A voice agent that checks 5,000 live products

01

Why a human VA fails at this task

• The showroom is open Monday to Friday, 10am to 5pm. 9 in 10 calls arrive outside that window, including $40,000 inquiries at 11pm and on Sundays. All of them hit a voicemail box nobody dealt with before Monday.
• They tried a virtual assistant first. Someone who cannot see the stock cannot answer the only question that matters: do you have this piece, and what does it cost.
• Stock changes every day, so a saved copy of the product list is out of date within 24 hours. Every answer has to come from live data.

02

How it works

Layer 1  Keeping the product list up to date

n8n’s built-in Shopify node only pulls 50 products at a time. Looping it enough times to get through all 5,000 kept hitting Shopify’s rate limit and switching the workflow off.

I replaced it with a custom request that pages through Shopify 250 products at a time, which is 5x faster and runs without me touching it.

It updates the existing record rather than adding a new one, so a price change edits the product instead of duplicating it.

Fig 1 Pulling every product out of Shopify, page by page

Fig 2 The same loop for orders, slowed down so Shopify does not block it

Layer 2  Finding the right watch while the caller waits

The agent can call 3 lookups in the middle of a conversation.

Each one does the same thing: take what the caller said, clean it up, search the product list, rank the matches, and hand back the best one in a form the agent can read out loud.

• 1Product nameLong and brand-heavy. This is what callers mispronounce and what speech-to-text mangles.
• 2PriceThe tiebreaker. When a name hits a cluster of near-identical pieces, price separates them.
• 3Model numberDigits, letters, dots, and a Preowned- prefix only some products carry.

Fig 3 Finding a watch by name and price

Fig 4 Finding a watch by model number

Fig 5 Looking up an order, with a live check if it is missing

It asks instead of guessing. If a model number matches more than one watch, the agent asks for the price and narrows it down. A pronunciation list keeps the brand names spelled correctly on the way in.

Layer 3  Turning each call into a lead the owner sees

Every finished call saves the transcript, the recording, the caller’s details and a link to the exact product they asked about.

A scheduled workflow gathers the overnight calls and emails them before the shop opens, because the owner was never going to log into a dashboard.

Fig 6 The email the owner gets before opening

03

The results, counted by hand

I listened to every call in a recent sample and tagged each one: when it came in, what they wanted, how it ended, and whether a lead was actually captured.

Stack

Retell AI n8n Airtable Shopify

More work