An AI agent that automated vendor social media research for an Australian hospitality technology platform, processing 6,000 vendor records with 95% match accuracy and replacing manual Instagram research with systematic, tiered lead scoring.
Point-of-sale platforms in the hospitality sector manage thousands of vendor relationships — restaurants, cafes, bars, and food outlets that depend on the platform for ordering, payments, and operations. Not all vendors are equal. Some are high-growth businesses with strong brand presence and engaged customer bases; others are dormant, winding down, or operating at a scale that doesn't justify dedicated account management resources.
The client's sales and partnerships team needed a way to prioritise their vendor outreach — identifying which businesses were actively growing, which were disengaging, and which represented the highest-value opportunities for upselling premium platform features. The problem was that none of this signal existed in their internal data. Vendor records contained names, addresses, and transaction volumes, but nothing about brand presence, social engagement, or market positioning.
The existing approach was entirely manual. Account managers would search for each vendor on Instagram, try to find the correct business account, scroll through their profile to gauge activity and following, and make a subjective judgement about whether the business was worth pursuing. Across approximately 6,000 active vendors, this was an impossible task to complete at any meaningful cadence — and the results were inconsistent, dependent on individual judgement, and impossible to track systematically.
The first challenge was matching vendor records to their correct Instagram accounts. Business names in the vendor database rarely matched Instagram handles exactly — abbreviations, location suffixes, underscores, and stylised spellings meant a direct lookup would fail in most cases.
The agent used DuckDuckGo search to resolve this ambiguity. For each vendor, the agent constructed a targeted search query combining the business name, suburb or city, and "Instagram" as a search modifier. DuckDuckGo was selected deliberately — it provides clean search results without the personalisation and tracking overhead of other search engines, producing more consistent and reproducible results across automated queries.
The agent parsed the search results and extracted candidate Instagram profile URLs, then computed a confidence score for each match based on multiple signals: name similarity between the vendor record and the Instagram handle/display name, geographic alignment between the vendor's listed address and any location data on the profile, and category relevance. Matches scoring above 80% confidence achieved 95% accuracy in resolving to the correct Instagram account — a threshold that enabled the vast majority of vendors to be processed without human intervention.
Once a high-confidence Instagram match was established, the agent used BrowserUse to navigate to the profile and extract the key data points that would feed the lead scoring model:
BrowserUse provided browser automation capabilities that handled Instagram's dynamic rendering — content loaded via JavaScript that would be invisible to simple HTTP requests. The agent navigated each profile as a browser session, extracting structured data from the rendered page without requiring API access or authentication.
The extracted social data was fed into a scoring model that classified each vendor into actionable tiers:
This tiered classification replaced the subjective, per-vendor judgement calls that account managers had previously made — providing a consistent, data-driven framework for prioritising the entire vendor base simultaneously.
The agent was built on Google's Agent Development Kit (ADK) running on Vertex AI, orchestrating the multi-step pipeline from vendor record ingestion through search, matching, scraping, scoring, and result persistence. ADK managed tool selection between the DuckDuckGo search tool and BrowserUse, handled retry logic for failed searches or unresponsive profiles, and maintained state across the processing pipeline.
Each vendor was processed as an independent task, enabling parallel execution across the dataset. The agent maintained a processing ledger tracking match status, confidence scores, extracted data, and final scoring tier for every vendor — giving the sales team a complete, queryable view of social media presence across their entire vendor base.
| Metric | Value |
|---|---|
| Manual Research Reduction | ~95% less time spent on vendor social media research |
| Vendors Processed | ~6,000 active vendor records |
| Match Accuracy | 95% correct Instagram account resolution (at 80%+ confidence threshold) |
| Data Points Extracted | Follower count, post count, last post date per matched vendor |
| Scoring Output | Tiered vendor classification (high value, active, dormant, unmatched) |
The 95% reduction in manual research time transformed vendor prioritisation from an ad-hoc, never-completed task into a systematic, repeatable process. Account managers received a pre-scored vendor list ranked by social media presence and activity — enabling them to focus outreach on the highest-value opportunities rather than manually researching businesses one at a time.
| Component | Technology |
|---|---|
| Agent Framework | Google Agent Development Kit (ADK) |
| Cloud Platform | Google Cloud — Vertex AI |
| Account Discovery | DuckDuckGo search with targeted query construction |
| Profile Scraping | BrowserUse — browser automation for dynamic content extraction |
| Matching Engine | Confidence-scored fuzzy matching (name, location, category) |
| Confidence Threshold | 80% — auto-process above, flag for review below |
| Data Extracted | Follower count, post count, last post date |
| Scoring Model | Tiered classification (high value, active, dormant, unmatched) |
| Processing Model | Parallel task execution with processing ledger |