70% of inbound customer service replaced by AI in 4 weeks for a fast-growing London Pilates group — without changing a single tool the team uses

The situation

A multi-site reformer Pilates group operating across central and Greater London, growing fast on the back of a strong intro offer and waitlists at peak times. Customer-service traffic was running at 50–70 new conversations a day before follow-ups, split across email and WhatsApp. The team replied manually — even with an inbox that offered AI-suggested replies, every send still required a human to read and approve. WhatsApp wasn't connected to anything at all: two separate WhatsApp Business accounts (CS and sales) being managed by hand, with no shared visibility. The Customer Service Lead trained both the CS and sales teams personally, which made her the single bottleneck on tone, on edge cases, and on response time. The MD wanted growth to stop adding headcount — every new site was adding inbound load the small CS team couldn't absorb, and the existing "AI" was capturing none of the time savings it implied.

The brief

Build an AI customer-service agent that genuinely handles repeated questions across email and WhatsApp end-to-end — no human in the loop on the easy stuff — without changing how the team works in Front, and without introducing a new system to learn. Get the team out of after-hours triage.

What we did

• Mapped inbound traffic into five message types — FAQ, Action Needed, Complaint, Human Needed, Casual — so the agent could decide what to handle alone, what to escalate, and what to ignore.
• Audited the existing stack — Front (email), WhatsApp Business (manual, two accounts), Mindbody (booking + member data), Attentive (CRM/SMS) — and explicitly chose not to add a new platform; the agent had to fit into Front, not replace it.
• Ingested the client's CS Bible, sales playbook, FAQs, tone guides and onboarding materials; reformatted them for retrieval, ran them past the CS Lead for accuracy, and corrected outdated content (credit expiry, intro-offer mechanics, escalation routing) before training the agent.
• Built the agent on a retrieval-augmented architecture — it only ever answers from Level Out's own documents. No general-purpose LLM responses. No hallucinations.
• Connected WhatsApp via Twilio's native Front integration — chosen over Respond.io specifically because it kept the team in their existing inbox. Three days to build, one day to test end-to-end.
• Configured automatic classification, tagging and escalation: FAQs go out without a human ever seeing them; Action Needed, Complaints and Human Needed land in Front already tagged, already routed, with the full thread and a draft reply.

Engagement length: 4 weeks (17 Mar → 10 Apr 2026) · Deliverable: Production AI customer-service agent on email + WhatsApp, integrated with Front

What we found

1. 10–15 questions were eating most of the inbox. Pricing, intro-offer mechanics, package pauses, pregnancy guidance, credit expiry and "I'm running late" accounted for the bulk of 50–70 daily conversations — every one of them being typed out by a human.
2. The existing "AI suggested replies" captured zero time. The inbox was generating drafts, but the team still had to read, approve and send each one. Every reply still cost a full human touch.
3. WhatsApp was off-grid. Two separate WhatsApp Business accounts being run from people's phones with no tagging, no audit trail, no shared visibility. Anything that came in over WhatsApp lived only in whoever picked it up.
4. The CS Lead was the bottleneck on the entire customer experience. She trained both the CS and sales teams personally, owned every edge case, and was the only person who could sign off on tone — meaning every new hire and every tricky reply funnelled through one person.
5. Roughly 14 hours/week of manual reply time was sitting inside the inbox waiting to be removed. 70% of ~60 conversations/day at ~3 minutes per touch, equivalent to roughly half an FTE before any growth in volume.
6. Mindbody's December 2025 webhook fix had quietly opened a new door. Until late 2025, automating actions inside Mindbody (pause package, cancel/rebook class) was a non-starter because the integration was unreliable. That's no longer true — making a phase-2 "agent that takes the action, not just the message" viable for the first time.

What we recommended

The outcome

• Live on Friday 10 April 2026, four weeks after the kickoff call, across email and WhatsApp simultaneously.
• FAQ-class messages now answered automatically on first contact, including overnight and at weekends — the operational pattern that was previously eating into the MD's and CS Lead's evenings.
• Action Needed and Human Needed messages auto-escalate inside Front with the full thread and a draft response. The team starts each ticket already in context — triage time effectively zero.
• Phase-2 scoped: AI taking the action, not just the message. Mindbody write-actions (pause package, cancel/rebook class) — the bulk of currently-escalated requests — unlocked by Mindbody's December 2025 webhook fix.
• CS Lead's role shifting from manual responder to agent-knowledge owner — meaning the next site opening doesn't add proportional CS headcount.

Status: 3 of 3 in-scope components live · Phase-2 (Mindbody write-actions) scoped, awaiting client go-ahead