Table of contents

Kevin Silberblatt’s team builds and maintains Merge’s integrations. They’re how customers connect to hundreds of tools through a single integration with Merge, so the team is always adding more.
Every new connector and feature needs a demo, and a demo on obviously fake data falls flat. Testing on it also lets bugs slip through to customers. “When you’re also QAing or testing it, you’re not really pressure testing,” Silberblatt says. “So are you waiting for the pressure test when the customer comes?”
His team’s answer is to build demos on data that looks like a real customer’s, using Merge for Workforce to decide what that data should be.
The team starts in Merge for Workforce and its Universal Context Layer, which has already indexed the places customer signal lives at Merge: Notion pages, Gong calls, support tickets, and product feedback requests.
The first question is which integrations belong in the demo.
Silberblatt asks the context layer what most customers will want to see in a demo. It searches across the product specs, docs, tickets, Gong calls, and feedback requests at once and comes back with a recommendation. This time it suggested Salesforce, plus ticketing and file storage, along with the fields that matter in each.
He uses the context layer search over a fresh crawl because it has already done the reading. "It'll be faster, cheaper, and more reliable," he says.
The next prompt asks for a plan doc: which sandboxes to set up, and what data each one needs to carry to show the feature properly. The team was demoing an inference feature, so the data needed enough depth, including a realistic org chart, for the inference to have something to work with.
The plan is saved to the team's demo SharePoint and doubles as a product requirements doc an engineer can pick up.
The plan gets split across the team. Each person runs a separate AI stream to seed the data, with Claude handling the sandbox setup.
Because the plan already says what each sandbox needs, the streams don't have to coordinate beyond following it.
The last step confirms the data landed the way the plan said it would. Silberblatt asks for a set of sample MCP calls to test it.
He runs those test cases in a new window. With the org chart in place, the team can show the inference working for different roles in the company, and test both the happy path and the mistakes to see how the product responds.
Testing against data shaped like a real customer’s turns up problems fake data never would. "This is great product feedback," Silberblatt says, "both working for the demo as well as making our product better."
It also moves the hard testing earlier, before a customer ever sees the feature.
Claude suggests the integrations and drafts the plan and test cases, and Silberblatt’s team reads all of it before anything gets built. They also run the MCP test calls themselves and check each response, because an answer can look reasonable and still be wrong for the role it was asked about.
This post is part of How Merge Uses Merge, a series on how our own teams use the products we build.