We built Truck Gear Nation's sales and support assistant. It answers "will this fit my truck?" by querying the store's live fitment database, shows matching parts as product cards, and hands customers to the team in Slack, where replies reach the chat in real time. It drops into the WordPress store with one script tag.
First commit to production
13 daysfrom the first commit to the first production conversation00Project overview
Truck Gear Nation sells tonneau covers, bed dividers, grilles, floor mats and more to US truck owners, from about 1,224 products on WordPress and WooCommerce. Codech audited the live store, designed and built an AI assistant that answers from the store's own data, launched it, fixed what customers reported, and handed it over to the client's next team.
The challenge
A tonneau cover fits one year range, one make and model, and often one bed length. The fitment data already sat in the store's database, but customers could only reach it through a picker form or a person. Support was manual, and when a customer was passed to a person they had to explain everything again.
Our proposal
We proposed function calling instead of a static knowledge base: the AI queries the live fitment table for every answer. Before writing code we checked the proposal against the live store and corrected seven assumptions, then the client picked one of four widget designs.
The solution
A brand-safe chat widget on every store page, run by Claude with 10 tools that query WordPress and WooCommerce, and a two-way handoff to the team's Slack. The store controls it from eight WordPress admin pages.
01The problem
One wrong bed length means a part that doesn't fit and a return. Answers had to come from the store's own fitment data, not from a knowledge base that goes stale.
Passing a customer to a person meant repeating the truck, the part and the question. The team already worked in Slack and wanted to answer from there.
It had to drop into the existing WordPress site, match the brand without the theme breaking it, and come with bot protection, cost limits and US privacy (CCPA) deletion.
02Fitment answers Live
Claude doesn't remember fitment. It calls tools that query the store's fitment table and WooCommerce through a WordPress plugin, so every answer matches the catalogue the moment it is asked.
Guided vehicle search
The assistant asks for the truck using only the options that exist in the catalogue, then shows matching parts as product cards with photo, SKU, price and stock. The truck is written back to the store's own vehicle picker, so the site updates without a reload.
Fitment check
On a product page the bot knows which product the customer is looking at. Asked about a truck, it runs an exact check against the store's fitment table and answers with the SKU and the bed length it matched.
Product specs
Feature bullets, attributes, warranty and install notes all reach the AI, so "can I mount a bike rack, and does it need drilling?" gets a direct answer from the catalogue instead of "that isn't listed".
03Slack handoff Live
When a customer asks for a person, or the bot can't help, the conversation moves to the team's Slack with everything attached. Staff reply in a thread and the customer reads it in the chat, live.
Two-way Slack
The handoff posts a thread to the team's private channel with the customer, the truck, the page, an AI summary and the recent transcript. Replies in the thread reach the customer's chat in real time, and typing \end hands the customer back to the AI.
Contact details first
Anonymous customers leave a name and email (phone optional) before the handoff, so the team can always follow up. Logged-in customers skip the form.
Out of hours
Business hours are set in WordPress. Outside them, the bot offers a "Leave us a message" form, which files an inquiry in WordPress, emails the team and alerts a Slack channel.
04Safe in public Live
The team runs the bot from WordPress. Bot checks, rate limits and cost caps stop abuse, and customer data is kept for 30 days, then purged.
WordPress admin
Settings, business hours, FAQ, contact email, excluded pages, inquiries, chat history and privacy tools sit in the WordPress admin the team already uses. FAQ entries go straight into the AI's instructions.
Abuse and cost controls
Anonymous visitors pass a Cloudflare Turnstile check and get a signed session token. Every message then meets a rate limit (30 per 5 minutes), a cost cap of $0.50 per conversation, a daily cap with Slack alerts, and at most 5 tool calls per turn.
Privacy (CCPA)
A CCPA deletion request is one form in WordPress: enter the email, and the data goes from both WordPress and the chat service. Conversations purge automatically after 30 days, and IP addresses are stored only as SHA-256 hashes.
The key design decision
The AI looks it up. It never remembers fitment.
A chatbot trained on a copy of the catalogue would be wrong the day the fitment data changed. So Claude gets tools, not knowledge: every fitment or product answer is a live query to the store's own database, through a signed request to WordPress.
Live data is only half of it. The four worst bugs after launch had one shape: the AI improvised a lookup key, the exact match failed, and "not found" reached the customer as a confident fact. So the tools now take the catalogue's exact keys, and they tell "not found" apart from "I didn't understand that input".
05How it works
Each chat turn streams over server-sent events. Claude's instructions are built from the store's FAQ and the customer's truck, page and details, and every tool call is a signed request to the WordPress plugin.
06Integrations
Nothing for the team to learn: the bot lives in WordPress, answers from WooCommerce, and hands over to Slack.
Store02
Team01
AI01
Platform03
07How we delivered
A written proposal, a live-site audit that corrected the plan before any code, a choice of four widget designs, then the build. Here are Truck Gear Nation's real ones.
Highlights from the proposal: the function-calling approach, the tools, the architecture and sample conversations.
Tell us about one workflow you'd like AI to handle. We'll send you a proposal like this one, with the AI design, the tools and the plan, plus an interactive prototype you can click through. Free.
There was no clickable prototype for this project. After the live-site audit, the client compared four widget designs on one page and chose B, Clean Light.
Real files from the engagement. The proposal's fee and cost pages are not shown. The UI directions used a placeholder navy and orange palette; the live widget uses Truck Gear Nation's own black and gold.
On 29 Apr we checked the proposal against the live store and corrected seven assumptions, from WordPress sitting on a separate server to vehicle cookies that already existed.
Eight build-phase pull requests merged on 29–30 Apr. On 2 May we dogfooded it on the live site and fixed about 14 bugs; the first production conversation came on 12 May.
Post-launch fixes from client reports between 28 Jul and 14 Sep, then a written handover to the client's new team on 10 Sep.
08Under the hood
A customer was told a cover didn't fit their 5.8 ft bed. It did: the AI sent "5.8", the exact match found nothing, and nothing became "doesn't fit". Spelling variants now map to the catalogue's value, but feet-and-inches are never converted (5'6" is 5.5 ft, and the catalogue has both), and input it can't read makes the bot ask.
The product tool returned only the short description. 1,057 of about 1,224 products had feature bullets the AI had never seen, plus warranty, install notes and attributes. Now they all reach it, and spec questions it couldn't answer before get answered from the catalogue.
Handoffs nobody picked up stayed open, so we added a 20-minute timeout. Before switching it on we found nine stuck conversations, up to 121 days old, closed them and handed them to the client for personal follow-up. One atomic update means no customer is emailed twice.
Reach out to us now. We'll look at the questions your customers ask, where the answers live in your store, and what an assistant that queries them live would change, in a free 45-minute call. Then we send you a free AI proposal and an interactive prototype.
Chat screens are recreated demo conversations; products, prices and stock are real public catalogue data from 12 Oct 2026. Slack and WordPress admin screens are recreations with demo data. The proposal's fee and cost pages are not shown.