Buying a WhatsApp customer-service assistant in Zimbabwe: a walkthrough

Customer service in Zimbabwe happens on WhatsApp. This guide walks through buying an assistant for that channel: what the numbers say about the channel, the one technical question that decides whether the thing is legitimate, four ways to buy it, the licence you must hold, and how to run a pilot that tells you the truth.

The channel, in the regulator’s numbers

POTRAZ’s fourth-quarter 2024 report put active mobile subscriptions at 15.68 million (a penetration rate of 102.26%) and active internet subscriptions at 12.49 million (81.49%). By the first quarter of 2025 active mobile subscriptions had reached 15.89 million, mobile data traffic was 114 petabytes for the quarter, and data earned operators 50.28% of their revenue — more than voice for the first time. A mid-2026 industry summary citing late-2025 data counted 16.2 million active cellular connections and 6.54 million individuals online (38.4% of people, a stricter measure than subscriptions). Whichever measure you take, the messaging channel is where your customers already are; the phone is a data device first.

The consequence for a buyer: the assistant’s home is a WhatsApp number, and every other channel (web chat, Facebook, email) is secondary.

The one technical question

There are two ways to put software on a WhatsApp number. The WhatsApp Business Platform (Meta’s official Cloud API) is the supported route: your business is verified by Meta, the number is registered to your business, conversations are metered and paid. The other route is unofficial — automation that drives a phone or a reverse-engineered client. It is cheaper and Meta can ban the number without warning, taking your customer history with it.

Ask every vendor: Is this the official WhatsApp Business Platform, and will the business verification and the number be in our name? If the answer is anything but a plain yes, the price does not matter.

Four ways to buy

Shape	What you get	Fits	Listed local reference price
Subscription tool	Chatbot builder, inbox, tags, maybe POS and invoicing	Micro and small firms; FAQs, hours, catalogue, order capture	Zimflow: US$20–50/month (listed); Ngaatec: US$99 or US$299 one-off packages with conversation caps (listed)
Pilot (S-02)	Four weeks on one number with a human watching every hand-off	Anyone unsure how their customers write	Quoted; fixed scope
Implementation (S-03) + Integration (S-04)	Assistant that reads stock, orders, bookings; logs to CRM	Firms with a POS/CRM and real volume	Quoted per milestone and per connector
Managed support (S-06)	Monitoring, monthly report, change budget	Once it faces customers 24/7	Retainer, usually with capped usage pass-through

Most readers of this site should buy the first row and stop. The second row is for anyone about to spend on the third.

What the assistant should and should not do

Should: answer the twenty repeated questions from a list you edit; collect name, location, product and quantity; hand over to a person with the whole thread; log everything. Should not: invent prices; promise stock it cannot see; loop; message people who never wrote to you.

The hand-off is where demos are won and lost. Insist on seeing the failure path: an angry message, a refund question, a Ndebele voice note. The good systems say “a person will reply” and mean it; the bad ones offer the menu again.

The licence you must not skip

If you hold customer names and numbers — which the assistant will, by design — you are a data controller under Zimbabwe’s Cyber and Data Protection Act. Statutory Instrument 155 of 2024 required data controllers to be licensed by POTRAZ by 12 March 2025, in tiers set by the number of data subjects (a threshold from 50 data subjects at the lowest tier), with annual licence fees the regulations place between US$50 and US$2,000 depending on tier, a licence valid for twelve months and renewable three months before expiry, and a named Data Protection Officer. Processing without a licence is an offence.

Practically: get your licence and DPO in order before go-live; write the vendor’s role (processor acting on your instructions) into the contract; ask where the conversation logs are stored and how you delete a customer’s data end to end. The data section of the question bank has the eight questions.

A four-week pilot plan

Week 0 — baseline. For five working days, count enquiries received and enquiries answered within one hour, by hour of day. That gap is your business case and your success test.

Week 1 — the list. Write the thirty most common questions and the approved answers, in the languages customers use. Decide the hand-off person and hours.

Weeks 2–3 — live on one number. Assistant answers; a person reads every conversation daily and notes every wrong or awkward answer. Fix the list, not the model.

Week 4 — measure. Same counts as week 0. Add: % of conversations resolved without a person; average hand-off time; number of wrong answers found. Decide: keep the tool, scale to the build, or stop.

Measure	Week 0	Week 4	Decision rule (example)
Enquiries answered within 1 hour	your count	your count	Keep if improved by 50%+
Resolved without a person	0%	your %	Above 40% justifies the tool; above 60% consider integration
Wrong answers found by review	—	your count	Any wrong price = fix the list before scaling
Hand-off time	—	minutes	Under 15 minutes in hours; a promise outside hours

Where to go next

The customer-service problem sheet has the vendor questions and the effort spec. The category explainer for what an AI agent is, versus a chatbot or a plain automation, is at aiagent.co.zw; this guide deliberately treats the assistant as a service shape rather than a category. And the providers listing shows who sells which shape today, with sources and dates.