Customer support on WhatsApp
Pricing, availability, delivery areas and lead times, opening hours, payment options, returns policy — answered instantly from your own source of truth, in the language the customer used.
Service · Customer experience
AI agents that answer your customers in seconds
Your customers are on WhatsApp at 21:40 asking whether you deliver to Kitengela. Right now that message waits until morning. We build agents that answer it immediately, correctly, from your real data — and know exactly when to fetch a human.
WhatsApp Business APIWebsite chatEnglish & KiswahiliLive system accessHuman handover
What is an AI agent for customer service?
An AI customer service agent is software that reads a customer’s message in natural language, works out what they need, queries your live business systems for the answer, and replies conversationally — escalating to a human when the question is outside its remit or the stakes are high. Unlike a menu-driven chatbot, it is not limited to pre-written replies and does not force customers to phrase things its way.
In Kenya this matters more than almost anywhere. Your customers will not fill in a web form and they will not email. They will WhatsApp you — and the business that replies first usually wins the order. An agent that answers in ten seconds at 21:40 is not a technology upgrade; it is a commercial advantage.
Done badly, though, this is the fastest way to annoy a customer ever invented. A bot that loops, misunderstands, refuses to fetch a human, or confidently invents a delivery date costs you more goodwill than having no bot at all. Most of our design effort goes into the failure paths, not the happy path.
An illustration of the interaction pattern we build: understand, check the real system, answer, offer a human.
In practice
What we build agents to do
Almost every deployment starts with one or two of these and expands once the team trusts it.
Order & delivery status
The agent looks up the real order in your system, gives the honest status, and offers the driver’s ETA. This one question is often a third of all inbound messages.
Lead capture & qualification
Ask the qualifying questions, score the lead, book the call into a calendar, and drop a clean record into your CRM — so your sales team calls people who are actually ready.
Internal knowledge assistant
Your staff asking “what is our policy on...?” and getting an answer with the source document cited, instead of interrupting a colleague or guessing.
Payments & account queries
Confirm a received M-Pesa payment, share a statement, or send a payment link — with strict permissions, so an agent can read balances but never move money without human approval.
Bookings & appointments
Check availability, take a booking, send the confirmation and the reminder, and handle the reschedule — for clinics, salons, workshops, tours and professional services.
The same answer, two ways
How an AI agent actually answers a question
It reads the question, goes and looks the answer up, then writes a reply.
When a customer messages, the agent first works out what they are really asking. “Hiyo order yangu iko wapi?” and “where is my delivery” are the same question, so both get the same treatment.
Then it goes and checks. This is the part that separates a useful agent from an irritating one: it does not guess from a script, it looks in your actual system — your orders table, your stock list, your price file — the same place your staff would look.
It writes the reply in your tone of voice, in the language the customer used, and it only says things it can back up. If it does not know, it says it does not know and offers to get someone. It never invents a delivery date to sound helpful.
Everything is recorded. You can read any conversation, see what the agent said and where it got the information, and correct it. Over time you tune it the way you would coach a new employee — except you only have to explain each thing once.
The rule we never break
Anything involving money leaving your business, a legal promise, or an unhappy customer goes to a human. The agent handles volume; your people handle judgement.
Tool-calling agent over a hybrid retrieval layer, with strict output grounding, scoped permissions and deterministic escalation rules.
Channel. WhatsApp Business Cloud API via a registered BSP, with webhook ingestion, session-window awareness and template messages for anything outside the 24-hour window. Web widget over websockets. Both normalise into one conversation model so behaviour does not fork per channel.
Understanding. No intent classifier maintenance treadmill. The model is given a
small set of typed tools — get_order(id), check_stock(sku, location),
get_price(sku, customer_tier), create_lead(...), escalate(reason)
— and chooses among them. Each tool validates its arguments server-side and enforces its own
authorisation; the model never gets raw database access.
Retrieval. Policy documents, FAQs and product copy chunked and embedded, queried with hybrid BM25 plus vector search and a re-ranking pass. Pure semantic search performs poorly on Kenyan proper nouns, SKUs and account numbers, so the keyword leg is not optional.
Grounding. The system prompt forbids answering outside retrieved context; responses carry source references internally for audit. A lightweight post-generation check flags replies containing prices, dates or commitments that do not appear in the retrieved set, and those are held or escalated rather than sent.
Escalation. Deterministic triggers, not model discretion: N failed resolutions, explicit human request, negative sentiment threshold, or any topic on the restricted list (refunds, complaints, credit, legal). Handover carries full transcript and extracted context into your helpdesk.
Evaluation. A regression suite built from real historical conversations with expected outcomes, run on every prompt, tool or model change. Production metrics: containment rate, escalation rate, first-response time, resolution rate, cost per conversation, and a sampled human quality score.
Where the agent sits: your channels on the left, the agent and its guardrails in the middle, your real systems on the right. It only knows what you let it read.
What you get
What you get
A live agent, the controls to run it, and the evidence that it behaves.
- A live agent on WhatsApp, your website, or both.
- Official WhatsApp Business API setup, including verification and green-badge application.
- Integrations into the systems that hold the real answers.
- A human handover path into your existing inbox or helpdesk.
- A tested escalation and out-of-hours policy you signed off.
- A conversation dashboard: volumes, containment, escalations, unanswered questions.
- An evaluation set built from your real conversations, so future changes can be tested.
Typical engagement
Timeline3–6 weeks to launch
Indicative investmentFrom KES 250,000 build, plus per-conversation running cost
Who we work withAny business with high inbound message volume
You ownCode, prompts, docs and accounts — outright
Every quote is fixed before work starts. If we scoped it wrong, that is our risk, not yours. See how pricing works.
Typical stack
WhatsApp Cloud APIMeta BSPClaudeGPTpgvectorPostgresPythonNode.jsTwilioHubSpotZohoFreshdesk
We choose tools your team can maintain, not the ones that make us look clever.
Questions
AI Agents & Chatbots — your questions answered
What is the difference between a chatbot and an AI agent?
A chatbot follows a script: it recognises a keyword or menu choice and returns a pre-written reply. An AI agent understands the question in natural language, decides what information it needs, calls real systems to get it — your stock table, your order database, your price list — and composes an answer from what it found. The practical difference customers notice is that a chatbot forces them to phrase things its way, while an agent handles “niaje, hiyo order yangu imefika?” and gives a real answer.
Can it speak Kiswahili and Sheng?
Yes. Modern language models handle Kiswahili well and cope with the English-Kiswahili code-switching that Kenyans actually type. We tune this deliberately: we test against real message samples from your own inbox, build a glossary of your product and industry terms, and set the tone — formal for a bank, relaxed for a retailer. We will be straight with you that Kiswahili quality is a step below English, so for high-stakes flows we recommend keeping the confirmation step in English or routing to a human.
How much does a WhatsApp AI agent cost in Kenya?
There are two costs. The build typically starts around KES 250,000 depending on how many systems it must connect to and how many flows it handles. Then there are running costs: Meta charges per conversation window through the WhatsApp Business Platform, a Business Solution Provider adds a platform fee, and the AI model itself costs a few shillings per conversation. For most businesses the running cost lands well below the cost of the staff hours it returns. We model this honestly for your volumes before you commit — see our cost guide.
Will it make things up?
Not if it is built correctly. The agent is grounded in your actual data and instructed to answer only from what it retrieves; when it cannot find an answer it says so and offers a human. We test this deliberately before launch with adversarial questions designed to tempt it into inventing prices, policies and delivery dates. We also keep it out of any conversation where a wrong answer is expensive — refunds, credit decisions, legal commitments — by routing those to a person by design.
Do I need the official WhatsApp Business API?
For any serious deployment, yes. Unofficial tools that automate the consumer WhatsApp or WhatsApp Business app violate Meta’s terms and get numbers banned, usually at the worst possible moment. We build on the official WhatsApp Business Cloud API through a registered Business Solution Provider, which also gets you a verified green badge, multiple agents on one number, and message templates you can legitimately use for notifications.
What happens when the AI cannot help?
It hands over to a person, with the whole conversation attached so the customer never repeats themselves. You decide the escalation rules — after a set number of failed attempts, when the customer asks for a human, when sentiment turns negative, or whenever a defined topic like a complaint or refund comes up. Outside working hours it takes a message, sets expectations honestly about response time, and queues it for the morning.
Keep exploring
Other ways we help
Most clients combine two or three of these.
AI Workflow Automation
We find the repetitive work eating your team's week — reconciliation, data entry, approvals, reporting — and hand it to software that never gets tired.
Explore AI Workflow Automation
AI Strategy Consulting
An honest map of where AI will make you money, where it will waste your money, and what to do first — costed, sequenced and tied to your numbers.
Explore AI Strategy Consulting
Custom AI Development
When off-the-shelf will not fit: document intelligence over your own files, forecasting on your own history, and integrations into the systems you already run.
Explore Custom AI Development