Case studies / eCommerce
AI agent setupNobody can memorise a catalog this big. So we stopped asking someone to.
OMG Cheers sells $37 bottles alongside five-figure collector releases across whiskey, wine, cognac, tequila, and champagne. Most first-time visitors don't know where to start: and the same question arrives a hundred different ways.
A catalog too big for any one person to hold in their head.
OMG Cheers is an online retailer selling whiskey, wine, cognac, tequila, and champagne, from $37 bottles to five-figure collector releases. A support VA handled customer questions, but only during a 40-hour workweek.
We replaced the VA with a conversational AI agent trained on the full catalog. Six months later, it handles about 600 conversations a week with 10-second replies, 24/7, for a one-time $750 setup instead of $1,100 per month.
Same job. Different arithmetic.
A support VA was handling inbound messages manually: a real person, doing the job well, within the limits of being a person. Here's what replaced them, and what it cost.
Support VA
$1,100 / month
- Coverage 9am to 5pm
- Hours per week 40
- Response time 20 to 30 min
- After hours Next day
- Catalog lookup Manual
- Six-month cost $6,600
Conversational AI agent
$750 one-time
- Coverage 24/7
- Hours per week 168
- Response time ~10 seconds
- After hours ~10 seconds
- Catalog lookup Instant, complete
- Six-month cost $750
4.2× the coverage at roughly a ninth of the six-month cost. But the cost saving isn't the interesting part: see below.
It isn't a cheaper person. It's a thing a person can't be.
The saving is real, but it's the least interesting part of this. What actually changed is that the agent does three things no human handling chat at this catalog size can do: not because they're not good at the job, but because the job has properties people don't.
Whiskey, wine, cognac, tequila, champagne. Hundreds of SKUs. Stock moving constantly.
-
Complete catalog recall, instantly
A person checking whether a specific cognac is in stock takes thirty seconds and occasionally gets it wrong. The agent knows the whole catalog at once: spirit type, flavour profile, price band, occasion, and matches against all of it in the same breath.
-
It pivots when something's out of stock
The most expensive moment in a chat is "sorry, we don't have that." Instead of leaving the customer stuck, the agent recommends a comparable bottle on the spot and sends the link. The conversation continues rather than ending.
-
3am is the same as 3pm
The VA covered forty hours a week. Browsing spirits is an evening and weekend activity. Every question asked outside business hours previously waited until the next day, by which point the customer had usually bought elsewhere or forgotten.
It wasn't good on day one.
Most AI case studies skip straight from setup to results. This one took four to six weeks of live conversations to reach full performance, and saying so is the point: anyone promising a perfect agent on launch day is describing a demo, not a deployment.
Functional
Trained on the full catalog and answering. Handling the common questions correctly, but recommendations were generic: matching on the obvious attributes rather than on what customers actually responded to.
Learning what lands
Live volume did the work. Which recommendations converted, which phrasings signalled budget, which substitutions customers accepted when the first pick was unavailable.
Full performance
The figures at the top of this page are from this period. Six months later it's still running, still handling roughly 600 conversations a week.
What we built.
Catalog-trained agent
Every SKU across six-plus categories, with attributes the agent can actually reason over.
Preference discovery
Spirit type, flavour profile, price range, occasion, asked conversationally, not as a form.
Out-of-stock substitution
Comparable alternatives offered in the moment rather than a dead end.
Direct product links
Once the customer settles, the agent sends the link so checkout is one tap away.
24/7 coverage
Evenings and weekends, when browsing spirits actually happens, handled the same as midday.
Post-launch tuning
Four to six weeks of monitoring live conversations and correcting what didn't land.
What went into it.
Three of our services, on one build.
What's your inbox doing at 3am?
If your customers browse in the evening and your support answers in the morning, there's a gap. Tell us what your inbox looks like and we'll tell you whether an agent is worth it: including if it isn't.
Setup from $750 · One-time payment · No monthly platform fee from us