The challenge
Priority One Wash Club sells unlimited carwash memberships in Australia. The offer is simple: Sedan for $59 per month, SUV for $69 per month, 4WD for $79 per month. Unlimited washes, fixed price. The company uses SMS marketing to convert existing customers into a membership.
The problem: more than 3,000 leads had been contacted via SMS, with a reply rate of 6%. That is a healthy percentage for SMS outreach in itself. But the number of conversions was zero. Literally zero.
The SMS agent in use had several structural problems. First: the agent forgot earlier replies. A customer who asked for a payment link in message three received another sales pitch in message four. The context window was too small, so the agent lost track of the conversation after a handful of messages.
Second: when customers explicitly asked how they could pay, the agent did not send a payment link. Instead, it kept on selling. That is the equivalent of a shop assistant ignoring a customer with their wallet in hand in order to hand over yet another product brochure.
Third: the agent could not handle "no". Customers who indicated they were not interested kept receiving messages. That led to frustration and unsubscribes.
Fourth: the agent crashed regularly on Twilio errors. Messages that could not be delivered (wrong number, blocked, timeout) caused fatal failures in the pipeline.
The approach
We rewrote the SMS agent from the ground up with a structured architecture that addresses each of these problems.
State machine. The conversation now follows four phases: WARMUP (greeting, retrieve context), DIAGNOSE (needs assessment, vehicle type), PRESENT (present the right membership with pricing) and CLOSE (send payment link, handle objections). The agent always knows which phase it is in and what the next step is.
Memory fix. The context window was expanded from 20 to 30 messages. That sounds like a small difference, but for SMS conversations that run across multiple days it is crucial. The agent now remembers what the customer said earlier, which vehicle type they have and whether they have already received a price.
Pre-agent router. Before an incoming message reaches the sales agent, it passes through a router that catches two situations: unsubscribe requests and wrong numbers. A customer who sends "stop" is unsubscribed immediately, without the sales agent ever seeing the message. A message from an unknown number is handled separately.
Post-agent classifier. After every conversation, a separate model classifies the lead as warm, cold or unqualified. Warm leads that do not convert right away automatically enter a follow-up workflow.
Automated follow-up. Leads that show interest but do not buy immediately receive a reminder after 24 hours. If there is still no response, a second message follows after 48 hours. No aggressive push, just a friendly check-in.
Twilio error handling. All Twilio-related errors are now caught and logged without crashing the pipeline. Messages that cannot be delivered are flagged and skipped.
The results
Conversion went from effectively zero to measurable results. That may sound modest, but the context is important: these are the same 3,000+ leads that had already been contacted earlier with a broken agent. Many of these contacts had already had a negative experience. The fact that the new agent still pulls conversions out of them says something about the quality of the conversation flow.
Warm leads are automatically re-engaged through the follow-up workflow. No manual work, no forgotten leads, no spreadsheets with "do I still need to follow up on this".
The Twilio crashes are fully resolved. The pipeline runs stably, even at high volumes and with messages that cannot be delivered.
What this means
Priority One did not have a lead problem. They had 3,000+ leads and a reply rate of 6%. That is 180+ people who responded. The problem was in the conversation itself. An agent that forgets context, does not send payment links and cannot stop selling converts no one.
This is a pattern we see often at companies that quickly set up an SMS or chat agent without properly thinking through the conversation logic. The technology works, but the conversation does not. By rewriting the agent with a clear state machine, memory and follow-up, the same leads still turn into revenue.