Starting point

Sample Home Systems Ltd. (fictional) is an HVAC business with 14 employees and a two-person office team that handles the entire customer communication alongside scheduling and invoicing. Customers reach out however they prefer: by phone for acute problems, by WhatsApp with a photo of the broken boiler, by email for quote requests.

The problem wasn’t any single channel, but switching between them: to reply on WhatsApp, the office team first had to check the ERP to see whether the customer was already a customer and what was last agreed, so a program switch, a search, back, then reply. At over 100 messages a week across all channels, that added up to a noticeable share of office time, without ever creating a feeling of control over open cases at the end of the day.

The setup (architecture as a list)

A Tuesday with the system

8:10 am: a WhatsApp message with a photo of a leaking boiler arrives. The conversation immediately shows: existing customer, last service 14 months ago. The AI suggests an appointment with the next available HVAC technician, the office team confirms it with two clicks.

10:45 am: an email inquiry for a quote on a bathroom renovation comes in. The AI draft already includes the standard assessment questions, the office team adds two individual points and sends it.

1:30 pm: a call reports “no hot water anymore”, the escalation keyword triggers, the conversation is shown with priority immediately, even before the office team goes through the list.

4:00 pm: at the end of the day, the office team sees a single list of all open cases across all three channels, instead of clicking through three separate inboxes one after another.

Model calculation

AssumptionValue
Incoming messages/week (call + WhatsApp + email)138
Answered from AI draft108 (78%)
Individually written from scratch (complaints, special cases)30 (22%)
Avg. time saved per message from AI draft (no system switching)5 min.
Office time saved/week≈ 540 min. → 9 hrs
Channels in one inbox3

The 9 hours are a model value based on the average time saved per message, not a stopwatch measurement. For the individually written 22%, the model assumes no time saving, even though the shared inbox shortens the search for the right context there too.

Limits & learnings

Draft mode was a deliberate choice from the start, not auto mode: at a home-systems business with technical follow-up questions, the team wanted to see every reply before it was sent. The WhatsApp connection needed about two weeks of lead time for the official Meta Business approval; that was communicated from the start, not discovered mid-project. And the escalation keywords needed one round of tuning after the first two weeks, because “no water” was initially too broad and also flagged harmless requests as priority.