyone else - RAG-based system, solid LLM, good prompts. Our bot sounded smart. Customers liked talking to it. But our customers couldn't actually trust it with their queue. Because finding the right document ≠ making the right decision. Real support tickets aren't "What's your refund policy?" They're "I was charged twice after upgrading mid-month, one in euros and one in dollars, fix it and downgrade me back." That requires: • Checking actual transactions across systems • Applying granular rules (region, plan type, contract term) • Executing actions correctly • Documenting what you did and why We rebuilt everything from scratch. No more doc search at answer time. Instead: compressed, structured knowledge layer that represents your entire business. Not fuzzy matching - deterministic workflows you can trace and audit. We call it RAGless. The result? Fini now resolves 80% of issues autonomously. CSAT up 10%. Support costs down 50%. More importantly: our customers in fintech and insurance can actually show compliance teams why the AI did what it did. If you're a CX leader and your AI sounds smart but can't be trusted with real work, the problem isn't your prompts. It's your architecture. We wrote the full technical breakdown in our blog: (link in comments)

Came back from vacation Monday. Inbox? On fire.🔥 Buried in the chaos: a customer story that stopped me in my tracks (and made me so happy). A Customer Support leader at a fast-growing financial services company used AI to transform his team - in just a few weeks. This leader works for a financial services company that’s in high-growth mode. Great news, right? Yes! For everyone except his Customer Support team… As the business grew faster, they were bombarded with repetitive questions about simple things like loan statuses and document requirements. Reps were overwhelmed. Customers faced longer response times. The company has been a HubSpot customer for nearly 10 years. They turned to Customer Agent, HubSpot’s AI Agent, and got to work: - Connected it to their knowledge base → accurate, fast answers - Set smart handoff rules → AI handles the simple, reps handle the complex - Customized the tone → sounds like them, not a generic bot (you know the type) In a short space of time, things changed dramatically: - Customer Agent now resolves more tickets than any rep - 94.9% of customers report being happy with the experience - For the first time, the team can prioritize complex issues and provide proactive support to high-value customers It’s exciting to see leaders using Customer Agent to not just respond to more tickets, but to increase CSAT and empower their teams to drive more impact. 2025 is the year of AI transformed Customer Support. I am stunned by how quickly that transformation is playing out!

Whilst Agentic AI is currently massively overhyped as the latest new shiny thing, it was good to see Bank of Singapore, Asia's Global Private Bank quietly publishing how they are applying Agentic in production to solve complex real-world compliance problems. Performing the Due Diligence on new Private Banking customers is an extremely time-consuming activity - requiring up to 40 different documents to be sourced from customers to understand their Source Of Wealth (SoW). Many days of effort are spent by bank staff to make sense of these documents, extract key information, benchmark the data and then draft a lengthy Source-of-Wealth memo justifying why a customer should be onboarded. Our Source of Wealth assistant allows front-line staff to simply drag & drop the relevant documents within CRM & have the Agentic AI assistant take over the process.   This was not an easy challenge – requiring multi-modal document intelligence including LLMs and VLMs to parse, validate and summarise a complex array of documents. Complex Agentic workflow was then required to handle the many different types of SoW workflows that can exist depending on client profile. Finally, extremely robust guardrails and evaluations were required to ensure accuracy and trustworthiness in such a critical process. Good AI is a team sport. Success wouldn’t have been possible without strong partnership across Technology, Business and Data Office to make it work. Too many to name but including  Chin Wong KAM, CAMS , Kelvin Chiang , Bryan Lee , Yixiang Teng , Srinivasan Thangamani, Kenneth Z. and Adrien Chenailler. Hard problems. Real-world benefits. It’s not about shiny toys.   Get the detail at: https://lnkd.in/gZtWhTuY

🛡️ Exploring the Intersection of Agentic AI and Compliance in Financial Services In the realm of financial services, the prevalence of false positives in KYC and sanctions screening, sometimes exceeding 90%, poses challenges such as burnout, onboarding delays, and regulatory risks. What if AI agents could revolutionize compliance practices while simultaneously enhancing oversight? Introducing Sardine's innovative Agentic Oversight Framework (AOF), a beacon for regulated financial institutions seeking to integrate AI agents securely. This framework goes beyond mere task automation, emphasizing precision, transparency, and auditability in operations. Key Features of the AOF:- ⏱️ Drastic reduction of KYC backlogs from 14 hours to a mere 41 minutes. 🧠 AI agents achieving a remarkable 100% precision rate in approved onboardings (complemented by human review. 🔁 Implementation of continuous feedback loops for ongoing performance enhancement. 🧾 Integration of audit trails, explainability, and governance measures. 🔐 Tailored for SR 11-7 and BSA/AML compliance standards. Rather than supplanting human roles, these AI agents function in a collaborative "copilot" capacity, augmenting both primary and secondary reviews. The AOF empowers AI agents to not only be potent but also accountable in their actions. The adoption of this framework enables financial institutions to expedite scalability, enhance customer experiences, and concurrently meet regulatory expectations. Delve deeper into the comprehensive insights provided in the full paper. #AgenticAI #ComplianceTech #AIinBanking #RegTech #KYC #AML #ResponsibleAI #FinTech #AIgovernance #AuditTrail #AIoversight