AI-Powered Insurance Claims Chatbot with Continuous Feedback Loop

allianz · Insurance · 2023 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Insurance claims customer support

Allianz Benelux needed to help customers find the correct insurance claim form or support destination across a complex product portfolio with many product variants and customer search terms.

Constraints:

• customers search for insurance help using many different terms
• insurance products map to different forms, support numbers, and destinations
• the team needed a fast implementation without a long custom software project
• the solution had to support Benelux localization across Dutch and German-speaking customers
• customer feedback needed to reach the operations team quickly enough to improve the bot
• the case study is vendor-authored and the described implementation appears rule-based rather than LLM-powered

Architecture

Allianz Benelux used Landbot's no-code chatbot platform to build a claims-navigation chatbot. The team analyzed more than 92,000 historical website search terms, categorized insurance keywords, linked them to insurance products, and mapped products to web destinations such as forms or support paths. The chatbot was monitored with Landbot analytics and drop-off analysis. Customer feedback flowed to Slack in real time and created Trello tickets so the team could convert issues into improvements. The bot was localized for Dutch and German-speaking Benelux customers through a 48-hour collaborative build.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  website-search-term-corpus["website-search-term-corpus"] --> keyword-categorization["keyword-categorization"]
  keyword-categorization --> insurance-product-map["insurance-product-map"]
  insurance-product-map --> landbot-flow-builder["landbot-flow-builder"]
  landbot-flow-builder --> customer-chatbot["customer-chatbot"]
  customer-chatbot --> destination-router["destination-router"]
  customer-chatbot --> landbot-analytics["landbot-analytics"]
  customer-chatbot --> feedback-capture["feedback-capture"]
  feedback-capture --> slack-channel["slack-channel"]
  feedback-capture --> trello-board["trello-board"]
  slack-channel --> improvement-team["improvement-team"]
  trello-board --> improvement-team
  improvement-team --> landbot-flow-builder
  landbot-flow-builder --> localization-workstream["localization-workstream"]

Components

Data flow

Historical website search terms were analyzed into insurance keyword categories. Those categories were linked to insurance products and web destinations, then implemented as a no-code Landbot conversation flow. Customers interacted with the chatbot to navigate to the correct insurance destination. Landbot analytics and drop-off analysis monitored where conversations failed. Feedback was sent to Slack and turned into Trello improvement tickets, giving the BTU and Customer Care teams a continuous improvement loop.

Why this architecture

The case supports a practical no-code, keyword-mapped design rather than an LLM agent. Allianz already had a large corpus of customer search terms, so categorizing keywords and mapping them to products gave the chatbot domain coverage without requiring generative reasoning. Landbot's no-code builder fit the three-week timeline and enabled regional localization. Slack, Trello, analytics, and drop-off analysis kept the deployed bot connected to an operational feedback loop.

Models used: Landbot no-code chatbot · keyword categorization · decision-tree chatbot

Scale

• The team used more than 92,000 historical website search terms as chatbot input.
  “the Allianz team gathered more than 92,000 unique terms people use in our search engine on the website.”article, stated
• The initial bot was developed, tested, and implemented within three weeks.
  “develop the bot, testing and implementing it within just three weeks.”article, stated
• Benelux localization was completed through a 48-hour hackathon.
  “They scaled the bot across the Benelux union serving Dutch and German speaking customers in an intensive 48-hour hackathon”article, stated
• The summary reports more than 18,000 customers served by the chatbot.
  “18,000+ customers used the chatbot”summary, stated
• The summary reports a 90% positive feedback rating.
  “90% positive feedback rating (exceeding their >85% target)”summary, stated
• The summary reports that 100 feedback points were converted into improvements within 24 hours.
  “100 feedback points converted into improvements within 24 hours”summary, stated

Alternatives considered (roads not taken)

Build a custom software project

Rejected because: The teams used Landbot's no-code builder and shipped within three weeks, which suited a rapid internal transformation effort.

Use a generative LLM or autonomous agent

Rejected because: The described implementation uses categorized keywords, product links, and destination routing rather than source-supported generative reasoning.

Review customer feedback in periodic offline batches

Rejected because: Allianz routed feedback into Slack and Trello in real time so issues became visible and actionable quickly.

Ship one generic language version

Rejected because: Benelux requirements pushed the team to build localized Dutch and German-speaking versions.

Patterns

data-quality-loop (learning)

Customer feedback and drop-off analytics are routed into Slack and Trello so the team can turn observed failures into chatbot improvements.

eval-driven-development (evaluation)

Landbot analytics, drop-off analysis, customer feedback, and positive-rating targets guided iterative changes to the chatbot flow.

human-in-the-loop-gate (control)

Feedback did not automatically alter the bot; it created Slack visibility and Trello work items for the Allianz team to review and implement.

Tradeoffs

Use no-code chatbot construction

Chose Landbot no-code bot builder over a bespoke claims-navigation application.

Rationale: The no-code builder let BTU and Customer Care teams implement and localize the bot quickly without a long engineering project.

Use keyword categorization and product routing

Chose categorized search terms linked to insurance products and destinations over open-ended generative conversation.

Rationale: Insurance claims routing requires sending customers to specific forms or destinations, and the available evidence supports structured keyword mapping rather than LLM generation.

Route feedback into team tools immediately

Chose Slack notifications and Trello ticket creation over manual review of chatbot analytics alone.

Rationale: Real-time feedback made customer issues visible to the operating team and turned them into tracked improvement work.

Gotchas (what bit them)

The implementation should not be mistaken for an LLM-powered agent.

Root cause: The source describes keyword categorization and destination mapping, while the broader Landbot page contains AI marketing language.

Resolution: Classify the case as a rule-based or NLU-style chatbot and mark it non-agentic in the extraction.

No-code speed still depended on substantial domain preparation.

Root cause: The chatbot needed years of customer search language to be analyzed, categorized, and mapped to product destinations.

Resolution: Treat the search-term corpus and product mapping as core system assets, not incidental setup work.

Vendor-reported success metrics lack methodological detail.

Root cause: The case study reports strong positive-rating and usage numbers but does not define the survey method, denominator, or share of total claims interactions.

Resolution: Preserve the metrics as reported, but keep promotional bias high and avoid treating them as independently audited outcomes.

Localization required concentrated cross-region work, not automatic scaling.

Root cause: The Benelux rollout needed a Belgian colleague, synchronous building, department testing, and language-specific versions.

Resolution: The team used a 48-hour hackathon and departmental testing to produce localized Dutch and German versions.

Evaluation

Allianz monitored the chatbot with Landbot analytics, drop-off analysis, direct customer feedback, positive-rating targets, and improvement tickets generated from feedback.

Metrics: drop-off points · customer feedback · positive feedback rating · number of customers served · feedback points converted into improvements · time to implement improvements

Outcomes

• The bot was developed, tested, and implemented within three weeks.
• The team used more than 92,000 customer search terms to structure the chatbot.
• The summary reports more than 18,000 customers served and a 90% positive feedback rating.
• Feedback was routed to Slack and Trello for continuous improvement.
• The bot was localized across the Benelux union for Dutch and German-speaking customers through a 48-hour hackathon.

Reported scale includes more than 92,000 unique search terms, a three-week build, a 48-hour localization hackathon, more than 18,000 customers served, a 90% positive feedback rating, and 100 feedback points converted into improvements within 24 hours.