CASE SUMMARY
AI did not replace adviser judgment. It organised large volumes of video, PDFs and sales knowledge into one trusted entry point.
The client is a Hong Kong insurance and financial advisory team. New hires revisit training videos, advisers prepare sales conversations and managers search across many product brochures. Different staff often repeated the same research.
We built a bilingual AI knowledge platform connecting training, conversation practice, official product documents, initial comparison and access control. AI supports search, summaries and practice, while people verify sources and retain final professional judgment.
As information grows, repeated searching and verification become the real cost.
A one-hour training video may contain only a few sections relevant to the current client. Without timestamps, staff replay it or search from memory. Product files also vary by language, version and date, so advisers must first confirm that each PDF is current before comparing anything.
A cloud folder can store files, but it does not combine role permissions, contextual questions, source references and review. The team needed an end-to-end knowledge workflow, not another filing cabinet.
Training was hard to reuse
Videos, slides and notes were scattered, making a single useful point expensive to retrieve.
Product versions caused ambiguity
Chinese and English names, official pages, PDF versions and dates had to stay aligned.
Roles needed different access
Internal staff, trainees, clients and administrators should not see identical functions or data.
Long videos became searchable knowledge, while static scripts became repeatable practice scenarios.
After an administrator uploads a training video or document, a background workflow processes the media, transcribes it and creates summaries, takeaways and application scenarios. The aim is to turn roughly one hour of material into three to five minutes of actionable reading while preserving the original source for verification.
The sales module organises techniques for first meetings, follow-up and objections. Staff can submit their own wording for analysis or practise through Cantonese or English role-play. A post-session review identifies missed questions and next steps, while managers retain control of training standards.
Roughly 370 plans were organised around official documents, with AI comparisons traceable to source pages.
A structured catalogue holds current plans from nine Hong Kong insurers. The Chinese interface prioritises Chinese names and PDFs, while the English interface uses English sources. After staff select plans, the system finds relevant pages on premiums, coverage, terms, limits and suitable profiles.
Results separate what the documents state from what they do not state. When adding a new product, AI can prefill the insurer, plan name and type, but a person must approve the record before it enters the live catalogue.
Product terms, coverage and client advice remain subject to current official documents, company compliance requirements and licensed professional judgment.
A usable interface is only half the work. Permissions, background processing and error tracing make the system sustainable.
The platform includes administrator, internal team, trainee and client roles, reinforced by row-level database access. Large videos and PDFs are processed in the background, so users can leave the page and return to a clear processing state.
The system uses React, TypeScript, Vercel, Supabase, n8n and a separate media-processing service. Realtime voice, transcription, summarisation, document extraction and research use task-appropriate services, with credentials kept in controlled environments.
The client estimated an 80% efficiency improvement and about 60 hours of monthly capacity released.
The estimate covers training-content preparation, product search, initial plan comparison, conversation preparation and internal questions. A less quantifiable gain is that the team now shares one source set, permission model and review method.
The 80% and 60-hour figures are internal estimates based on old and new workload. Ongoing validation can use active time, completion rates, feature usage and task handling time.
Frequently asked questions
Does the AI provide insurance advice directly to clients?
No. It supports internal organisation, search, comparison and training. The team confirms final content against official documents, compliance rules and professional judgment.
How does the system reduce use of the wrong version?
The catalogue stores language, official page, PDF source and version data. Comparisons link back to source pages, and new documents require administrator approval before activation.
Is the 80% efficiency gain guaranteed?
No. It is the client's internal estimate based on its workload. Other teams will vary with content volume, data quality, adoption and approval processes.
EXPERT AUTOMATION CONSULTATION
Does your team have a repetitive, error-prone workflow?
Tell us how the process works today, its monthly volume and the common exceptions. We will identify the part worth automating and define a testable first phase.