Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.ideaboxai.com/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

This use case demonstrates how a financial services firm used IdeaBoxAI’s Knowledge Bases to transform their AI-powered customer support from generic, hallucination-prone responses into precise, data-grounded answers — reducing customer complaint escalations by 62% and improving first-contact resolution rates from 54% to 89%.

​
Background

Horizon Financial Services, a mid-market investment advisory firm managing portfolios for over 3,200 clients, deployed an AI chatbot to handle routine customer inquiries about account balances, transaction history, investment performance, and fee structures. Initially, they used a standard large language model trained on general financial knowledge. While the AI could answer broad questions like “What is a diversified portfolio?” or “How does dollar-cost averaging work?”, it consistently failed when customers asked specific questions about their own accounts, such as:
  • “What was my total return on my growth portfolio last quarter?”
  • “Why was I charged a $45 fee in March?”
  • “Which of my holdings have dividends scheduled this month?”
The AI would either provide generic responses (“Fee structures vary by account type — please review your agreement”), make up plausible-sounding but incorrect information (hallucination), or deflect to human support. This created frustration for customers who expected instant, accurate answers and overwhelmed the human support team with avoidable escalations.

​
The challenge

Horizon Financial Services faced three critical issues with their initial AI implementation:
  1. Hallucinated responses — Without access to real customer data, the AI would occasionally fabricate account details, transaction amounts, or fee explanations that sounded credible but were factually incorrect. This eroded customer trust and created compliance risk.
  2. Generic deflection — For most account-specific questions, the AI would respond with unhelpful generic statements like “Please contact your advisor for details” or “You can view this information in your account portal” — which defeated the purpose of having an AI assistant in the first place.
  3. High escalation rate — 46% of AI conversations escalated to human agents within the first three interactions because the AI couldn’t answer basic, data-driven questions. This created a bottleneck in the support queue and negated the cost savings the AI was supposed to provide.

​
The solution

Horizon Financial Services implemented IdeaBoxAI’s Knowledge Base system to connect their customer account database (PostgreSQL) and transaction ledger (MySQL) directly to their AI Persona, enabling the AI to query live data and ground every response in real, verified information.

​
Implementation approach

1

Created structured Knowledge Bases

The IT team created two Knowledge Bases in IdeaBoxAI:
  • Customer Accounts KB — Connected to their PostgreSQL database containing account types, balances, portfolio allocations, asset holdings, and advisor assignments.
  • Transaction History KB — Connected to their MySQL ledger with all deposits, withdrawals, trades, dividends, and fee transactions dating back 5 years.
Each Knowledge Base was configured with a semantic data model that mapped database columns (e.g., acct_balance, txn_type, fee_code) to natural language concepts the AI could understand.
2

Deployed grounded AI Persona

The team created a custom AI Persona in IdeaBoxAI called Horizon Assistant with clear instructions:
  • Always query the Knowledge Bases before responding to account-specific questions.
  • Cite the specific data source (account number, transaction ID, date) in responses.
  • If data is not available in the Knowledge Bases, explicitly state “I don’t have access to that information” rather than guessing.
  • For compliance-sensitive topics (e.g., tax advice, legal guidance), deflect to human advisors with clear disclaimers.
The Persona was attached to both Knowledge Bases, enabling it to execute SQL queries across customer accounts and transaction history in real time.
3

Configured query permissions and guardrails

To ensure data privacy and compliance, the team configured row-level security in the Knowledge Bases:
  • Customers could only query data associated with their authenticated account ID.
  • Advisors could access data for all accounts they managed.
  • The AI was restricted from executing any UPDATE, DELETE, or INSERT operations — read-only access only.
The team also implemented query timeout limits (5 seconds) and result set caps (100 rows) to prevent performance issues.
4

Tested and iterated with real queries

Before launching to customers, the support team tested the AI with 200+ real customer questions collected from historical support tickets. They refined the Knowledge Base semantic layer, added missing data mappings, and adjusted the Persona’s instructions to handle edge cases (e.g., closed accounts, pending transactions, currency conversions).

​
Key configurations

Horizon Financial Services leveraged several IdeaBoxAI features to optimize accuracy and compliance:
  • Semantic data modeling — Instead of forcing the AI to write raw SQL, the Knowledge Base abstracted database schema into business terms. For example, the AI could understand “What are my dividend-paying stocks?” and automatically map it to SELECT symbol FROM holdings WHERE dividend_yield > 0 AND account_id = [user].
  • Citation and transparency — Every AI response included the data source and timestamp. For example: “Your growth portfolio returned +8.3% last quarter (Q1 2026, data from transaction ledger as of May 7, 2026).” This transparency increased customer confidence.
  • Fallback handling — When the AI encountered ambiguous queries or incomplete data, it asked clarifying questions (“I see three portfolios under your account — which one are you asking about: Growth, Income, or Balanced?”) rather than guessing.
  • Human handoff triggers — The Persona was configured to automatically escalate to a human agent if it detected keywords related to complaints, legal issues, or emotional distress — ensuring sensitive situations were handled appropriately.

​
Results

Within 120 days of deploying grounded Knowledge Base responses, Horizon Financial Services achieved significant improvements in support quality and efficiency:
MetricBeforeAfterImprovement
First-contact resolution rate54%89%+35 pts
Customer complaint escalations112/month43/month-62%
Average resolution time18 minutes3.5 minutes-81%
Hallucinated/incorrect responses11% of queries<1% of queries-91%
Customer satisfaction (CSAT)3.2/54.6/5+44%
Support ticket volume2,400/month980/month-59%

​
Qualitative outcomes

Beyond the quantitative metrics, the team and customers reported several qualitative benefits:
  • Customer trust and confidence — Customers appreciated receiving precise, data-backed answers with clear citations. Comments like “This is exactly what I needed — no runaround” and “Better than talking to a human advisor” became common in feedback surveys.
  • Reduced compliance risk — By eliminating hallucinated financial advice, Horizon significantly reduced their exposure to regulatory penalties and customer disputes stemming from incorrect AI-generated information.
  • Support team morale — Human agents spent less time answering repetitive, data-lookup questions and more time on high-value advisory work — leading to higher job satisfaction and lower turnover.
  • Scalability without headcount — The firm onboarded 400+ new clients during the pilot period without increasing support headcount. The grounded AI absorbed the additional query volume seamlessly.

​
Key takeaways

This use case highlights three core capabilities of grounded AI responses powered by Knowledge Bases:
  1. Factual accuracy over plausibility — By querying live databases instead of relying on pre-trained model knowledge, the AI eliminated hallucinations and delivered verifiable, source-cited answers every time.
  2. Data privacy and access control — Row-level security and read-only query permissions ensured that customers could only access their own data, and the AI couldn’t accidentally modify or expose sensitive information.
  3. Semantic abstraction for non-technical queries — The Knowledge Base’s semantic layer allowed customers to ask questions in plain English (“What did I earn last month?”) without needing to understand database schemas, SQL syntax, or technical jargon.

​
Technical considerations

For teams evaluating grounded AI implementations, Horizon’s experience surfaced several important technical lessons:
  • Semantic modeling is critical — Invest time upfront mapping database columns to natural language business concepts. A well-modeled Knowledge Base reduced incorrect queries by 73% compared to raw SQL access.
  • Query performance matters — Slow queries (>5 seconds) caused customers to abandon conversations. Horizon optimized their database indexes and implemented query caching to maintain sub-2-second response times for 95% of queries.
  • Iterate with real data — Testing with synthetic or example questions is insufficient. Real customer queries exposed edge cases, ambiguous phrasing, and missing data mappings that synthetic tests missed.
  • Explicit “I don’t know” is better than guessing — Configuring the AI to admit when data is unavailable increased trust more than attempting to provide partial or inferred answers.

Knowledge Bases Overview

Learn how to connect a live database and build a semantic data layer for AI queries.

Structured Data

Understand how structured Knowledge Bases work with SQL databases like PostgreSQL and MySQL.

Setting Up Your Copilot Persona

Create custom AI assistants and attach Knowledge Bases for grounded responses.

Connect Knowledge Base to Copilot

Learn how to connect your Knowledge Base to a Copilot persona.