Presto Drycleaners is an established, referral-driven business with a large customer base built up over three decades and no paid advertising. Two problems were converging: a flood of inbound WhatsApp enquiries that a small customer-service team could not keep up with, and an ageing point-of-sale system that notified customers of ready orders by SMS, an increasingly expensive and low-engagement channel.
We did two things in parallel. We moved Presto's order notifications off SMS and onto WhatsApp by wiring their POS directly into the messaging platform, and we built a team of specialist AI agents to handle the resulting inbound conversations across twelve outlets, twenty-four hours a day. This is an account of how that system was built and hardened. It runs live across all twelve outlets, handling real customers around the clock, and the record of what broke early on and how we fixed it is the most useful part of this story.
Presto is a multi-outlet garment-care business: laundry and dry cleaning, plus specialist services such as curtains, carpets, leather, sofa covers and alterations through on-site tailors at selected outlets. Customers drop off and collect at any of twelve outlets, or arrange home pickup and delivery. The business is seasonal, with demand spiking ahead of the year-end and festive periods.
Growth has come almost entirely through word of mouth and a very large legacy contact base accumulated over the years, rather than through advertising. That base is an asset most businesses would envy, but it also means a high steady volume of inbound questions: pricing, turnaround times, outlet locations and hours, order status, collection, and the handling of delicate or specialist items.
Crucially, the existing custom POS already sent customers an automated message when their order was ready. It just did so over SMS, and it could not, on its own, hold a conversation when a customer replied. That gap, between a one-way notification and a two-way conversation, is where the project lived.
A small customer-service team was managing a heavy weekly conversation volume by hand. The same questions recurred constantly, across twelve outlets each with its own hours, payment methods and quirks, and the team could not be a confident expert on every service line at once while also keeping pace with the inbox.
The SMS notification was doing only half a job. Order-ready alerts went out by text, which is costly and low-engagement, and when a customer replied to one there was no system to catch and answer them. Meanwhile a referral business runs on responsiveness and reputation, but the team could only answer during working hours, so serious after-hours and weekend enquiries were simply being missed.
The design principle, as with our other deployments, was specialists over a generalist front desk, with deterministic routing and a human always one step away. We work as the operator of the system, reviewing real conversations daily and tuning against what customers actually send.
A Receptionist greets every incoming WhatsApp and Instagram message, identifies what the customer needs, and points it to the right specialist. Behind it sit specialists for pricing and services, bookings and specialist items, outlet guidance, and escalations, each kept to a tight scope with a human always one step away.
Every agent shares conversation state and hands off silently, so the customer never sees the internal routing and never has to repeat themselves.
The larger structural change was moving order notifications from SMS to WhatsApp. Presto's POS now calls the messaging platform directly by API to fire a templated, pre-approved "ready for collection" message the moment an order is done, with the outlet name and invoice number filled in per customer. Because customers reply to these on WhatsApp at far higher rates than they ever opened an SMS, the same notification now becomes the start of a conversation the AI team can carry, rather than a dead-end text.
Each completed-order message also increments an order count against the contact, which drives lifecycle tagging automatically: first-time customer on the first order, returning customer on the next, and a loyalty tier beyond a set threshold, all without the team keying anything in.
It is easy to read a line like "a team of specialist agents" as a handful of prompts. It is closer to the truth to say each agent is a digitised slice of how Presto already runs.
Take the Escalation Handler, the agent whose entire job is knowing what it cannot answer. Behind it sits a routing matrix we built directly from the client's own escalation practice, where different problems take very different paths:
Each goes to the specific person who handles it, with the right level of urgency attached, so a chargeback never lands in the same place as a forgotten umbrella. On top of that sits its own logic for when a conversation may be closed and when it must stay open, after-hours scripting and a follow-up cadence, all executed silently so none of it ever surfaces to the customer.
We are deliberately not reproducing those rules here; they are the client's operational IP and ours. The point is the shape of the work. Every agent carries this kind of structure, and arriving at it meant sitting with how the business actually operates: who owns which decision, where the urgency thresholds sit, which items must never be priced from a photo, long before a word of agent instruction was written. Writing the agents was the fast part. Understanding the operation well enough to write them was not.
Delivery ran in clear phases, with a weekly written status report to the client at every step, from setup and discovery through to go-live and ongoing stabilisation.
A deployment at this scale, on a live consumer inbox with a legacy POS and a third-party messaging platform underneath, surfaces a lot. The value has been in catching and fixing issues fast, usually the same day. A representative sample:
The agents repeatedly mishandled dates: offering a public holiday as a delivery slot, miscalculating express-ready dates, or saying an outlet was open on a day it was closed.
We moved to a hard date pre-check that validates every date against the full public-holiday list before the agent is allowed to respond, plus a deterministic time and date check pinned in the agent instructions.
The most instructive problem was conversational bleed. A Receptionist whose only job was to route would sometimes answer the question itself and hand off at the same time, producing a double reply.
The fix was to make routing far more deterministic, silo each agent strictly to its own scope, and add explicit duplicate-prevention and "not your scope" boundaries so it routes silently instead of improvising.
WhatsApp coexistence is powerful but has sharp edges: a reconnect once logged the client out of all linked devices, a changed channel ID broke the POS calls until updated, and getting the POS notifications flowing took genuine debugging of payloads and contact lookups.
We also confirmed the hard way that there is no reliable way to check in advance whether a number is on WhatsApp, so non-WhatsApp contacts route to a human or an SMS fallback.
The system rode out incidents beyond our control, including an upstream model-provider wobble and a platform outage that affected several hundred organisations. Because human takeover is built in, the team could step in and conversations were reassigned once service returned.
The conclusion we reached with the client is the right one: an AI front desk that knows its boundaries beats one that tries to answer everything.
Our customers get an answer the moment they message now, at any hour. The AI handles the routine questions across all twelve outlets, and my team steps in exactly where a person makes the difference.Weitian Chan · Founder, Presto Drycleaners
The points below are directional, drawn from Presto's own operation rather than a controlled study. They describe the shape of the change.
Directional, drawn from Presto's own operation to date rather than a controlled study. Specific figures relating to revenue and database size have been withheld under our confidentiality commitments.
Presto's point-of-sale system now calls the messaging platform by API to fire a pre-approved "ready for collection" message the moment an order is done, with the outlet name and invoice number filled in per customer. Customers reply on the channel they actually read, turning a dead-end text into a thread the AI team can carry. SMS stays on as a quiet fallback for anyone not on WhatsApp.
Prepared by Zelix Labs. The engagement is live and in active stabilisation; outcomes are directional and reflect Presto's own operation to date. Specific figures relating to revenue, customer or database size, and other commercially sensitive details have been deliberately withheld under our privacy and confidentiality commitments. Client quote and this case study published with the permission of Presto's owners.
We design the team, train it on your products and tone, deploy it across the channels your customers actually use, and stay on as the operator.
