GWS Connect 24: Adding AI to a Live B2B Marketplace

The problem

GWS Connect 24 connects verified retailers with closeout and surplus stock from manufacturers and brands. These are large, time-sensitive lots that need to reach the right buyers quickly. Before this work:

• Visitors with questions had to wait for a person to reply.
• Deals lived on the marketplace, but sales outreach happened in a separate CRM. Admins retyped deals by hand and couldn't see what outreach had already happened.
• Emails were written and tracked elsewhere, so opens, clicks and replies weren't visible next to the deal.
• Buyers struggled to find products: search needed exact spelling, country data was inconsistent, and there was no volume pricing.

The challenge was adding AI to a system with real buyers, suppliers and orders, without putting any of that at risk.

What I delivered

A 24/7 AI sales agent

An AI sales agent, powered by SalesCloser.ai, now appears on every page. On the Book a Demo page, visitors can choose a call with the team or an instant meeting with the agent. It answers questions about the platform, registration, fees and the mutual NDA, and hands over to a person when the conversation turns to prices or terms.

• It loads without slowing the page, and it can be switched off with one setting without leaving any broken screens.
• I built and tested, ahead of go-live, the connection that records every agent conversation event and sends qualified leads into HubSpot, so the sales team picks up where the AI stopped.
• Clear boundaries: the agent never names a supplier or buyer, never quotes a price, and hands anything commercial to the sales desk. A test plan checks each rule with deliberately tricky questions before launch.

A two-way connection to the CRM

• Deal to CRM: one click (with a confirmation) sends any deal from the marketplace to the AgentlyLeads CRM. Sending it again updates it rather than creating a duplicate, and any errors are shown clearly.
• CRM back to the deal: the deal page shows the campaigns, emails and calls from the CRM, without keeping a second copy of that data. If the CRM is unreachable, the page shows a simple message instead of breaking.

The loop is closed: an admin sends a deal, the sales team works it in the CRM, and the history shows up on the deal page.

AI that drafts and people who approve

I built AI workflows on top of the marketplace's own data to automate the slowest part of selling: deciding who should see a deal and what to say to them. No email is ever sent automatically.

• Market a deal: ranks the best-fit buyers for a deal from their interests and history, and writes a personalised email draft for each one, ready for review.
• Follow up: finds buyers who have gone quiet past a set time and drafts relevant follow-ups for review.
• Deal intake: reads supplier emails, PDFs and spreadsheets and turns them into structured deal listings. I used it to add more than 40 live deals and catalogues, and completed the product categories so every deal lands in the right place.

Email outreach with tracking

Emails are reviewed, edited and sent one by one, with tamper-proof click tracking, open tracking and automatic reply matching. Publishing a deal can create a campaign for every buyer interested in that category, and a campaign problem can never undo the deal being published.

A better experience for buyers and admins

• A redesigned buying experience with clear volume pricing, quick product previews, a "take the whole lot" option and minimum order rules, with stock reserved per colour and variant.
• Search that forgives spelling mistakes, consistent country filters, and counts that always match what the buyer sees.
• Admin tools: outreach review with autosave, an in-browser viewer for the original supplier emails, and internal notes.
• A mobile and accessibility review across public, buyer and supplier pages, with an automated checker for every user type.
• I found and fixed a security weakness that could have let harmful code be saved through a text editor.

Key decisions

• People stay in control. In wholesale, the wrong email to the wrong buyer costs trust. AI ranks and drafts; a person sends.
• Connect, don't copy. The CRM stays the single source of truth for outreach.
• Safe to retry. Every sync and campaign can be repeated without creating duplicates.
• Fail gently. If a connected service is down, the page explains it instead of breaking.
• Careful changes to a live system. Separate development branches, testing against a local copy first, a written plan before each feature, and database changes checked against the live setup. That check caught a data type mismatch, which I fixed quickly.

The results

The AI sales agent is live on every page, around the clock. Deals reach the CRM in one click and their outreach history is visible on the deal page. Finding buyers and writing personalised outreach went from manual research to an AI-drafted, human-approved workflow. More than 40 live deals were added through AI-assisted intake, and buyers get forgiving search, accurate filters, volume pricing and a mobile-friendly, accessible experience.