AI WhatsApp customer assistant, English and isiZulu
Sophia. AI that answers your customers on WhatsApp.
Trained as an agent of your business. Sophia works out what the customer is actually asking and reasons its way to the right answer, for South African businesses, in English and isiZulu.
Sophia is a Synchrotech product: we build it, own it and run it, and make it available to businesses under a support plan.
An example conversation
Customer
Hi, is the Main Road branch open on a public holiday?
Sophia
Yes, Main Road trades on public holidays from 8am to 5pm, then resumes normal hours the next day.
Customer
Great, and can I still use my savings card there?
Illustrative
Not a scripted chatbot
Not a scripted WhatsApp chatbot.
A standard chatbot
Matches keywords to scripted answers
Sends a generic reply from a prompt
Breaks when the customer goes off-script
Treats every conversation the same
Sophia
Understands the question in context, including follow-ups and unclear wording
Trained as an agent of the business, answering on its behalf
Thinks through new questions instead of replying "I don’t understand"
Shapes each response to the customer's situation
In production
Already talking to customers
Sophia answers questions on WhatsApp for a multi-branch South African retailer, in English and isiZulu, covering store hours, savings cards, specials and returns. Between 5 and 7 percent of chats reach a person; Sophia resolves the rest on its own.
5.59m
Campaign messages sent
161 campaigns since November 2024
104 556
Opted-in contacts
Reachable in a single send
55 823
Customer conversations
About 182 000 messages answered
96%
Of replies written by AI
Measured on a 240 000 message sample
Inside Sophia
The console the business works in
Sophia is not a black box. Every conversation, limit and permission sits in a console your team signs into.
Campaigns: every message template approved before it can send, customer copy redacted
Handover: the conversation a person takes over, customer details redacted
Authoring: a new template composed against a live preview
Limits: message thresholds and spend caps, per customer and per day
Access: roles and permissions, so only the right people review conversations
How a reply is made
Five steps, and a check before anything is sent
No trade secrets here, just the design. Each step is a separate job, so a weak answer is caught before a customer sees it.
Understand
Sophia works out what the customer is really asking, in the context of the conversation.
Look it up
The answer is retrieved from the business's own approved knowledge base, found by meaning rather than keywords, and never scraped from the open internet.
Draft
A dedicated model writes the reply using only what was found.
Check
Independent inspectors verify the draft is backed by the source material, including one running on a different AI provider, so no single model marks its own homework.
Send or escalate
Verified replies go out in seconds. Anything doubtful or sensitive goes to a person.
What holds it in check
Limits the system enforces
Grounded, not improvising
Every answer is built from the business's approved playbook: branch details, specials, policies. If the playbook cannot back an answer, Sophia does not invent one.
Guardrails and circuit breakers
Opt-outs are detected and honoured automatically. Per-customer and platform-wide send limits stop runaway messaging. Public social replies wait for a person to approve them.
Teach it today, roll it back tomorrow
The knowledge base is versioned, 48 saved versions so far. An update is prepared, scheduled and activated, and any previous version can be restored, so when a special ends the old promise ends with it.
Everything leaves a paper trail
More than 170 000 audit entries record who and what touched each conversation. Every reply is stored with its category and the reasoning behind it, so any answer can be traced afterwards.
Roles decide who may change what
Staff access is controlled by role, with system roles kept separate from the roles the business defines itself, and the same roles govern what Sophia may do on its behalf.
Built to absorb a busy day
Sophia runs serverless across multiple isolated data centres with automatic failover, defined by 21 infrastructure blueprints. Queues absorb campaign days of 100 000 messages, and failed deliveries are retried rather than lost.
What it takes off the team
The inbox, before and after
Before
Two hours a day, four on Mondays
One person worked through the inbox by hand. Weekend questions waited until Monday morning and piled into the week's biggest backlog, and peak season doubled everything. Baseline supplied by the team.
After
Between 5 and 7 percent reach a person
Sophia answers in a median of 14 seconds, measured across 5 736 reply pairs, day and night and over weekends. The team only sees the handful of chats that genuinely need a person. Nearly two in five questions arrive outside trading hours.
650 to 700
Staff hours back a year
Estimate, from the team's own baseline at the measured escalation rate
About R900
Monthly running cost
Estimate, about 37 cents per conversation
11.9m
WhatsApp events processed
Over the past 14 months
Method. Measured figures are pulled read only from the production account on 16 September 2026. Figures marked Estimate carry their assumptions: the staffing baseline was supplied by the team, and hours saved apply the measured escalation rate of 5 to 7 percent to it.
Built with care
Limits, handover and a record of every conversation
Clear limits on what Sophia can do and say
Send limits per customer and across the platform, opt-outs honoured automatically, and nothing posted publicly without a person approving it.
Handover to a person when a conversation needs one
Anything doubtful or sensitive goes to a person, and between 5 and 7 percent of chats do.
Conversation records the business can review
Every conversation is recorded with its category and the reasoning behind each reply, visible to the staff whose role allows it.
Questions
Straight answers about Sophia
Which languages does Sophia speak?
English and isiZulu, in production on WhatsApp for a multi-branch South African retailer. It works out what a customer is asking in the context of the conversation, including follow-ups and unclear wording.
What happens when Sophia is not sure?
It hands the conversation to a person. Anything doubtful or sensitive is escalated rather than guessed, and between 5 and 7 percent of chats reach a person; Sophia resolves the rest on its own.
What does Sophia cost to run?
About R900 a month at the measured volume, roughly 37 cents per conversation, as an estimate from the production account. Sophia is a Synchrotech product made available under a support plan; the process review that scopes it is free.
How does Sophia handle personal information and opt-outs?
Opt-outs are detected and honoured automatically. Sophia answers only from the business's own approved knowledge base, never the open internet, and every conversation is recorded with its category and reasoning so it can be reviewed by the staff whose role allows it. Personal information is handled in line with POPIA, as with every system Synchrotech runs.
See what Sophia could do for your customers.
Book a free process review with the team who would build it.