CreditAssociates Deploys Agentic AI to Boost Sales Conversion and Retention

Using Agentic AI to Drive Customer Acquisition & Retention With CreditAssociates — Ai4

Ai4

CreditAssociates links its AI strategy to a unified data foundation and explainable lead recommendations to capture millions in value from a single percentage point gain in client retention.

AI-assisted summary of Using Agentic AI to Drive Customer Acquisition & Retention With CreditAssociates — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published by Ai4 on YouTube on August 15, 2026, this presentation featuring David Truty, VP of Corporate Reporting and Analytics at CreditAssociates, explores how Agentic AI drives customer acquisition and retention. The presentation details how CreditAssociates leverages advanced analytics and agentic AI while maintaining key operational constraints around trust, explainability, compliance, and consistency.

## TL;DR David Truty outlines CreditAssociates' strategy for solving sales queue inefficiency and optimizing customer acquisition and retention through Agentic AI. Built upon a solid, centralized data foundation, their Next-Best-Action agent ranks leads and delivers explainable recommendations to human sales reps. This advisory approach ensures regulatory compliance while significantly improving rep performance and conversion rates. Ultimately, targeting key high-impact metrics like a 1% boost in retention delivers multi-million dollar ROI and lays the groundwork for enterprise-wide AI expansion.

## Atmosphere & style The presentation carries an engaging, self-deprecating, and pragmatic tone, with the speaker balancing lighthearted humor—acknowledging he stands between the audience and lunch—with deep industry focus. It feels grounded and practical rather than overly hyped, framing complex AI integration around real-world human adoption, strict regulatory limits, and measurable business ROI. The narrative follows a logical journey from foundational data cleanup to production implementation and future enterprise scaling.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=4s) CreditAssociates operates in the consumer debt resolution industry, helping clients navigate debt through personalized savings plans and settlement negotiations. - [01:19](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=79s) The company views itself as being in the "trust business," making customer trust a primary design constraint for implementing AI. - [03:21](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=201s) Acquisition and retention rely on finding prospects with sufficient debt, hardship, and high probabilities of both enrolling and graduating from the program. - [03:56](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=236s) CreditAssociates tracks three interconnected KPIs—Marketing Efficiency Ratio (MER), Cost Per Opportunity (CPO), and Cost Per Graduate—to evaluate marketing spend. - [06:32](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=392s) A single percentage point increase in client retention delivers millions of dollars in financial value to the company. - [06:52](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=412s) CreditAssociates' AI strategy follows a three-phase roadmap: establishing a clean data foundation, deploying narrow-scope agents, and scaling across the enterprise. - [07:31](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=451s) A centralized, sanitized data foundation provides the required "memory" for AI agents to answer complex dynamic state questions about prospects in real time. - [08:41](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=521s) The Next-Best-Action agent integrates marketing attribution, behavioral data, trade line information, and credit scores to rank leads and guide sales reps. - [09:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=580s) Intentional feedback loop design is required to keep predictive models from overfitting and to ensure continuous model learning without alienating sales leadership. - [10:21](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=621s) Using AI in an advisory role with clear, explainable reason codes yields higher sales conversion rates than pulling blindly from a standard queue. - [11:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=700s) AI systems in regulated financial industries must satisfy two distinct audiences: sales reps who need trust, and regulators who require auditability. - [13:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=820s) The future vision for CreditAssociates includes expanding agentic AI into customer servicing, settlement tracking, financial governance, and outbound dialing. - [14:11](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=851s) Building a solid data foundation represents the hardest and most essential portion of an AI deployment project. - [14:23](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=863s) Explainability should be treated as a performance-enhancing feature rather than an administrative compliance tax. - [14:35](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=875s) AI initiatives should be anchored to a specific core metric where a 1% improvement translates to massive business ROI.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=4s) Introduction and the Queue Management Problem

David Truty, Vice President of Corporate Reporting and Analytics at CreditAssociates, opens his talk by introducing himself and his company. Based in Plano, Texas, CreditAssociates is a consumer debt resolution company. Truty shares that it is his first time attending the Ai4 conference and expresses how amazing it is to be surrounded by people who are "exponentially smarter than me" from whom he can learn. Acknowledging that he stands between the audience and lunch, he jokingly promises to get everyone out on time, if not a little early.

Truty frames the core operational challenge by describing the company's sales floor. Every day, sales representatives log in to a phone queue filled with hundreds of inbound and outbound calls, but they only have an eight-hour workday to navigate them. For most of CreditAssociates' history, reps simply grabbed whichever lead was sitting at the top of the queue. The underlying question driving his presentation—and the entire project—is determining who to call, when to call them, and what to say. He outlines four key stops for the presentation: an overview of CreditAssociates, what makes their problem difficult, their AI roadmap and first agentic deployment, and how they plan to scale agentic AI across the enterprise.

- **Company Location:** Plano, Texas. - **Core Sales Challenge:** Reps managing hundreds of calls in an 8-hour shift traditionally pulled sequentially from the top of the queue. - **Agenda:** Company background, industry-specific constraints, AI roadmap & agent setup, and future enterprise scaling.

---

[01:19](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=79s) Debt Resolution and the Trust Business

Clients come to CreditAssociates overwhelmed by credit card debt, and the company works to help them regain control by creating personalized debt plans with structured savings and negotiating debt settlements on their behalf. Over its history, CreditAssociates has resolved over $1 billion in debt across hundreds of thousands of customer accounts. The company's mission is to instill financial confidence using financial literacy and compassion, guided by the belief that people live better debt-free.

Truty emphasizes that CreditAssociates does not view itself as being strictly in the "debt business," but rather in the "trust business." Customers reach out at the lowest point of their financial lives, entrusting the firm to guide them back to stability. If the company fails to earn and maintain that trust, nothing else it does matters. This focus on trust acts as a critical binding constraint on every decision regarding how AI is implemented across the company.

- **Track Record:** Resolved over $1,000,000,000 in debt across hundreds of thousands of accounts. - **Core Philosophy:** Guided by compassion and financial literacy, treating customer trust as a fundamental engineering constraint.

---

[03:21](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=201s) Balancing Acquisition, Retention, and Financial KPIs

The primary goal of CreditAssociates' acquisition and retention strategy is to acquire and retain the right customers at an optimized spend level while personalizing the client journey throughout their lifecycle. Truty points out that acquisition and retention are essentially the same goal with different timing—retention does not start upon enrollment, but rather with choosing the right customer to enroll in the first place. A qualifying customer has a sufficient level of debt, faces genuine financial hardship, can fund the structured program, and shows a high probability of both enrolling and graduating.

Because every marketing dollar spent is a bet on a prospect the company has not yet met, CreditAssociates tracks three interconnected KPIs to evaluate those bets: 1. **Marketing Efficiency Ratio (MER):** Evaluates overall portfolio-level marketing allocation and efficiency. 2. **Cost Per Opportunity (CPO):** Derived from a Multi-Touch Attribution (MTA) model to measure the cost efficiency of getting a lead to talk to a debt consultant. 3. **Cost Per Graduate:** Measures the total cost required to shepherd a client through the entire length of the program.

Looking at these metrics in isolation can be misleading because optimizing one can harm another. For instance, the easiest way to boost MER is to purchase cheap leads, whereas the easiest way to maximize graduation rates is to be extremely selective about enrollments. Personalization bridges this gap by delivering the right message at the right time. Truty notes that a client suffering from job loss and maxed-out credit cards requires a completely different conversation than one struggling with unexpected medical debt. Prior to this project, sales metrics were missing targets, lack of personalization drove up serving costs and eroded margins, and retention improvements had hit a plateau after all traditional process changes, scripting, and training were exhausted. Because a single percentage point increase in retention is worth millions of dollars to the company, funding was secured for an AI-driven approach.

- **Key Metrics Tracked:** - **MER** (Marketing Efficiency Ratio) for portfolio allocation. - **CPO** (Cost Per Opportunity) via Multi-Touch Attribution modeling. - **Cost Per Graduate** for long-term program efficiency. - **Financial Impact:** A single 1% boost in client retention translates to millions of dollars in ROI.

---

[06:52](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=412s) Three-Phase AI Roadmap and Data Foundation

Reflecting on a keynote speech from Dataiku earlier at the conference, Truty agrees that successful AI companies are not necessarily those that move the fastest, spend the most, or shout the loudest, but those that establish a solid foundation. CreditAssociates structured its AI strategy into three phases: - **Phase 1: Foundation (Data Sanitation & Single Source of Truth):** CreditAssociates spent most of the previous year building a centralized, sanitized data foundation with analytics layered on top, aligning marketing, sales, and operations on a single source of truth. - **Phase 2: Agent Production & Narrow Scope:** Deploying agentic AI specifically focused on acquisition and retention to prove out ROI before expanding. - **Phase 3: Enterprise Scale:** Expanding agents to customer servicing, settlements, and financial governance.

Truty explains that the data foundation serves as the "memory" of the agentic system. To make intelligent decisions, the agent must evaluate complex "state questions" for every prospect—such as how many times they have been emailed, whether they viewed a digital ad or received direct mail, how recently that occurred, if they responded, and whether they fall within state calling windows. Without a clean data foundation, answering these dynamic state questions accurately in real time would be impossible.

- **Phase 1 Focus:** Centralized and sanitized data truth across marketing, sales, and operations. - **Role of Foundation:** Provides the memory required for the agent to evaluate complex, multi-touch prospect states.

---

[08:41](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=521s) Next-Best-Action Agent Architecture and Sales Integration

The Next-Best-Action agent integrates inputs from Media Mix Modeling (MMM), Multi-Touch Attribution (MTA) models, behavioral data, demographic data, trade line utilization, and credit scores (where compliantly allowed). The system scores and ranks leads, providing explicit reasons for the ranking, and instructs sales reps on the next best action to take to guide clients toward enrollment and eventual graduation.

``` +-----------------------------------------------------------------------+ | DATA INPUTS | | • Media Mix Model (MMM) • Multi-Touch Attribution (MTA) | | • Behavioral & Demographic Data • Credit Score & Trade Lines | +-----------------------------------------------------------------------+ │ ▼ +-----------------------------------------------------------------------+ | NEXT-BEST-ACTION AGENT | | • Evaluates state questions & compliance constraints | | • Ranks leads and generates explainable action reasons | +-----------------------------------------------------------------------+ │ ▼ +-----------------------------------------------------------------------+ | HUMAN-IN-THE-LOOP EXECUTION | | • Reps receive recommendations + reasoning | | • Reps execute personalized outreach → Enrollment → Graduation | +-----------------------------------------------------------------------+ ```

Truty details several practical insights gained while bringing this agent into production: 1. **Answering Complex Queue Dynamics:** While standard predictive models can rank lead quality, an agentic system dynamically evaluates competing queue factors (e.g., deciding whether to call a prospect who opened an email yesterday versus 400 other leads waiting in the queue). 2. **Managing Sales Feedback Loops:** As the model learned who actually enrolls, it narrowed down lead selection. However, to prevent model drift and ensure continuous learning, the team had to intentionally feed the model some leads it did not particularly like, leading to "fun, awkward conversations" with sales leadership. 3. **Advisory Role vs. Full Automation:** CreditAssociates deliberately chose an advisory setup where the agent advises and human reps make the final decision. Internal testing showed that sales reps who received explainable reason codes alongside recommendations closed at a significantly higher rate than reps pulling blindly from a traditional queue.

- **Input Signals:** MMM, MTA, behavioral indicators, credit score, and trade line utilization. - **Human-in-the-Loop Results:** Reps who understood the agent's explanation and trusted the recommendations achieved far higher closing rates.

---

[11:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=700s) Trust, Compliance, and Consistency as Engineering Principles

Operating in a heavily regulated industry means that any deployed AI system must satisfy two distinct audiences: sales reps, who must trust the advice enough to act on it, and regulators, who may demand an audit of why a decision was made. Therefore, trust and explainability are treated as fundamental engineering principles rather than optional features.

``` ┌───────────────────────────┐ │ ENGINEERING PRINCIPLES │ └─────────────┬─────────────┘ │ ┌───────────────────────┴───────────────────────┐ ▼ ▼ ┌───────────────────────┐ ┌───────────────────────┐ │ SALES REPS │ │ REGULATORS │ │ • High trust & clear │ │ • Full explainability│ │ reason codes │ │ • Regulatory & policy│ │ • Prevents going │ │ compliance │ │ "rogue" │ │ │ └───────────────────────┘ └───────────────────────┘ ```

Truty notes that if a sales rep loses trust in the model, goes rogue, and happens to be right a few times, organizational trust in the system vanishes. To maintain alignment, the agent provides clear reason codes (such as high digital engagement or high enrollment likelihood) while strictly enforcing state, local, and internal policy regulations. Crucially, the agent brings consistency to the sales floor, ensuring that similar leads are handled with the same high standard every time.

While the models and agentic layer are actively running, Truty notes that the final impact number (left as "X" on his slides) is still being earned. Because graduation occurs months or years after initial enrollment, the team must play a patient game before publishing long-term ROI metrics, promising to present the final numbers at the following year's conference.

- **Two Key Audiences:** Sales representatives (trust/adoption) and regulators (auditability/explainability). - **Core Benefit:** Replaces inconsistent rep behavior with repeatable, policy-compliant execution.

---

[13:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=820s) Enterprise Expansion and Key Takeaways

Looking forward, CreditAssociates plans to expand its agentic AI architecture from acquisition and retention into other business functions, including customer servicing, settlement tracking, financial spend governance, and AI-powered outbound dialers. The overarching goal remains delivering an elevated, personalized experience at every touchpoint along the client lifecycle.

Truty summarizes his presentation with three main takeaways, plus a partner shout-out: 1. **The Data Foundation is the Real Work:** "The boring data work really was the project for us." Establishing a clean data foundation was the necessary prerequisite for agentic AI. 2. **Explainability is a Feature, Not a Tax:** In a regulated environment, explainability drives user trust, compliance, and higher conversion performance. 3. **Target Your 1% Metric:** Identify the single metric in the business where a 1% shift translates to massive financial value (for CreditAssociates, a 1% gain in retention equals millions of dollars). This clear ROI provides the launchpad and funding for AI initiatives. 4. **Partner Acknowledgment:** Truty gives a "shameless plug" to their analytics partner, Kaizen Analytics (located at booth 500), thanking them for their multi-year collaboration and mentioning their golf tees before wishing the audience a great rest of the conference.

- **Future Scaling Areas:** Customer servicing, debt settlement, spend monitoring, and AI outbound dialers. - **Final Advice:** Build a solid data foundation, treat explainability as an asset, and anchor AI investments to high-ROI business metrics.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=4s) exponentially smarter than me — David Truty

> [00:04](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=4s) who to call, when to call them, and what to say — David Truty

> [01:19](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=79s) debt business — David Truty

> [01:19](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=79s) trust business — David Truty

> [06:52](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=412s) State questions — David Truty

> [08:41](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=521s) fun, awkward conversations — David Truty

> [11:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=700s) X — David Truty

> [13:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=820s) The boring data work really was the project for us. — David Truty

> [13:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=820s) Explainability is a Feature, Not a Tax — David Truty

> [13:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=820s) Target Your 1% Metric — David Truty

> [13:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=820s) shameless plug — David Truty

## Mentioned David Truty, CreditAssociates, Ai4, Plano, Texas, Dataiku, Kaizen Analytics

## Video description David Truty - VP, Corporate Reporting and Analytics, CreditAssociates

It has always been critical to acquire the right customers, optimize customer-related spend, and personalize engagement at every stage of the customer lifecycle. Agentic AI creates a huge opportunity for companies to increase customer acquisition and retention.

This presentation will cover how Agentic AI is helping CreditAssociates take their customer acquisition and retention to the next level, building on their existing advanced analytics capabilities and factoring in key themes of trust, explainability, compliance, and consistency.

[Source](https://www.youtube.com/watch?v=jrabW5hKrmQ)

中文简报

CreditAssociates部署Agentic AI提升销售转化与留存 CreditAssociates将AI战略与统一数据底座及可解释的潜客推荐相结合，通过仅一个百分点的客户留存率提升获取数百万美元价值。 David Truty概述了CreditAssociates通过Agentic AI解决销售队列效率低下、优化获客与留存的战略。基于稳固的集中化数据底座，其“下一步最佳行动”（Next-Best-Action）智能体对潜在客户进行优先级排序，并为人工销售代表提供可解释的建议。这种辅助决策模式在确保合规的同时，显著提升了销售代表的人效与转化率。最终，聚焦于留存率提升1%等高价值关键指标，不仅带来了数百万美元的投资回报率（ROI），也为全企业范围内的AI拓展奠定了坚实基础。 CreditAssociates 将客户信任视为核心工程约束，帮助背负沉重债务的客户制定个性化还款方案并完成债务协商，目前已累计解决超过 10 亿美元债务。 (1:19) 公司将获客与留存视为统一目标，通过综合跟踪营销效率比（MER）、单次获客机会成本（CPO）和每位结业客户成本三项指标平衡支出，客户留存率每提升 1% 即可创造数百万美元价值。 (3:21) 团队将 AI 战略划分为三阶段，首先构建集中且规范的数据底座作为智能体的“记忆”，用于实时评估潜客的复杂动态状态。 (6:52) 上线“下一步最佳行动”（Next-Best-Action）智能体，整合归因模型、行为特征与信用评分对线索进行排序并提供解释，采用人机协同建议模式显著提升销售转化率。 (8:41) 未来 CreditAssociates 计划将智能体扩展至客户服务、债务结算跟踪及外呼拨号系统，并强调可解释性与锚定高 ROI 业务指标是落地 AI 的关键。 (13:40)

Reading summary

David Truty of CreditAssociates details how the debt resolution firm implemented an advisory, next-best-action agentic AI system to resolve sales queue inefficiencies, arguing that sustainable acquisition and retention gains require a clean data foundation, human-in-the-loop explainability, and adherence to strict regulatory constraints. Aligning Acquisition KPIs with Long-Term Retention (3:21) CreditAssociates has resolved over $1 billion in debt, but its sales representatives historically navigated daily phone queues of hundreds of leads by simply dialing from the top down. Truty argues that customer acquisition and retention represent the same core objective separated by time: retention begins by enrolling individuals with verifiable financial hardship, sufficient debt, and the capacity to complete structured savings programs. Optimizing acquisition solely through standard metrics can create conflicting operational incentives. While the Marketing Efficiency Ratio (MER) measures high-level spend efficiency, Multi-Touch Attribution Cost Per Opportunity (CPO) evaluates sales lead costs, and Cost Per Graduate tracks the ultimate expense of carrying a client through the multi-year program. Maximizing one metric in isolation can harm another, as purchasing low-cost leads improves MER while depressing graduation rates. Prior to deploying AI, the firm found that conventional interventions—such as updated sales scripting and process training—had hit a performance ceiling. Because a single percentage point increase in client retention represents millions of dollars in financial value, the company pursued an algorithmic approach to deliver personalized outreach tailored to specific hardship profiles, such as medical debt versus sudden job loss. Building Data Foundations for Dynamic State Evaluation (6:52) CreditAssociates organized its enterprise AI roadmap into three distinct stages: establishing a sanitized data foundation, deploying targeted agents in acquisition and retention, and scaling architectures across servicing and settlement operations. Truty notes that data hygiene represented the primary operational challenge, requiring nearly a full year to align sales, marketing, and operational metrics into a single source of truth. This unified architecture functions as the operational memory for agentic workflows. To make dynamic lead recommendations, an agent must assess real-time prospect states, including cross-channel touchpoint histories across email, direct mail, and digital advertising, user response recency, and legal calling windows across state jurisdictions. Truty emphasizes that without establishing this underlying data foundation first, agents cannot accurately evaluate the state conditions needed to route calls effectively. Human-in-the-Loop Architecture and Recommendation Explainability (8:41) The company's Next-Best-Action agent synthesizes signals from Media Mix Modeling, Multi-Touch Attribution, demographic indicators, behavioral trends, credit scores, and trade line utilization. Rather than automating outreach, CreditAssociates implemented an advisory framework that presents prioritized leads and explicit reason codes to human debt consultants, who retain final operational control. Internal tests revealed that sales representatives provided with transparent reasoning closed prospects at significantly higher rates than those working from traditional, unranked queues. To prevent algorithmic drift and sustain model learning, the analytics team deliberately routed a subset of low-scoring leads to sales representatives, maintaining broad training exposure despite initial pushback from sales leadership. Embedding Trust and Governance for Enterprise Scaling (11:40) In a regulated consumer finance environment, system design must satisfy both internal sales staff and external regulators. Truty asserts that algorithmic explainability operates as a fundamental engineering requirement rather than a compliance burden. Clear rationale prevents sales representatives from bypassing system recommendations, while providing auditable trails to verify adherence to federal, state, and internal outreach rules. CreditAssociates plans to expand its agentic framework into customer servicing, debt settlement negotiation workflows, financial spend tracking, and automated dialers. The long-term objective is standardizing representative performance across the organization while ensuring consistent, policy-compliant customer interactions across every stage of the account lifecycle. Truty acknowledges that long-term program ROI remains unfinalized, as measuring client graduation requires months or years of follow-up observation. CreditAssociates maintains that foundational data infrastructure and explainable advisory systems—rather than rapid automation—are the essential prerequisites for capturing multi-million-dollar retention returns.

精炼阅读摘要

CreditAssociates 企业报告与分析副总裁 David Truty 指出，在受严格监管的债务重组行业中，将智能体（Agentic AI）作为“人机协同”的建议层，能够精准优化线索队列并显著提升转化率，但其成功的核心前提是可解释性以及坚实的数据底座。销售队列痛点与留存的商业杠杆 (0:04) CreditAssociates 位于得克萨斯州普莱诺，专注于消费者债务重组，历史上已累计为数十万客户解决超过 10 亿美元的债务。在传统业务流程中，销售代表在 8 小时的排班中面对包含数百通进出电的电话队列，往往只能盲目从队列顶端顺序拨打。然而，债务重组的本质建立在信任之上，客户在经历失业、信用卡刷爆或突发医疗支出等财务困境时接入系统，需要高度个性化的沟通，而不是千篇一律的话术。针对获客与留存，该公司密切追踪三项相互制约的指标：衡量整体营销组合效益的营销效率比率（MER）、基于多触点归因（MTA）的单次机会成本（CPO），以及衡量客户全程履约总成本的单结业客户成本（Cost Per Graduate）。Truty 指出，若孤立优化任一指标往往会损害其他环节，例如低成本购买劣质线索虽能推高 MER 却难以完成债务计划。由于常规培训与脚本优化已达到瓶颈，而客户留存率每提升 1 个百分点即可为公司带来数百万美元的价值，这一高投资回报率成为启动 AI 项目的直接商业依据。三阶段规划与数据底座的记忆功能 (6:52) CreditAssociates 将 AI 战略划分为三个阶段：第一阶段聚焦于数据清洗与建立营销、销售及运营的“单一事实来源”；第二阶段针对获客和留存定向部署智能体以验证投资回报；第三阶段将扩展至客户服务、债务结算和财务支出治理。Truty 强调，许多企业急于求成，但构建集中且规范的数据底座才是整个项目的核心实质。在技术架构中，数据底座实质上充当了智能体的长期“记忆”。智能体在决策时必须动态评估多触点状态，例如客户接收营销邮件的次数、最近是否浏览过数字广告或收到直邮、是否有过互动响应，以及外呼操作是否符合各州设定的呼叫时间窗口。若缺乏清洗完备的数据底座，系统将无法在毫秒级时间内准确判断这些复杂的状态变量。下一步最佳行动智能体的工程落地与协同 (8:41) 该公司落地的“下一步最佳行动”（Next-Best-Action）智能体整合了媒体组合模型（MMM）、多触点归因（MTA）、用户行为数据、人口统计特征以及合规允许范围内的信用评分与信用记录利用率。系统不仅对线索进行评分与动态优先级排序，还生成具体的推荐理由，指导销售代表开展个性化沟通以促成客户签约与长期履约。在实施层面，团队采取了“人机协同”（Human-in-the-loop）的顾问模式而非完全替代人工。Truty 透露，向销售代表提供具备可解释性的推荐理由，使得线索的最终转化率大幅高于传统盲拨模式。此外，为了防止模型发生漂移并维持持续学习能力，团队有时必须故意将部分模型不看好的线索输入系统，这在实践中曾引发与销售管理层的业务摩擦与沟通；但通过动态平衡即时互动线索与海量积压线索，系统展现了常规预测模型难以企及的队列决策灵活性。可解释性与企业级规模化扩展 (11:40) 在受到严格监管的金融服务领域，智能体必须同时满足两大受众的诉求：一是销售代表必须充分信任推荐建议以避免擅自违规操作（go rogue），二是监管机构要求每一项推荐决策都具备合规可追溯性。因此，CreditAssociates 将“信任”与“可解释性”视为系统底层的工程设计原则，而非事后负担，从而确保一线团队在面对相似线索时能维持高度一致且合规的执行标准。在未来规划上，CreditAssociates 计划将这一智能体架构逐步拓展至客户服务体系、结算跟踪、财务支出管控以及 AI 外呼拨号工具中，以覆盖客户的全生命周期。Truty 总结了三项核心经验：打磨数据底座是 AI 实施中最关键的脏活累活；可解释性是驱动转化与合规的战略特性；锚定能带来数百万美元回报的核心关键指标（如 1% 的留存率改善）是 AI 举措获得立项与资金支持的关键。尽管该智能体已在生产环境中正常运行并展现出更高的转化率，但 Truty 指出，债务结业周期通常需要数月乃至数年，因此最终量化的长期投资回报率指标仍在积累与验证之中，完整数据计划于下一年度公布。

Detailed notes — zh

## Background 该视频由 Ai4 于 2026 年 8 月 15 日发布在 YouTube 平台上。视频记录了 CreditAssociates 企业报告与分析副总裁 David Truty 的主题演讲，介绍了公司如何在现有高级分析能力的基础上引入 Agentic AI，以优化获客与留存，并兼顾信任、可解释性、合规性与执行一致性。 ## TL;DR 面对销售队列机械抓取与多重 KPI（营销效率比、商机成本与结案成本）博弈的业务挑战，CreditAssociates 推出了聚焦获客与留存的 Next Best Action 智能体落地项目。团队历时一年夯实数据底座构建长期记忆，坚持“智能体建议，人类决策”的人机协同模式，并通过赋予模型极强的可解释性来赢得一线坐席与监管机构的信任。未来公司计划将智能体进一步扩展至客户服务与智能外呼等全生命周期触点，以此撬动留存率哪怕微小提升所带来的巨大商业回报。 ## Atmosphere & style 演讲风格务实清醒、风趣干练，没有空洞的 AI 概念炒作，而是充满了一线实操的真实洞察与真诚经验。从体贴观众午餐时间的幽默开场，到坦言“枯燥的数据治理才是工程本质”与暂不公布未完全收敛数据的克制，演讲者展现出对金融强监管行业的高度敬畏。整场分享节奏明快，既有深度的业务架构解析，又有充满人情味的落地心法，临场感极强且极富说服力。 ## Key takeaways - [01:19](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=79s) 债务减免业务本质是“信任行业”，由于客户身处绝望的财务低谷期，赢得并维持信任构成了技术选型与 AI 系统设计的底层硬性约束。 - [02:24](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=144s) 获客与留存是同一目标的连续阶段，客户最终能否履约留存并非始于签约之后，而是在选择签约何种资质客户时便已注定。 - [02:24](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=144s) 企业需协同平衡营销效率比 (MER)、每商机成本与每毕业客户成本三大 KPI，孤立追求廉价线索会直接损害长期的客户成功与结案毕业率。 - [02:24](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=144s) 债务化解的终极指标“毕业（Graduation）”具有长达数月乃至数年的业务滞后性，给模型验证与短期归因带来了独特的评估挑战。 - [05:42](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=342s) 在常规运营优化见顶后，留存率每变动 1% 即可带来数百万美元的商业价值，寻找并锚定高杠杆关键指标是争取 Agentic AI 预算与立项的战略起点。 - [05:42](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=342s) 枯燥的数据治理才是工程本质，建立跨部门的“单一事实来源”构成了智能体追踪用户多触点交互历史的“长期记忆系统”。 - [08:37](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=517s) Next Best Action 智能体融合了归因模型、实时数字行为与信贷合规数据，能够精准决断复杂的全局状态并指导坐席在最佳时机开展外呼。 - [08:37](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=517s) 模型部署中必须平衡“探索与利用（Exploration vs. Exploitation）”，主动向销售分配部分非高分样本以防算法过拟合并维持持续学习进化。 - [08:37](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=517s) 坚持“智能体建议，人类决策”的人机协同原则，A/B 测试证明为一线坐席提供清晰可解释的推荐理由能大幅提升信任度与实际转化率。 - [11:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=700s) 在强监管行业中可解释性是一项核心功能（Feature）而非合规税负（Tax），它不仅满足监管审计，更能防止一线代表怀疑系统并擅自脱轨操作。 - [11:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=700s) 智能体架构有效补齐了过往缺失的“执行一致性（Consistency）”，确保每一个线索在全生命周期各触点均按照统一的高标准规范进行跟进。 - [13:01](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=781s) 在获客留存场景验证后，企业计划将智能体架构横向拓展至客户服务履约（Servicing Agent）、支出风控以及 AI 智能外呼拨号器等全域板块。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=4s) 销售排队的困境与项目初衷 CreditAssociates 的企业报告与分析副总裁 David Truty 在 Ai4 大会上开启了他的分享。作为一家总部位于德克萨斯州普莱诺（Plano, Texas）的消费者债务减免公司，其业务一线每天都面临着极具代表性的挑战：销售代表每天登录系统后，面对的是排着数百通待呼叫电话的队列，而全天只有 8 个小时的工作时间。在公司过去的大部分时间里，团队的做法非常简单粗暴——直接从队列最顶端抓取下一条线索进行跟进。究竟该联系谁、何时联系、在电话中该沟通什么内容，构成了本次分享与整个 Agentic AI 落地项目的核心母题。 ```mermaid flowchart LR accTitle: Sales Queue Prioritization Shift subgraph Traditional["传统模式"] A["数百条待呼叫线索"] --> B["从队列顶部机械抓取"] B --> C["缺乏针对性沟通"] end subgraph Agentic["Agentic AI 模式"] D["多维数据信号输入"] --> E["Next Best Action 智能体"] E --> F["高转化率/高毕业率精准触达"] end ``` - 演讲者承诺高效演讲，不耽误听众的午餐时间，并表达了向现场顶尖 AI 同行学习的谦逊态度。 - 本次分享主要围绕四个核心站点展开：CreditAssociates 的业务背景、业务痛点所在、AI 演进路线图、首个落地的智能体架构，以及如何在整个企业层面进行规模化推广。 --- ### [01:19](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=79s) 业务本质：建立在信任之上的债务化解在 CreditAssociates 的业务模型中，客户通常深陷信用卡债务泥潭、倍感无助，公司通过提供定制化储蓄方案并代表客户进行债务协商，帮助其摆脱财务困境。在公司的发展历程中，他们已累计化解了超过 10 亿美元的债务，服务了数十万个账户。然而，David Truty 强调团队从不单纯认为自己处于“债务处理行业”，而是身处“信任行业（Trust Business）”。客户在人生中最绝望的财务低谷期向他们寻求帮助，如果失去了客户的信任，其他一切工作都毫无意义。这种对信任的执念，也成为了后续 AI 决策与落地过程中不可动摇的刚性约束。 --- ### [02:24](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=144s) 行业特殊约束与三大 KPI 的张力博弈债务化解行业面临三大独特的业务约束：第一，一个错误的决策并不仅仅意味着错失了一笔潜在订单，而是意味着一个身处困境的人没有获得救助；第二，公司需要在错综复杂的美国联邦、州及地方监管法规下运作，所有操作必须具备完全的可解释性与可审计性；第三，整个项目的终极指标是“毕业（Graduation）”——即客户完成全部债务清偿、走向财务自由。然而，“毕业”是一个在初始签约入组（Enrollment）后数月乃至数年才会显现的滞后信号。在营销与运营层面，公司将获客与留存视为同一目标的两个不同阶段，并依靠三项核心 KPI 来衡量每笔营销支出的有效性： ```mermaid flowchart TD accTitle: Balancing Three Core KPIs subgraph KPIs["核心衡量指标"] K1["营销效率比 (MER) 投资组合层面的资金分配"] K2["每商机成本 (Cost per Opportunity) 基于多触点归因 MTA 评估"] K3["每毕业客户成本 (Cost per Graduate) 全程服务与结案的总成本"] end K1 <-->|追逐廉价线索 vs 严审准入资质| K3 ``` - **获客与留存的连续性**：留存并不是从签约后才开始的，而是从“决定签约谁”的那一刻便已注定。理想客户需具备符合条件的债务规模、真实的财务困境、项目履约资金能力，以及极高的入组并最终“毕业”的可能性。 - **孤立看指标的危险性**：若单独看某一个指标往往会产生误导。为了优化 MER，团队很容易倾向于购买低成本的劣质线索；而为了确保高毕业率，又必须极度挑剔客户资质。过度偏向任何一方都会破坏另一方，必须协同平衡。 - **个性化破局**：因失业刷爆信用卡与因突发医疗债务陷入困境是截然不同的情境，唯有在正确的时间向特定客户传递针对性的信息，才能在控制服务成本的同时拉动留存。 --- ### [05:42](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=342s) 业务挑战与三阶段 AI 路线图在推行 AI 之前，由于缺乏精细化运营，仅按队列顺序作业导致销售指标难以达标；“千人一面”的服务方式拉高了服务成本并压缩了利润空间。更关键的是，公司此前已通过流程改造、话术优化和人员培训拿下了所有“低垂的果实（Low-hanging fruit）”，留存率哪怕只提升一个百分点，其带来的商业价值都高达数百万美元。常规手段见顶后，引入 Agentic AI 成为突破瓶颈的必然选择。 ```mermaid flowchart LR accTitle: Three Phase AI Roadmap P1["第一阶段：数据基石 单一事实来源、状态追踪与记忆构建"] --> P2["第二阶段：智能体落地 聚焦获客与留存的 Next Best Action"] P2 --> P3["第三阶段：全域扩展 推广至客户服务与债务清算等板块"] ``` - **数据基础决定成败**：正如会议演讲所强调的，AI 成功的关键不在于谁最快、谁花钱最多，而在于底座是否扎实。团队花费了去年整整一年的时间打磨第一阶段，统一销售、营销与运营部门的口径，建立集中且清洗干净的“单一事实来源（Source of Truth）”。 - **解答“状态问题（State Questions）”**：智能体要解决的不仅是客户好坏，而是复杂的全局状态判定——例如该客户被邮件营销触达了几次、是否看过数字广告、是否收到过直邮、最近一次互动在何时、是否有转化响应等。底层数据基石实际上充当了智能体的长期“记忆系统”。 - **阶段范围控制**：第二阶段严格聚焦于获客与留存验证，待模式成熟并积累足够业务势能后，再向客服与清算板块拓展。 --- ### [08:37](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=517s) Next Best Action 智能体落地与实战心法在现阶段的架构中，系统汇集了来自多触点归因（MTA）、媒体混合模型（MMM）、用户行为、人口统计学，以及在合规前提下获取的信贷额度使用率（Trade Line Utilization）和信用分等数据信号。模型对线索进行评分排序并给出可解释的推荐理由，由 **Next Best Action 智能体** 接管并指引下一步具体动作，助力实现从触达到毕业的全周期目标。在复盘智能体落地历程时，David 提到了三点关键洞察： 1. **底座价值释放与复杂状态决断**：传统专家模型或许能给出单一预测，但智能体能处理复杂的组合场景。比如面对一个昨天刚打开邮件并点击了行动号召（CTA）、仍处于该州合法外呼时间窗口内、同时排在 400 人呼叫队列中的特定线索，智能体可以精准评估并指导销售应该优先打给谁。 2. **销售协同与模型探索（Exploration vs. Exploitation）**：起初向销售团队宣讲时充满理想主义色彩，承诺大家都能轻松达标。但随着模型逐渐收敛，只挑高转化群体，团队不得不经历与销售管理层的“尴尬对话”——为了防止模型过拟合并持续学习，必须主动向销售分配一部分模型评分并不高的数据进行探索，这需要销售团队的理解与配合。 3. **坚持“智能体建议，人类决策”**：尽管分析团队追求快速高效的自动化决策，但 CreditAssociates 明确选择让智能体作为辅助顾问，最终决策权交给一线坐席。这绝非出于胆怯或保守，而是 A/B 测试的真实结果证明：当销售人员拿到附带清晰“推荐理由”且具备高度解释性的线索列表时，他们建立了对系统的信任，实际成交转化率远高于传统的机械呼叫。 ### [11:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=700s) 信任、可解释性与合规：将信任确立为核心工程原则 CreditAssociates 深刻认识到自身所处的是一个“以信任为核心”的行业，因此团队所部署的任何技术与模型系统，都必须严格通过两类受众的审视与检验：其一是第一线的销售代表，他们必须发自内心地信服系统给出的建议并据此采取行动；其二是外部监管机构，他们随时可能要求公司对其决策逻辑进行合规性解释。正因如此，信任绝非可有可无的附加项，而是从一开始就作为系统设计的硬性工程原则。在智能体（Agent）系统的落地过程中，团队必须同时对信任度与模型收益进行优化。因为一旦业务代表对系统产生怀疑，开始脱离推荐逻辑擅自操作（“go rogue”），即便他们碰巧做对了几次，系统好不容易建立的信任机制也会瞬间瓦解。为了解决这一问题，系统强化了可解释性（Explainability），明确向销售代表展示决策背后的具体缘由——例如客户展现出“高签约意向”或是触发了特定的线上数字行为信号。与此同时，系统确保所有交互严格符合州法律、地方法律以及公司内部政策的合规要求。最后，智能体架构还补齐了过往长期缺失的关键一环：执行的一致性（Consistency），确保每一个线索在每次触达时都能按照相同的高标准与规范路径进行跟进。 ```mermaid flowchart TD accTitle: Dual Audience Trust Gate and Agentic Principles subgraph Engine["智能体决策与策略引擎"] A["潜在线索与数字信号输入"] --> B["Agentic AI 决策与推荐"] end subgraph TrustGates["双重受众审视卡点 (Trust Gates)"] B --> C["一线业务代表 (Sales Rep)"] B --> D["合规监管机构 (Regulator)"] end subgraph CoreValues["核心工程落地价值"] C --> E["可解释性 (提供明确推荐原因)"] D --> F["政策合规 (州/地方法规与内控)"] E & F --> G["全流程一致性执行 (Consistency)"] end ``` - **双重受众审视**：销售代表需要确信并执行推荐；监管机构需要清晰透明的规则解释。 - **防止脱轨行为**：销售代表一旦对推荐失去信任，就会自行其是；因此算法优化中“信任度”与业务指标提升同等重要。 - **可解释性信号**：向业务前端清晰展示客户的高签约倾向评分及在线行为特征。 - **标准化一致性**：消除人为跟进差异，实现线索处理流程的高效统一。 --- ### [13:01](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=781s) 落地进展与全生命周期扩展蓝图目前，底层核心预测模型已经构建完毕并投入稳定运行，上层的 Agentic 智能体层也正在紧锣密鼓地搭建与部署中。谈及具体带来的业绩提升数字时，演讲者幽默地指出幻灯片上的影响数值之所以暂时标注为“X”，是因为团队不想拿虚假的数据应付观众，真实的成果仍处于验证和赚取阶段。由于公司此前曾推动过留存率的变动，非常清楚留存率每提升一个百分点对整体业务意味着多少真金白银的价值，因此接下来的重点就是精确量化这一增量。不过，由于债务重组客户从签约到最终完成履约（“graduation”）的业务信号存在较长的滞后周期，团队目前必须保持耐心，预期在来年的大会上才能完整公布最终的成效数据。在业务扩展规划上，团队目前刻意将试点范围保持在较窄的切口，但在验证成功之后，计划迅速将智能体架构横向拓展至整个客户生命周期： 1. **客户服务与履约（Servicing Agent）**：部署专门的客服智能体协助处理日常服务请求。 2. **支出与成本控制**：利用智能体监控并防止营销与运营支出超标失控。 3. **AI 外呼拨号器（AI Outbound Dialers）**：作为模型系统的一部分，目前已在着手构建智能外呼能力。整套方案的核心愿景是在坚实的底层数据之上，不断扩大智能体覆盖的业务表面积，为客户旅程的每一个触点提供高度个性化且体验卓越的服务。 --- ### [14:28](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=868s) 三大核心心得与合作伙伴致谢在分享的尾声，演讲者提炼了三条给企业实施 Agentic AI 项目的核心洞察与实践心得： 1. **枯燥的数据治理才是工程本质**：“boring data work really was the project”——清洗、整合与打通数据基座占据了绝大部分精力，但它是支撑后续一切智能决策的真正基石。 2. **在强监管行业中，可解释性是一项功能（Feature），而非额外负担（Tax）**：透明可追溯的逻辑不仅是满足合规的必备机制，更能反哺一线销售的采纳度。 3. **在业务中寻找那价值连城的“1%”**：找到能够撬动巨大商业价值的关键单点指标，正是明确了这 1% 留存率的商业价值，团队才成功申请到项目预算并将其打造为业务腾飞的发射台。演讲者特别向长期合作伙伴 Kaizen Analytics 表达了感谢，坦言若没有他们过去数年间的紧密协作，CreditAssociates 无法顺利完成这一复杂的 Agentic AI 落地。最后，他幽默地为合作伙伴做了现场展位推广（展位号 500），并邀请大家前去交流并领取高尔夫球座纪念品，在掌声中结束了本次演讲。 ## Notable quotes > [01:19](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=79s) 我们从不单纯认为自己处于债务处理行业，而是身处信任行业（Trust Business）。 — David Truty > [02:24](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=144s) 留存并不是从签约后才开始的，而是从决定签约谁的那一刻便已注定。 — David Truty > [05:42](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=342s) AI 成功的关键不在于谁最快、谁花钱最多，而在于底座是否扎实。 — David Truty > [05:42](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=342s) 底层数据基石实际上充当了智能体的长期记忆系统。 — David Truty > [08:37](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=517s) 为了防止模型过拟合并持续学习，必须主动向销售分配一部分模型评分并不高的数据进行探索。 — David Truty > [08:37](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=517s) 当销售人员拿到附带清晰推荐理由且具备高度解释性的线索列表时，他们建立了对系统的信任，实际成交转化率远高于传统的机械呼叫。 — David Truty > [11:40](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=700s) 一旦业务代表对系统产生怀疑，开始脱离推荐逻辑擅自操作（go rogue），即便他们碰巧做对了几次，系统好不容易建立的信任机制也会瞬间瓦解。 — David Truty > [14:28](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=868s) 枯燥的数据治理才是工程本质（boring data work really was the project）。 — David Truty > [14:28](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=868s) 在强监管行业中，可解释性是一项功能（Feature），而非额外负担（Tax）。 — David Truty > [14:28](https://www.youtube.com/watch?v=jrabW5hKrmQ&t=868s) 在业务中寻找那价值连城的 1%：找到能够撬动巨大商业价值的关键单点指标。 — David Truty ## Mentioned David Truty, CreditAssociates, Ai4, Kaizen Analytics, Next Best Action, MTA, MMM ## Video description David Truty - VP, Corporate Reporting and Analytics, CreditAssociates It has always been critical to acquire the right customers, optimize customer-related spend, and personalize engagement at every stage of the customer lifecycle. Agentic AI creates a huge opportunity for companies to increase customer acquisition and retention. This presentation will cover how Agentic AI is helping CreditAssociates take their customer acquisition and retention to the next level, building on their existing advanced analytics capabilities and factoring in key themes of trust, explainability, compliance, and consistency. [Source](https://www.youtube.com/watch?v=jrabW5hKrmQ)

Source quotes

[1:19] David Truty: 我们从不单纯认为自己处于债务处理行业，而是身处信任行业（Trust Business）。 [2:24] David Truty: 留存并不是从签约后才开始的，而是从决定签约谁的那一刻便已注定。 [5:42] David Truty: AI 成功的关键不在于谁最快、谁花钱最多，而在于底座是否扎实。 [5:42] David Truty: 底层数据基石实际上充当了智能体的长期记忆系统。 [8:37] David Truty: 为了防止模型过拟合并持续学习，必须主动向销售分配一部分模型评分并不高的数据进行探索。 [8:37] David Truty: 当销售人员拿到附带清晰推荐理由且具备高度解释性的线索列表时，他们建立了对系统的信任，实际成交转化率远高于传统的机械呼叫。 [11:40] David Truty: 一旦业务代表对系统产生怀疑，开始脱离推荐逻辑擅自操作（go rogue），即便他们碰巧做对了几次，系统好不容易建立的信任机制也会瞬间瓦解。 [14:28] David Truty: 枯燥的数据治理才是工程本质（boring data work really was the project）。 [14:28] David Truty: 在强监管行业中，可解释性是一项功能（Feature），而非额外负担（Tax）。 [14:28] David Truty: 在业务中寻找那价值连城的 1%：找到能够撬动巨大商业价值的关键单点指标。

Source quotes

[0:04] David Truty: exponentially smarter than me [0:04] David Truty: who to call, when to call them, and what to say [1:19] David Truty: debt business [1:19] David Truty: trust business [6:52] David Truty: State questions [8:41] David Truty: fun, awkward conversations [11:40] David Truty: X [13:40] David Truty: The boring data work really was the project for us. [13:40] David Truty: Explainability is a Feature, Not a Tax [13:40] David Truty: Target Your 1% Metric [13:40] David Truty: shameless plug

Original transcript

[0:04] So, my name is David Trudy. I run corporate reporting and analytics for Credit Associates. [0:09] Um, we're a consumer debt resolution company based out of Plano, Texas. [0:14] Um, it's my first time at AI4. Uh, it's amazing to be surrounded by so many people that are [0:19] exponentially smarter than me that I get to learn from. Um, so great conference so far. [0:25] I'm super happy to be here and uh I know I'm the last thing probably between you and lunch. So, [0:29] you have my promise to get you out of here on time, if not a little early. Um, I want to start [0:36] with our sales floor. Um, every day we have reps log in to a queue filled with hundreds of calls [0:43] um and only eight hours in the day to work through those. For most of our company's history, [0:49] we just always grab from the top of the queue. That's the next one that we take. So the question [0:52] of of who to call, when to call, what to say, that was really this this talk and and that was was [0:59] really this project. Four stops today. Uh we'll briefly just talk about who Credit Associates is, [1:07] what made our problem hard. Um we'll touch on our AI road map, take a look at the first agent that [1:13] we've been developing, and then talk about how we'd like to scale throughout the enterprise. [1:19] So, at Credit Associates, people come to us overwhelmed uh with credit card debt, [1:24] and we help them get out from under it. And we do that by giving them a personalized plan with [1:28] structured savings, negotiate their debt for them. Uh we've resolved over a billion dollars worth of [1:33] debt in in our company's existence, hundreds of thousands of accounts. Uh so many people helped. [1:38] Our goal is to instill financial confidence in our clients using literacy and compassion. Um [1:46] it it guides everything that we do. We truly believe that people live better debt-free. Um [1:53] I don't we don't often talk about being in the debt business though. We we'd like to think that [1:58] and say that we're in the trust business because people come to us at the lowest point of their [2:03] financial lives and they're entrusting us um to help get them out of that. And so if [2:07] we don't have their trust, nothing else that we do truly matters. Um and that trust is is really [2:15] we'll we'll see it pop up in later slides. um it's a binding constraint on a lot of [2:18] the decisions that we've made um on on how we decide to implement AI at our company. [2:24] So um three things kind of specific to our industry that that have made this hard. Um [2:31] one is a bad decision isn't somebody that we don't enroll. It's not a missed client. Um it's somebody [2:35] that doesn't get help and and we hate that. Um secondly, we work across a patchwork of of state [2:44] and and local regulation. And so everything that we do has to be explainable and auditable. [2:51] Uh and then lastly, the the goal of our program is graduation. It's you enroll, [2:57] we settle everything, and you get on a path to financial freedom. So graduation is the key, [3:02] but it's also a signal that comes months, if not years, after we make that initial enrollment. So, [3:08] um, you'll hear me say graduation is is the key and and it really is. It it drives, [3:14] um, everything that we do. So, what's our goal? Our goal is to acquire and retain [3:21] the right customers at an optimized spend and to be able to personalize the journey [3:25] uh, all throughout the client life cycle. So, acquisition and retention may look [3:31] like two separate goals, but to me, they're really the same, just with different timing. [3:36] Retention doesn't start when a client enrolls. It it it starts with who we enroll. [3:42] Um and and the right customer for us is somebody that has a qualifying level of debt. They have a [3:46] true financial hardship. They're able to fund our program and they have a high likelihood to not [3:51] just enroll but enroll and graduate. And that's the goal of our program. So every marketing dollar [3:58] that we spend is a bet on somebody we haven't met yet. Um and so we use three KPIs to really [4:07] keep track of whether that was a good bet or not. Um the first one at at sort of a portfolio level [4:12] is our marketing efficiency ratio. It's how we allocate and uh we don't just report on this, [4:17] we optimize against it. And so it's that that's a a good portfolio level KPI. Uh next is our costs [4:25] per opportunity. This is a result of an MTA model that we've had in place and uh really just tells [4:31] us how efficient are we at getting people to make a call, talk to a debt consultant. [4:37] And then lastly, what does it cost us to work somebody through the entirety of the program? [4:41] So what is that cost per graduate? U if you look at those in isolation, they can kind of mislead [4:48] you. They can lie to you. Um which is why we look at those those three together. When we look at [4:53] the right customer at that kind of optimized spend, those two things are actually at odds, [5:00] the the best way to optimize your MER is to go after cheaper leads. And [snorts] the best way to [5:06] ensure graduation is to be pick your enrollment. And so you push too hard on one and it damages the [5:11] other. And so that's that's a problem that, you know, that we we've got to solve. Um, [5:17] and then lastly, the the retaining the client and and and getting them through [5:21] graduation. And that's where where personalization kind of pulls its weight. You know, [5:25] the right message at the right time given the right situation is is critical for us. So, [5:32] you know, cuz somebody's struggling with job loss and and maxing out their their credit [5:36] cards is a completely different situation, somebody struggling with medical debt. So, [5:42] so some of the challenges that we that we encountered along the way, um, you know, [5:47] when we're just working the next lead in the queue and not the one that converts, our sales metrics, [5:52] you know, they they're below their targets. Um, without personalization, we treat every client [5:59] the same. It increases our cost to serve. It it it drives our our cost up and our margins down. Um, [6:07] and then retention has been hard to move because we kind of the lowhanging fruit was already taken [6:13] and retention is one of our our strategic pillars of the company. Um, [6:18] it's if this wasn't a a pet project for analytics. This we've moved this number before and and the [6:24] ROI there is real. Um, it's worth millions to us that single point of retention. So that's also [6:33] part of the reason why this this project got funded because we need to move that retention [6:37] up another point and the process changes, the scripting, the training. We'd done all of that [6:43] and so it was clear that we needed another solution to try and in and get to that next level. [6:52] So I'll touch quickly on our road map. Uh yesterday during one of the keynotes, I think [6:56] it was the data IQ guys, you know, they said the the companies that are going to be successful with [7:01] AI aren't the ones that get there fastest um spend the most money or the loudest. It's the ones that [7:07] that have the foundation. And so when I look at the three phases of our AI road map, um [7:13] you know, I sat there and was like, it's easy to nod to, but it's hard to do. Um, [7:17] we spent most of last year on phase one, which was just getting our data correct. You know, [7:24] getting our centralized and sanitized source of truth with our analytics on top of it, [7:31] getting marketing and sales and ops all aligned to what is the truth? Um, [7:37] you know, because we're we're asking a lot of what I call state questions [7:40] of of our agent. It's not just is this a good prospect or a bad prospect. It's, [7:45] you know, how many times have we emailed them? Have they seen a digital ad? Have we sent [7:50] them direct mail? You know, how how recent was that? Did it work? Did they respond? Um, and so [7:56] without that foundation, you know, those types of questions get really difficult to answer. So, [8:03] um, so the foundation wasn't just to make the the Agentica possible. It's really the the memory of [8:10] it. Um, so which leads us to phase two where we're at now in and trying to get our our agent into [8:17] production. Um, we've kept our scope deliberately narrow here. Just focusing on acquisition and [8:23] retention first and proving that out and then and then, you know, getting enough traction to [8:27] be able to deploy to other areas of the business or customer service side or uh settlement side. So [8:37] So here's where we're at right now. [8:41] We've got all of our signals in. So, our media mix model, our multi-touch attribution model [8:46] feeds this behavioral data, demographic data. Um, in some cases, we have trade [8:51] line utilization, credit score, that type of data. Anything that that we are allowed to have [8:56] compliantly, we will use to feed to the model and it's going to rank and give reasons uh why [9:04] we're going to rank and work those leads in that order. Um, and then our next best action agent [9:10] takes over and tells us what to do next. Um, and we get to achieve our goal, [9:17] which is the acquisition and retention of our clients and seeing them through to graduation. So, [9:24] this is obviously a very simplified view of of what we've built. When I kind of take a step back [9:30] and think about the journey to this point, um, a couple different things come to mind, [9:39] you know, one is the foundation paid off in a really big way. Um, you know, I said that a [9:46] couple slides ago, we had a lot of state questions and it's, you know, an expert model can tell you, [9:51] okay, what's the next best thing to do? But the agent can answer the harder question of, [9:57] "Okay, what do we do with this lead? We emailed them yesterday. We know that they opened the [10:01] email. They followed the the call to action. Um, they're still within their states calling window. [10:07] And uh we've also got 400 other people in the same queue, you know, so so who do we go after?" [10:13] Um, so that's one thing. And and then that that the foundation just made that possible. The second [10:18] thing, um, I think back to conversations with our sales team. You know, at the start of this, [10:24] it's it's, hey, we're going to tell you what to work, when to work it, how to work it. Um, [10:29] everyone's going to hit their goals. We're all going to sing kumbaya together, and it's going [10:32] to be great. Um, but then as the model gets in place, it starts to narrow down the pool, [10:38] uh, as it gets smarter and learns who who's actually going to enroll. And so [10:43] um there are some awkward conversations of we still have to feed leads to the model that it [10:49] doesn't particularly like to make sure that it keeps learning. So that led to some fun [10:54] conversations with sales leadership. Um and then the last thing the intelligent decisioning. Um [11:02] coming from an analytics background I want to help the company make smartest the smartest decisions [11:06] possible as fast as we can. Um, so I want to get that into their hands and then the execution and [11:11] and automation next. But our decision here was to let the agent advise and let our people make [11:18] the decision. And that's not from a place of fear or timidity at all. It's really we tested [11:25] and the agents that we gave the list to that gave the reasons why it was explainable and they had [11:31] trust with it, they closed at a much higher rate than the ones just were working in the queue. [11:40] So, I mentioned earlier that we're in the trust business and everything that [11:45] we deploy has to pass through two audiences, right? It has to pass through a sales rep [11:50] that needs to believe it and act upon it and then it has to pass through a regulator that may [11:56] ask us to explain it. Um, so this wasn't just a nice to have for us. [12:02] that trust was really an engineering principle that needed to be a part of this. [12:09] And speaking of of trust, you know, it's it's [12:12] that trust and explanability with our agents is something that we have to optimize for [12:17] because the minute an agent doesn't trust it, goes rogue and is actually right a couple times, [12:21] then then we've lost that trust. And so you have to optimize for trust almost as well as asyft. So, [12:29] so explainability, you know, give the reasons why, you know, um, you know, high likelihood to enroll, [12:35] searching, you know, digital signals online. Um, give them the reasons and then making sure [12:41] that we stay within our state, local, and our internal policy. Make sure we we are [12:47] compliant there. And then lastly, and I think the the key to what's been missing for a long time [12:53] is just consistency for us. So, working the same lead the same way every time. um has been huge. So [13:01] the models are built and running. The agentic layer is going up right now. Um [13:06] the impact number which I left as as X there um is still being earned honestly. So I didn't want [13:12] to put a a fake number up there. So we know what a point of retention is worth to our company because [13:18] we moved it before. We've watched that we've watched that number. And so we're going to measure [13:23] it again. And right now we're kind of having to play the patient game because like I said earlier [13:27] the the signal of graduation um happens long after the enrollment. So that'll take months and I think [13:34] um you know next year when I'm here I'll have the slight finish so we can talk about it then [13:40] um scaling across you know our company we kept the scope deliberately narrow but you know I think as [13:47] we continue to prove out and see success from what we're doing now would really love to see this go [13:52] into the servicing get a servicing agent get an agent to help on the customer service side um [13:58] getting an agent to help us just keep track of keep our our our spend from from going out of [14:03] line. You know, we're already starting to build AI outbound dialers to be part of this model. So, [14:10] that's something we're working on right now. But again, the whole goal here is just making [14:13] sure that we have an elevated experience personalized at every step of the journey. [14:19] Um, you know, and building on that foundation, the same foundation, just more surface area. [14:28] So, if you take nothing else away from this, um the three things that that kind of come to mind [14:34] to me are are the boring data work really was the project for us getting that foundation built. [14:40] Um and then in a regulated industry and business like we're in, the explanability uh is a feature. [14:47] It's not a tax. And you know, go find that 1% number in your business. that's what got [14:54] us funded and and kind of the launch pad for for what we'd like to do. Um and then lastly, [15:01] I should have added this as a bullet, but um much of of what we've done wouldn't have been possible [15:07] without a great analytics partner that we had. Um we've been working with Kaizen Analytics the last [15:12] several years. They've been great partners. Again, much of this isn't possible without them. So, [15:18] shameless plug for them. Go see them in in uh booth 500. But I think they've got some [15:22] really cool golf tees and uh uh really appreciate the time. Thank you Kaisen [15:27] for the invitation to to come share this and uh enjoy the rest of the conference.