Smart Business Analyst: Automating Competitor Analysis in Medical Device Procurement

philips · Healthcare · year n/a · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (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: Procurement competitor intelligence for medical devices

Philips needed to automate secondary research for competitor analysis where analysts manually gather precise feature, cost, and market data across many internal and external sources.

Constraints:

• Competitive medical-device data is niche and scattered across PDFs, Word documents, SharePoint, databases, websites, and subscriptions.
• A feature comparison can require filling hundreds of fields across devices and competitors.
• Generic ChatGPT may provide relative comparisons without precise numerical feature values.
• The system needs current internet information, multilingual access, and conversational follow-up memory.

Architecture

Philips' Smart Business Analyst uses a multi-agent GenAI workflow for secondary research: the application breaks business questions into subqueries, uses search agents with web scraping and vector storage to retrieve top-ranked URLs, connects internal and external data sources including Google Search API for fresh data, and uses a report-generation agent to create display-ready feature comparisons, charts, and conversational answers with multilingual support and memory.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Business analyst query"] --> B["Subquery generation agent"]
  B --> C["Search agents"]
  C --> D["External data sources"]
  C --> E["Google Search API wrapper"]
  C --> F["Vector store"]
  G["Internal data sources"] --> H["Report generation agent"]
  F --> H
  I["Conversational memory"] --> H
  H --> J["Analyst output"]

Components

Data flow

A business analyst asks a procurement or competitor question. A subquery generation agent splits it into many retrieval queries, search agents scrape and store relevant web results into a vector store, connected internal and external sources are queried, and a report-generation agent assembles numerical comparisons or narrative insights in a display-ready format.

Why this architecture

The system decomposes queries and uses multiple data sources because competitor information is niche, sparse, and time-sensitive, and the output must fill precise numeric fields rather than generic comparative prose.

Models used: LLM APIs · pre-trained model · GenAI chatbot

Scale

• A single analyzer comparison can involve 25 features across eight competitors, or about 200 fields.
  “that means 25 features into eight competitors that is around 200 fields that are need to be filled”transcript, stated
• The chatbot can fill a comparison that would be manual work within a few seconds.
  “if this has been enabled with the chatbot uh gen chatbot application this could be filled up within a few seconds”transcript, stated
• The system supports many source types: journals, annual reports, news feeds, regulatory data, trade associations, financial databases, and academic research.
  “the data sources could be it could be some of the subscriptions of journals and magazines it could be annual reports or some new speed and daily updates government or regulatory data”transcript, stated

Alternatives considered (roads not taken)

Manual secondary research

Rejected because: Manual collection across many sources is time-consuming and can leave fields empty when analysts cannot find information.

Generic ChatGPT comparison

Rejected because: The presenter says ChatGPT gives broad comparative statements rather than exact feature values.

Rely only on LLM training data

Rejected because: The application needs latest internet information after the model training cutoff.

Patterns

multi-agent-orchestration (agent-architecture)

The workflow uses distinct subquery generation, search, and report generation agents.

rag-retrieval (retrieval)

Search agents scrape and store web results into a vector store for report generation.

query-rewriting (retrieval)

The system splits a business query into many subqueries to retrieve maximum competitor information.

agent-memory (agent-context)

Conversational memory lets the application answer follow-up questions without repeating the initial question.

Tradeoffs

Use an agentic secondary-research tool.

Chose subquery/search/report generation agents over manual analyst research over many sources.

Rationale: The agentic workflow can fill many feature fields in seconds and reduce manual effort.

Fetch live web data.

Chose Google Search API wrapper over LLM-only responses from training data.

Rationale: Competitor analysis needs information that may have appeared after the model was trained.

Gotchas (what bit them)

Generic LLM answers may lack precise procurement numbers.

Root cause: General chat outputs tend toward relative statements when the business needs exact feature values.

Resolution: Train/connect Smart Business Analyst to data sources that return exact comparison values.

Sparse niche data can leave manual comparison cells empty.

Root cause: Analysts may not find information across scattered secondary sources.

Resolution: Split queries and search top-ranked URLs plus multiple source classes.

LLM training cutoff misses current market information.

Root cause: LLM APIs are trained at a point in time.

Resolution: Plug Google Search API into the LLM workflow for latest internet results.

Evaluation

The presentation demonstrates value through illustrative feature-matrix completion, exact numeric comparisons, speed relative to manual work, and feature support such as multilingual queries and conversational memory; it does not provide a formal offline accuracy benchmark.

Metrics: fields completed · time to fill comparison matrix · exact numerical value retrieval · multilingual support · conversational follow-up support

Outcomes

• A 25-by-8 competitor feature table can be filled in seconds in the illustrated workflow.
• The system can return exact numeric values rather than only relative comparisons.
• The tool supports live web search, multilingual questions, conversational memory, and pluggable data sources.

The concrete scale example is 25 features times eight competitors, or about 200 fields.