FINANCIAL SERVICES

Max AI Chatbot

AI Chatbot Qualification: 25% Efficiency Gain for Benefits Access

How IncomeMax achieved 25% efficiency gains with an empathetic AI chatbot for benefits qualification. RAG system, human-in-the-loop design, Salesforce integration.

Max AI Chatbot

"Everyone at our org is joyous and buzzing. Max is already handling most of our customer support inquiries and has saved us a ton of time and money. Riteeka and the team are all geniuses and we love partnering with you."

Leigh Thompson, Director of Services, IncomeMax

THE CHALLENGE

The problem.

IncomeMax helps vulnerable populations access financial benefits they're entitled to but often miss. Their human advisors guide customers through complex qualification processes for energy grants, disability benefits, and other assistance programs. The challenge: demand far exceeded capacity. Every hour spent on initial screening meant fewer people getting the help they needed.

The constraint wasn't advisor skill. It was advisor availability. Every minute spent qualifying an ineligible customer meant one less person getting benefits advice.

Automation seemed obvious. But the population IncomeMax serves is vulnerable. Many are elderly, disabled, or financially stressed. They need empathy, not efficiency theater. Any AI solution had to maintain the warmth and trustworthiness of human advisors while handling the qualification workload.

The technical challenge was equally complex. Benefits eligibility involves interpreting dense regulatory documents like the 400-page Disability Rights Handbook. The AI needed to provide accurate, citation-backed advice. Hallucinations weren't just wrong, they could cost someone critical financial assistance.

THE SOLUTION

What we built.

RAG System for Regulatory Accuracy

We implemented Retrieval Augmented Generation grounded in authoritative sources. The 400-page Disability Rights Handbook was dismantled and structured for AI retrieval. When Max answers a question about disability benefits, it pulls relevant passages and cites sources.

This approach reduces hallucinations. The AI doesn't generate advice from its training data. It retrieves verified information from IncomeMax's knowledge base. If the answer isn't in the knowledge base, Max says so.

We also integrated web search capabilities for real-time information about energy companies and utility programs. This Perplexity-style search keeps Max current without manual knowledge base updates.

Salesforce Integration for Full Transparency

Every conversation Max has gets transcribed to Salesforce. Human advisors see the complete interaction history before engaging with a customer. This serves two purposes: quality control and seamless handoff.

Advisors can review Max's guidance, catch errors, and understand the customer's context. The AI doesn't replace human judgment. It prepares the ground for it.

Persona Design: The Polite British Advisor

Technical accuracy wasn't enough. Max needed to sound right. IncomeMax's human advisors use a warm, professional tone. They're patient, empathetic, and never condescending.

We engineered Max's persona through careful prompt design. The result: a 'polite British advisor' who mirrors the brand's existing voice. Max asks clarifying questions, acknowledges customer frustration, and explains complex topics in plain language.

We tested persona variations with IncomeMax's team. They flagged moments where Max sounded too formal or too casual. We adjusted the prompts until the tone felt natural. The AI needed to be helpful without being overly cheerful, professional without being cold.

Phased Rollout: Managing Risk with Feature Flags

We didn't launch Max to all customers at once. We used feature flags to control deployment partner by partner.

The pilot started with one IncomeMax partner. This limited blast radius. If Max made errors or customers reacted poorly, the impact stayed contained. We monitored conversation quality, customer feedback, and advisor handoff success rates.

The pilot validated the approach. Max successfully qualified customers, maintained appropriate tone, and provided accurate information. Advisors reported spending less time on initial screening and more time on qualified leads.

This phased strategy also built internal confidence. IncomeMax's team saw real results before committing to broader deployment. Feature flags gave them control. They could enable Max for new partners when ready, not when we said so.

Multi-Channel Deployment: Web and WhatsApp

Max launched on two channels: IncomeMax's web app and WhatsApp. This meets customers where they are.

Some customers prefer web chat. Others find WhatsApp more accessible. The underlying AI system is the same, but the interface adapts to user preference.

WhatsApp deployment required additional consideration. Message threading, media handling, and notification management differ from web chat. We built the architecture to handle both channels without duplicating logic.

This multi-channel approach expands reach. Customers who wouldn't use a web form will text on WhatsApp. Accessibility matters when serving vulnerable populations.

HOW IT WORKS

The details.

Answers Grounded in Real Policy, Not Training Data

We structured the 400-page Disability Rights Handbook for AI retrieval and built Max around it. When Max answers a question about benefits, it pulls relevant passages and cites its sources. If the answer is not in the knowledge base, Max says so. The AI does not generate advice from memory. It retrieves verified information and points to where it came from. We also added web search for real-time information about energy companies and local programmes so Max stays current without manual updates.

Every Conversation Goes to the Human Advisor

Every conversation Max has is transcribed to Salesforce. When a human advisor picks up, they see the full history of what Max discussed. They can review Max's guidance, catch anything that needs correction, and understand the customer's context before saying a word. The AI prepares the ground for the human, not the other way around.

A Persona That Matches the Brand

Technical accuracy was not enough. Max needed to sound like an IncomeMax advisor: warm, professional, patient, and clear. We designed the persona through careful prompt engineering and tested variations with the IncomeMax team. They flagged moments that sounded too formal or too casual until the tone felt right. The result is a polite, helpful voice that explains complex topics in plain language without being condescending.

Starting Small to Manage Risk

We did not launch Max to all customers at once. We used feature flags to start with one partner organisation and watched closely: conversation quality, customer feedback, and how well handoffs to human advisors went. The pilot validated the approach. Advisors reported spending less time on initial screening and more time on qualified leads. Only after seeing real results did we expand to new partners.

Available on Web and WhatsApp

Max launched on two channels. Some customers prefer a web chat interface. Others find WhatsApp more natural. The underlying AI is the same on both channels, but the interface adapts to where the customer is. For a service that works with vulnerable populations, meeting people where they already are matters. Customers who would not fill in a web form will send a message on WhatsApp.

OUTCOMES

What shipped.

Almost 25% operational efficiency gain in pilot rollout

400-page Disability Rights Handbook queryable in seconds

Full Salesforce transcription and advisor oversight

Deployed on web and WhatsApp channels

Phased rollout with zero service disruption

KEY TAKEAWAYS

What we learned.

  • RAG architecture is essential for regulated domains. Grounding AI responses in authoritative sources like the Disability Rights Handbook reduces hallucinations and builds trust with human advisors reviewing conversations.
  • Persona design matters as much as technical accuracy when serving vulnerable populations. We engineered Max's tone through prompt iteration until it matched IncomeMax's empathetic, professional brand voice.
  • Full transparency enables human oversight. Salesforce integration provides complete conversation transcripts, allowing advisors to review AI guidance before engaging with customers.
  • Phased rollout with feature flags manages deployment risk. Starting with one partner limited blast radius and built internal confidence before broader launch.
  • Multi-channel deployment expands accessibility. Launching on web and WhatsApp meets customers where they are, particularly important for vulnerable populations with varying tech comfort levels.
  • Web search integration keeps knowledge current. Perplexity-style search for energy company and utility program information eliminates manual knowledge base updates.
  • Efficiency gains compound when automation targets the right bottleneck. By automating qualification, advisors focus on qualified leads, helping more people without increasing headcount.

IN SUMMARY

Bottom line.

In summary, IncomeMax transformed their customer qualification process from a manual bottleneck to an automated, empathetic system. As a result, Max handles initial screening with accuracy and appropriate tone, while human advisors focus on qualified leads. The pilot delivered almost 25% efficiency gain, proving the approach scales impact without sacrificing service quality. The combination of RAG-based accuracy, persona-driven empathy, and full human oversight created an AI system that vulnerable populations can trust. IncomeMax now helps more people access the benefits they're entitled to, without proportionally increasing costs. Furthermore, that's automation that matters.

FAQ

Frequently asked.

How did you ensure the AI chatbot maintained empathy when dealing with vulnerable populations?
We implemented a human-in-the-loop design where the AI focuses on gathering information empathetically while human advisors handle all final recommendations and sensitive conversations. The chatbot was trained on real advisor conversations and call recordings to learn natural, compassionate communication patterns that resonate with vulnerable populations. Additionally, we built strict guardrails to prevent the AI from making financial recommendations or decisions. When complex emotional situations arise or the conversation requires nuanced judgment, the system seamlessly hands off to a human advisor who has full context from the chat history.
What was the process for training the AI on advisor conversations and call recordings?
The training process involved analyzing hundreds of real advisor-client conversations to identify effective communication patterns and common qualification scenarios. We extracted successful dialogue flows, empathetic phrasing, and question sequences that advisors naturally used when working with clients. This conversational data was then used to fine-tune the RAG system, ensuring the chatbot could retrieve relevant context and respond in ways that mirrored the organization's established advisory approach. The training focused on qualification questions rather than advisory responses, maintaining the human-centered approach for actual benefits recommendations.
How do you handle situations where the AI doesn't know the answer or makes mistakes?
The system is designed to acknowledge uncertainty rather than guess or provide incorrect information. When the AI encounters a question outside its knowledge base or confidence threshold, it immediately escalates to a human advisor while maintaining conversation context. We implemented confidence scoring for all AI responses, and any response below a certain threshold triggers human review. The chatbot also tracks all conversations for quality assurance, allowing advisors to review interactions and provide feedback that continuously improves the system. This human-in-the-loop approach ensures vulnerable populations never receive incorrect guidance.
Why did you choose a phased rollout approach instead of launching to all partners at once?
A phased rollout allowed us to validate the chatbot's performance with real users while minimizing risk to vulnerable populations and partner relationships. Starting with a single partner gave us concentrated feedback and the ability to quickly iterate on issues before broader deployment. This approach also helped build organizational confidence in the AI system. By demonstrating success with early partners and achieving measurable efficiency gains, we could address skepticism and refine our implementation process. Each phase provided valuable learnings about integration challenges, user behavior, and optimization opportunities that informed subsequent rollouts.
What guardrails did you implement to prevent the AI from giving incorrect financial advice?
The primary guardrail is that the AI never provides financial advice or makes recommendations—it only gathers qualification information. The chatbot is strictly limited to asking questions, collecting client data, and routing to appropriate human advisors who make all advisory decisions. Technically, we implemented content filtering, response validation, and strict prompt engineering to prevent the AI from straying into advisory territory. The RAG system retrieves factual information about programs and eligibility criteria but never interprets how those apply to individual circumstances. All outputs are logged and reviewable, with automatic escalation triggers for any responses that approach advisory language.
How did you overcome initial concerns about AI accuracy for advisory capabilities?
We addressed accuracy concerns by clearly positioning the AI as a qualification assistant rather than an advisor, which reframed expectations appropriately. The human-in-the-loop design meant that all actual advisory work remained with trained professionals, while the AI handled the more straightforward task of information gathering. We also provided extensive transparency into how the system works, including the RAG architecture and confidence scoring mechanisms. The phased rollout approach allowed skeptics to see real performance data and measurable improvements in efficiency without compromising service quality. Demonstrating that the AI reduced advisor workload by 25% while maintaining high customer satisfaction helped convert initial skepticism into support.
What KPIs do you track to measure chatbot success beyond efficiency gains?
We track multiple dimensions of chatbot performance including conversation completion rate, escalation frequency, user satisfaction scores, and accuracy of information collected. These metrics help us understand both operational efficiency and user experience quality. Additional KPIs include time-to-advisor-handoff, conversation length, repeat usage rates, and partner satisfaction scores. We also monitor quality metrics like the percentage of conversations requiring correction by advisors and the accuracy of qualification data captured. Integration health metrics track Salesforce sync success rates and data consistency to ensure seamless operation across systems.
How does the handoff from AI to human advisor work in practice?
The handoff is designed to be seamless and context-preserving. When the chatbot completes qualification or encounters a situation requiring human expertise, it summarizes the conversation and transfers all collected information directly into Salesforce where advisors can access it immediately. Advisors receive notifications with full chat history, qualification data, and any flags the AI identified as requiring attention. This means clients never need to repeat information, and advisors can start the conversation with complete context. The system also allows for warm handoffs where the chatbot can schedule calls or immediately connect clients to available advisors based on urgency and availability.

LET'S TALK

Bring us the hard problem.

We'll bring the team that ships.

Book a call Back to home
Get in touch