How to Automate Customer Questions With AI: One Answer Base for WhatsApp, Website, Instagram and Email

Customers ask the same questions in different places. The same "do you deliver to Port Harcourt?" arrives as a WhatsApp message, an Instagram DM, a website chat and an email, and in many Nigerian businesses each channel is answered by a different person with a slightly different answer. AI can remove the repetition, but only if there is a single, correct set of answers for it to draw on.

This article is about that core: how to build an answer base that works across channels, how to decide what AI should and should not answer, and how to keep it accurate. Channel-specific detail (WhatsApp rules, website widgets) is covered in companion articles; here the channels are endpoints of one system.

What automating customer questions involves

Automating customer questions means an AI assistant reads a customer's question, finds the relevant approved information, composes an answer in natural language and sends it, or, when it should not answer, routes the question to a person with context. The essential difference between this and a traditional FAQ page is that the customer does not have to find the right question; the AI matches their wording, including typos and Pidgin, to the right answer.

A working system has three layers:

• The answer base: approved answers with facts, sources and exceptions, plus live data connections for things like prices, stock or order status.
• The assistant: a language model with instructions, retrieval from the answer base and rules for escalation.
• The channels: WhatsApp (via the Business Platform), a website chat widget, Instagram and Facebook Messenger (via Meta's APIs), and email (via an inbox integration), all connected to the same assistant.

Building the answer base is most of the work. The assistant and channels are increasingly standard components.

Step 1: Collect the real questions from every channel

Do not start from what you think customers ask. Collect what they actually ask.

1. Gather two to four weeks of messages from WhatsApp, Instagram DMs, Facebook, email, website forms and, if possible, notes from phone calls and walk-ins.
2. Anonymise names, numbers and other personal details before analysis. This is both good practice and consistent with the Nigeria Data Protection Act 2023.
3. Extract the questions in the customer's own words. Keep the wording; it becomes test data and shows how people phrase things ("how much last", "abeg do you deliver to PH", "is it original").
4. Record the channel, the time and how it was answered, and whether the answer needed a lookup, a decision or a person.

An AI assistant can help extract and cluster questions from anonymised exports. Check its clusters by hand; the errors are informative.

Step 2: Group questions and grade answerability

Group the questions into topics (pricing, delivery, availability, payment, process, policy, identity and trust, complaints, advice). Then grade each group on answerability.

Most Nigerian SMEs find that grades A and B cover the majority of their volume. That is the launch scope. Grade C questions are worth writing rules for because they are often high-volume (delivery fees especially).

Step 3: Write answer units, not FAQs

A traditional FAQ is a question with a paragraph. An answer unit is a structured record that an AI can use reliably. For each topic in grades A to C, write:

• Approved answer: the wording you would be happy to see sent to a customer, in Standard Nigerian English, short and direct.
• Facts and source: the specific facts (fees, times, addresses) and where they come from (a price list, a policy document, a system). If the fact lives in a system, note the lookup rather than the value.
• Variants: the different ways customers ask (from Step 1), including Pidgin and typos. These improve retrieval.
• Exceptions and escalation: when the answer does not apply and what the AI should do instead.
• Owner and date: who maintains this unit and when it was last checked.

Example of an answer unit for a delivery question (illustrative): approved answer states delivery days and fee bands by zone; facts point to the delivery table; variants include "how much to Ajah", "you dey deliver to Ibadan?"; exceptions say that any address outside the listed zones is escalated for a manual quote; owner is the operations lead; checked monthly.

Also write the "identity" units that Nigerian customers ask for: CAC registration status, physical address, whether payment on delivery is available, and how to verify the business. These answer the "is this real?" question that precedes many purchases.

Step 4: Decide what AI must never answer alone

This is a policy decision, not a technical one, and it should be written down and approved by the business owner.

Typical restricted categories for a Nigerian business:

• Health, legal, financial or safety advice. A pharmacy's bot can state opening hours and whether a product is stocked; it must not advise on dosage. A lab's bot can explain test preparation; it must not interpret results.
• Complaints and disputes. Acknowledge, apologise, collect details, route to a person.
• Payment confirmation when the AI cannot verify the payment through a gateway or bank data.
• Negotiation. "How much last?" gets a polite fixed-price answer or a handoff, not an AI-invented discount.
• Regulated claims. Anything about product efficacy, certifications or guarantees not in the approved answer base.
• Personal data changes. Address, account or contact changes require verification by a person or a secure process.

For each restricted category, write what the AI should say and where it should route. Test these paths deliberately.

Step 5: Deploy one assistant across channels

With the answer base ready, connect it to the channels. The principle is one assistant, one answer base, many endpoints.

• WhatsApp: through the WhatsApp Business Platform, with a shared inbox for escalations. Respect Meta's template and opt-in rules for anything you initiate.
• Website: a lightweight chat widget that loads after the main content and can hand off to WhatsApp or email.
• Instagram and Facebook Messenger: through Meta's messaging APIs, often via the same provider as WhatsApp. For Instagram-dependent Nigerian brands, this is frequently the highest-volume channel.
• Email: an integration that drafts replies for a person to approve, or auto-replies for grade A questions with a clear "reply to reach a person" line.

Configure the assistant's instructions once: tone, disclosure that it is automated, the restricted categories, and the escalation wording. Channel-specific differences (message length, buttons, links) are handled by the channel layer, not by different answer bases.

Escalations from every channel should land in the same inbox with the same tagging, so a person sees one queue rather than four apps.

Step 6: Keep the answer base current

An answer base decays as prices, policies and stock change. Build the maintenance into the operation.

• Weekly transcript review: the knowledge owner reads a sample of AI answers and every escalation, marks wrong or missing answers, and updates units.
• Change triggers: any price change, new product, policy update, new delivery zone or promotion updates the relevant unit before it is announced.
• Monthly accuracy check: run the original test questions from Step 1 through the assistant and score the answers.
• Quarterly re-grading: move topics between grades as data connections improve (for example, once order status is integrated, "where is my order" moves from D to B).

Assign the owner role explicitly. In many small businesses it is the operations manager; in larger ones it may be a customer service lead.

What changes for Nigerian businesses

Instagram DMs are a primary channel. Many Nigerian consumer brands receive more questions on Instagram than anywhere else. Include those DMs in the collection and connect the assistant to Instagram messaging, not just WhatsApp.

Price questions dominate and need a policy. Decide whether the assistant states exact prices, ranges or "from" prices, and how it handles "how much last". Nigerian customers disengage when price is hidden, so where you can state it, do.

Trust questions are frequent. "Are you real?", "do you have a shop?", "can I pay on delivery?", "send your CAC". Write these units carefully and truthfully.

Language mix. Collect and test Pidgin and mixed-language variants. Standard Nigerian English answers are usually right for the reply; understanding the question is where variants matter.

Delivery answers are conditional. Write the rules (zones, weight bands, partner timelines) and let the assistant apply them. Escalate unusual addresses rather than guessing.

Connectivity. Hosted assistants keep answering during office outages; make sure escalations reach a person with mobile data.

Data protection. Questions may include personal data; publish a privacy notice, limit transcript access, and check where your platform stores data. NDPC guidance is the reference.

Example (hypothetical): a diagnostic laboratory in Abuja

Example (hypothetical): a diagnostic laboratory with two branches in Abuja receives questions by WhatsApp, Instagram, website form and phone: test prices, whether fasting is required, how long results take, whether home sample collection is available in a given district, branch hours and locations, and, frequently, requests to interpret results.

Collection and grading produce this picture:

• Grade A: hours, locations, payment methods, home-collection districts.
• Grade B: test prices and packages (from a price list), result turnaround per test type.
• Grade C: home-collection fees by district and time slot; preparation instructions per test (fasting, timing).
• Grade D: bulk or corporate pricing, urgent turnaround requests.
• Grade E: anything asking what a result means, whether a symptom is serious, or which test to take for a condition.

The lab writes answer units for A to C with the medical director approving the preparation instructions, and a restricted-category script for E: the assistant explains that results are interpreted by a clinician, offers to book a consultation or connect to the lab's staff, and never comments on values. One assistant is deployed to WhatsApp, Instagram and the website, with escalations landing in a shared inbox monitored by the front desk.

The weekly review catches two early issues: customers writing test names in abbreviations the answer base lacked, and a district added to home collection without updating the unit. Both are fixed by editing the answer base, not the technology.

The point of the example is the restricted category: for a healthcare business, deciding what the AI must not answer is as important as what it can.

Costs and tools

Indicative 2026 ranges; actual quotes vary with scope, vendor and exchange rate. Compare two or three written quotations on identical scope.

A hosted platform is the quickest route for grades A and B across channels. A custom build makes sense when you need lookups from your own systems, strict control over restricted categories, or data hosted where you choose.

Mistakes to avoid

• Starting from an imagined FAQ. Collect real questions; the gap between what you think and what they ask is large.
• A different answer base per channel. Inconsistency confuses customers and doubles maintenance.
• Letting the model answer from general knowledge. It will invent prices, policies and, in regulated sectors, dangerous advice. Restrict it.
• No restricted-category policy. Decide what AI must not answer before launch, and test those paths.
• No owner. The answer base is out of date within weeks without one.
• Ignoring Instagram. For many Nigerian brands it is the biggest source of questions.
• Hiding the human path. Every channel needs a clear way to reach a person, and someone must answer.
• Never re-testing. Run the original questions monthly; accuracy drifts as the business changes.

Conclusion

Automating customer questions with AI succeeds or fails on the answer base. Collect the real questions from every channel, grade them for answerability, write structured answer units with facts, variants, exceptions and owners, decide explicitly what the AI must never answer alone, and deploy a single assistant to WhatsApp, the website, Instagram and email with one escalation inbox. Then keep it current with a weekly review and a monthly re-test. Get the base right and the channels become straightforward.

If you want help building an answer base and deploying one AI assistant across your customer channels with the right restrictions for your sector, Linestech can scope it with you.

Frequently asked questions

Do I need a separate chatbot for each channel?

No. The better design is one assistant with one answer base connected to WhatsApp, website, Instagram, Facebook and email. Channels differ in formatting and rules, but the answers should be identical. Many platforms support multiple channels from one configuration, and custom builds are designed the same way.

How many questions should I write answers for before launching?

Enough to cover the majority of your volume, which for most businesses means the top twenty to forty question groups in grades A to C. Launch with those, route everything else to people, and add answer units from the weekly review. A small, accurate base beats a large, unchecked one.

Can AI answer pricing questions accurately?

Yes, if prices come from a maintained list or system that the assistant looks up, and if the policy on ranges, "from" prices and negotiation is written. It cannot answer accurately if it is left to guess. Date the price list and assign an owner.

What about questions in Pidgin or local languages?

Include real Pidgin and local-language variants in the answer units so retrieval matches them, and test the assistant with those messages. Reply in Standard Nigerian English unless your brand voice and audience call for something else. Route local-language conversations to people if accuracy is poor.

How do I handle questions the AI cannot answer?

The assistant should say clearly that it does not have that information, collect any needed details, and route the question to a person with a promised response time. The weekly review then decides whether the question deserves a new answer unit. Silent failures or invented answers are the two outcomes to avoid.

Is a FAQ page on my website still useful?

Yes. It serves search engines and customers who prefer to read, and it can be generated from the same answer base so it stays consistent with the assistant. It also supports FAQ structured data where the questions and answers are genuinely visible on the page.

How long does it take to set up?

Collecting and writing the answer base typically takes two to four weeks of part-time effort. A hosted platform can be connected to channels in days once the base exists; a custom build with data connections takes one to three months. The writing, not the technology, sets the pace.

Sources and further reading

Figures, platform rules and regulations change. These are the primary references behind this article and the places to check before you act on it.

• Corporate Affairs Commission (CAC)
• Nigeria Data Protection Commission (NDPC)
• WhatsApp Business Platform — Meta for Developers
• OpenAI platform documentation
• Anthropic developer documentation