A bike web store lost night-time shoppers to one unanswered question: will it fit? We built an AI agent grounded in specs, size charts and compatibility tables, live on site chat and WhatsApp. It now self-serves 83% of pre-sale questions.
Typical results
6 sec
Median first answer, from 4 hrs
2.1x
Chat-to-order conversion
-33%
Wrong-size returns
Figures reflect typical results for this kind of automation.
The challenge
Bikes are bought with questions — frame size for a given height, whether a rack fits a specific model, which tube matches which wheel. Answers waited a median of four hours for a human, and evening shoppers rarely waited: carts were abandoned, and wrong guesses came back as returns, 9% of them size-related.
Before and after
Questions needing a human
−83%Conversion after asking a question
2.1×Wrong-size returns
−33%The automation
We built an AI agent grounded strictly in the store's own data — catalog specs, frame-geometry charts and parts-compatibility tables, re-indexed nightly. It answers fit and compatibility questions on site chat and WhatsApp in seconds, quotes the exact spec row it relied on, and links the correctly sized variant straight to the cart. When confidence drops, the thread hands to a human with full context, and unanswerable questions feed a weekly catalog-gap report.
How it works
- 01
Ground the agent
Catalog specs, frame-geometry charts and parts-compatibility tables are indexed nightly into the agent's knowledge base.
- 02
Answer with receipts
On site chat and WhatsApp, the agent answers fit and compatibility questions, quoting the exact spec row it used.
- 03
Link the right variant
Every answer includes the correctly sized variant, one tap from the cart.
- 04
Escalate honestly
When confidence drops — custom builds, price haggling — the thread hands to a human with full context.
- 05
Close the gaps
Questions the agent couldn't answer feed a weekly report that patches missing catalog data.
Starts from
- Site chat & WhatsApp questions
- Specs, size charts & compatibility tables
Delivers
- Spec-backed instant answers
- Low-confidence human handoff
- Sized add-to-cart links
The impact
The agent now self-serves 83% of pre-sale questions, with a median first answer of six seconds instead of four hours. Shoppers who chat convert at 2.1x the site average, wrong-size returns fell from 9% to 6%, and the support team finally works on builds, not size charts.
Questions about this automation
Where does the agent get its answers from?
Only from the store's own data: catalog specs, frame-geometry charts and parts-compatibility tables, re-indexed every night. Each answer quotes the spec row it used, so shoppers and staff can check it. If the data does not cover a question, the agent says so and offers a human instead of guessing.
When does a human take over the conversation?
When the agent's confidence is low or the request falls outside its scope, such as a custom build or a price negotiation, the thread passes to a staff member with the full conversation and the products viewed. Questions it could not answer feed a weekly report that points to missing catalog data.
Which channels and models does it run on?
It runs on the website chat widget and on WhatsApp Business, using an OpenAI model with the store's data as its only source. Every answer links the correctly sized variant straight to the cart. It is a multi-step build, usually rolled out in stages, starting with the most common product questions.