What an AI Assistant Should and Should Not Answer on a Clothing Store
By The Velocity Wear Team
A clothing store gets the same questions over and over: will this fit, when will it arrive, how do I wash it, can I return it. Those are worth automating. The questions underneath them — can you make this in navy, will my logo work at that size, my order is wrong — are not, and the difference matters more than the technology.
The good candidates
Automate questions where the answer exists, is stable, and is the same for everyone.
- **Sizing against published measurements.** If your size chart is accurate, an assistant reading it is genuinely helpful and available at midnight.
- **Delivery expectations.** Standard timeframes and destinations. Anything that depends on a specific order should hand off to the order record rather than be guessed.
- **Care instructions.** Fixed per product and frequently asked after purchase.
- **Returns policy.** Stable, written down, and the answer people most want at 11pm before buying.
- **Stock and availability,** where it can read live inventory rather than infer.
The bad candidates
These are the ones that cause damage, because a wrong answer is worse than no answer.
- 1**Bulk quotes.** Pricing depends on quantity, garment, decoration method, colour count and placement. An assistant that improvises a per-unit price has made a commitment you then have to break.
- 2**Artwork suitability.** Whether a file will print well is a judgement about resolution, colour and edges. Getting this wrong produces a rejected job after payment.
- 3**Lead times for a specific order.** These depend on artwork approval, stock and current capacity. A confident invented date is the fastest route to a complaint.
- 4**Complaints and faults.** Someone with a damaged order does not want to be triaged by software. Route straight to a person.
- 5**Anything about a specific existing order** unless the system genuinely reads that order.
The rule that prevents most of the harm
Ground every answer in a source, and make the assistant say when it does not have one. An assistant that answers only from your published size charts, policy pages and product data — and hands off otherwise — is useful. One that generates plausible text about your business is a liability, because it will invent a lead time, a discount or a policy that you then have to honour or deny.
Make the handoff good
Most implementations put the effort into the answering and none into the handoff, which is backwards. The handoff is where the difficult, valuable conversations start.
Carry the conversation across so the customer does not repeat themselves. Set an honest expectation of when a person will reply. Make the escape hatch visible from the first message rather than after three failed attempts — someone with a bulk enquiry is the most valuable visitor on the site, and making them argue with software is an expensive way to lose them.
What to measure
Not deflection rate. Deflection counts conversations that ended, which includes everyone who gave up and left. Measure resolution — did the customer get what they needed — and track whether enquiry volume and order value went up or down after launch. Those are the numbers that tell you whether it worked.
