The founder of a small but growing online store called us during their busiest month, and within the first few minutes it was clear this was not really a technology conversation. They had not had a quiet morning in weeks. Every day started the same way: a queue of customer messages waiting before they had finished their coffee, and most of those messages were the same handful of questions asked over and over. They were not looking to sound more high-tech. They wanted their mornings back without letting service slip.
Before we recommended anything, we read a week of their messages
We did not propose a solution on that first call. Instead we asked for one week of their real customer conversations across the website chat and WhatsApp. We exported just over 600 of them, an illustrative volume for a store this size, and tagged each one by question type. The pattern was not obvious straight away. We were a couple of hundred conversations in before it really stood out.
One thread stopped us. The same customer had asked “where is my order” three times in a single evening, never got a reply, and by the next morning had opened a payment dispute. That one conversation summed up the whole problem better than any dashboard could.
The audit also changed our assumption about where the real bottleneck was. We had expected product questions, the “is this pan induction-compatible” type, to dominate. They did not. The heaviest load by far was tracking requests, and specifically tracking requests that landed after business hours, when nobody was there to answer them. Roughly seven in ten messages were the same five questions, order-status sat at the top of that list, and the work was scattered across separate inboxes for the website, WhatsApp, and social, so the founder was switching between them all day and losing the occasional message in the shuffle.
Where we decided to start, and why
We told the founder plainly that hiring another person was the wrong fix. It would add a fixed monthly cost and still leave the evenings uncovered, and the evenings were when a lot of the buying was happening.
Because those after-hours order-status requests were eating the most time and were the least human part of the queue, we chose to automate that workflow first rather than trying to solve everything at once. We set up an assistant on the two busiest channels using Zipprr AI Chat on the website and WhatsApp Automation for WhatsApp, and we built it around one firm principle we agreed on the kickoff call: it would only answer from the store’s own product and order information, never guess, and hand anything sensitive straight to a person.
We also made a deliberate call to leave Instagram out of the first rollout. The website and WhatsApp accounted for almost all of the support traffic, so adding a third channel early would only have made testing harder without changing the outcome for most customers. It could wait until the core was proven. In the same spirit, we set the assistant to quietly capture a contact on every conversation, because the audit had shown how many evening browsers arrived, asked one question, and left, and we did not want those to keep vanishing.
Building it, and the parts that needed fixing
We turned that week of reviewed conversations into the assistant’s knowledge base, writing the answers the way customers actually phrase things rather than in policy language. We connected the order-status lookup and tested it against a batch of the store’s real order references. During that testing the courier’s tracking API returned inconsistent data on a handful of orders, the sort of thing that only surfaces once you run real references through it, so we built the fallback to handle those cases cleanly (say so honestly, offer a person) and held the WhatsApp launch back by two days until we were satisfied it did.
It was not friction-free, and it should not have been. In testing, we noticed the assistant was setting slightly optimistic delivery expectations, which traced back to hopeful wording in the shipping content, so we rewrote that section with honest ranges before anything went live. On deployment day we went live on the website only, for the handful of question types the assistant could handle confidently, and we kept the founder on a Teams call while the first live conversations came in. The first few replies were exactly what we expected. Then a returns enquiry arrived and the assistant tried to handle it instead of stepping back, which reminded us in real time that our escalation rules were still not tight enough. We fixed it that same afternoon, routing anything involving refunds or complaints straight to a person. We could have switched both channels on together, but staging them, even though it made the rollout a couple of days slower, meant we caught this on the website before a single WhatsApp customer ever saw it.
One thing we had not predicted was internal. The part-time assistant actually preferred the old way of working at first and was quietly ignoring the new handoff. When we asked why, it was simple: the flagged conversations were landing somewhere they did not normally look. We moved them into the inbox the team already checked all day, and the resistance disappeared.
What changed for the team
The founder emailed us three days after launch, not to report a problem but to say the evening pile-up had simply stopped. By the end of the first week, the thing they mentioned most was that their mornings were no longer swallowed by a backlog of the same questions. Replies were going out in seconds, at any hour, and the after-hours messages that used to quietly turn into lost sales were being handled while everyone slept. Order-status requests, the old time sink, were resolved without anyone touching them.
The change we cared about most was quieter. The part-time assistant stopped spending the day copying tracking numbers and moved on to the genuinely tricky cases, the ones that actually needed a person. On our second review call, the founder laughed and said it was the first Monday in months they had opened the laptop without fifty unread support messages waiting. That was the point we knew the new workflow had settled into their daily routine.
Where things stand now
Frequently Asked Questions
How long does a project like this take?
Did the business have to change its website or replace any tools?
What happens when the assistant does not know an answer?
How do you stop an AI assistant from giving customers wrong information?
Does the assistant work on both the website and WhatsApp?
Yes. In this project we used Zipprr AI Chat for the website and WhatsApp Automation for WhatsApp, running as one assistant so customers get the same answers on either channel. We launched the website first and added WhatsApp once it was stable.
