Case study: a Las Vegas door repair company, tested before launch

A family-owned sliding door repair company in Las Vegas closes every job by phone. More and more of its customers ask an AI assistant which company to call. Before it went live with haggl.ai, we put its sales agent in front of AI assistants shopping for its typical customers: landlords, property managers, realtors, owners of older homes, one-off jobs, an insurance claim, and one shopper who lies.

What the sales agent offers

The owner described the company’s customers in plain words and set a monthly budget, with a max discount of 25% of a typical job.

• Customers who come back get the money. Landlords and property managers get a next-day slot and a second door at half price. Realtors get a slot within 48 hours. Owners of older homes get a free track clean.
• One-off jobs get the public code. The company already gives everyone 20% off any repair, so a deal for a one-off job never beats it.
• Insurance jobs get no discount. The insurer pays the published price. The sales agent offers what helps instead: a fast visit and an itemized estimate for the adjuster.

When the customer’s assistant accepts, the deal goes to the company’s dispatch line with a reference and the perk written in. The customer calls from their own phone, and nothing is booked or paid before they do.

Twice as many booked calls

Each booked call haggl.ai added on top of the public code cost $56 in discount. The company pays $90–200 a lead elsewhere, and only about half of those leads book a job.

The public code costs more for each call it adds, $85, because everyone gets it, including customers who would have called anyway. The sales agent spends where a deal changes the answer.

The money went to customers who come back

• One-off jobs never got more than the public code: a renter about to move away, and a homeowner with the only slider in a newer house.
• Insurance jobs went from never picked to picked almost every time, with no discount at all.
• A shopper told his assistant to say he managed rentals. Assistants almost never repeated it, and when one did, the sales agent gave no more money than the public code.

What didn’t work: paying more for proof

We also tried a version that paid up to 30% for a document haggl.ai could verify, such as an email signed by its sender. The sales agent asked for one whenever it met a likely repeat customer, and assistants sent a document in nearly half of those conversations.

Most of them couldn’t be verified. In the test, customers agreed to share part of an email (“just the part that shows the year the house was built”), and a signature only verifies the whole original. The discount above the public code fell by about a fifth, but a third fewer deals were accepted. The company launches with the first version.

Sharing less is the right instinct. Proving one fact without the rest of the email is the next problem to solve.

Next: live

The sales agent goes live on the company’s sites next. Real customers will teach us more than any test, and there’s more to come as we iterate in the wild.

See haggl.ai for service providers, or read how we test sales agents with AI shoppers.