How to Build an AI Business Assistant for a Nigerian Company
A general chatbot knows nothing about your business. It cannot tell you which Lagos distributor has not paid in 45 days, how many bookings the Abuja branch has next week, or whether stock of your best-selling item will run out before the next shipment clears at Apapa. An AI business assistant is different: it is connected to your own data and tools, so its answers come from your numbers and its drafts use your context.
This article is written for owners and managers of Nigerian SMEs and mid-sized companies who want an assistant that works for the management side of the business. It is not about customer-facing chatbots, and it is deliberately narrower than an "AI agent" that runs whole processes on its own. For an assistant that staff use to look up HR policies and procedures, see the article on building an internal company assistant; for one that must keep data entirely on your own infrastructure, see the article on private AI assistants.
What is an AI business assistant?
An AI business assistant is a language-model-powered application that is connected to a company's own systems, such as sales records, inventory, accounting, CRM and calendars, so that a manager can ask questions, get summaries, draft communications and trigger routine actions through a chat interface. The difference between it and ChatGPT is the connection: the assistant answers from your data and can act within your tools, under permissions you set.
Three kinds of work sit naturally with a business assistant:
- Ask: "How many units of 50kg cement did we sell in Ibadan last week?" "Which customers have not ordered in 60 days?" "What is tomorrow's delivery schedule?"
- Draft: a payment reminder to a specific customer using their actual balance; a weekly summary for the board; a reply to a supplier email; a job advert based on your existing templates.
- Do (with limits): create a task, log a call note in the CRM, schedule a follow-up, generate an invoice PDF from a confirmed order. Actions with financial or legal effect should require a human tap to confirm.
What it is not: an accounting system, a replacement for your CRM, or a decision-maker. It sits on top of those systems and makes them easier to use.
Which tasks should it handle first?
The first version should handle two or three tasks that are frequent, currently slow, and answerable from data you already have in a system. Use this framework to pick.
| Criteria | Good first task | Poor first task |
|---|---|---|
| Frequency | Asked daily or weekly | Asked once a quarter |
| Data location | Already in a database, accounting tool or clean spreadsheet | Lives in WhatsApp chats and people's heads |
| Risk if wrong | Low; a human checks before acting | High; money moves automatically |
| Time saved | 20–60 minutes a day across the team | A few minutes a month |
Typical strong candidates for Nigerian SMEs: daily sales and cash position summaries, overdue receivables with draft reminders, stock-level questions with reorder alerts, and "summarise everything that happened with this customer". Weak candidates: anything requiring judgement about pricing, credit or hiring.
The building blocks of an AI business assistant
Every business assistant, regardless of who builds it, is made of the same six components. Understanding them lets you brief a developer and evaluate a quotation properly.
- Interface: where you talk to it. For Nigerian owners the practical answer is usually WhatsApp (through the WhatsApp Business Platform) plus a simple web dashboard for longer outputs like reports. Telegram, Slack or Microsoft Teams work where the company already uses them.
- Language model: a hosted model from a provider (OpenAI, Anthropic, Google and others) or an open-weight model you host yourself. The model interprets requests and writes responses; it does not store your data.
- Tools (function calling): defined functions the model may call, such as
get_sales(period, branch),list_overdue_invoices(),draft_reminder(customer_id),create_task(...). Each tool has strict inputs and returns structured data. - Data connectors: the code behind each tool that queries your database, accounting software API, CRM API, Google Sheets or POS export. Read-only wherever possible.
- Knowledge base (optional): documents the assistant can search, such as price lists, supplier contracts, SOPs and policies, retrieved on demand. This is the "retrieval" layer; the knowledge base article in this library covers how to build it properly.
- Permissions, logging and confirmation: who can ask what, which actions need a confirmation tap, and an audit log of every question, tool call and answer.
The model is the least important choice. Connectors, tools and permissions determine whether the assistant is useful and safe.
Step-by-step: building the assistant
The core steps are: define the tasks, inventory the data, design tools, choose the interface, build the orchestration, add guardrails, pilot with two people, then expand. Each step is explained below.
- Write the top ten questions you ask your team every week. Group them into "ask", "draft" and "do". Pick three for version one.
- Inventory where each answer currently lives. Accounting software? POS system? A Google Sheet the operations manager updates at night? A WhatsApp group? Anything not in a system must be moved into one first, or the assistant will guess.
- Design the tools. For each task, define a function with precise inputs (dates, branch, customer) and outputs (numbers, lists, drafts). Keep tools narrow; a tool that "does anything with sales" is hard to control.
- Build the connectors. Query the accounting API, a read-only replica of the sales database, or a Sheets range. Never give the model direct database credentials; it should only call your functions.
- Write the system prompt. Describe the business, units, currencies (₦ by default), branch names, product categories, and the tone for drafts. State clearly what the assistant may not do.
- Choose and connect the interface. WhatsApp for quick questions on the move, a web page for tables and reports. Both talk to the same backend.
- Add permissions and confirmation. Map each user to the tools they may use. Any "do" action sends a confirmation ("Send this reminder to Chukwudi Ventures for ₦1,240,000? Reply YES") before executing.
- Log everything. Question, tools called, data returned, final answer, user, time. This is how you find wrong answers and how you prove what happened.
- Pilot with two people for two weeks. The owner and one manager. Compare answers with the actual reports every day. Fix connectors and prompts.
- Expand tasks and users gradually. Add a task a fortnight. Resist adding actions with financial effect until the "ask" and "draft" layers are trusted.
Connecting the data sources most Nigerian businesses actually have
Most Nigerian SMEs keep their data in a mixture of accounting software, POS exports, spreadsheets and WhatsApp. The assistant can only be as good as the sources it can reach, so plan the connections deliberately.
- Accounting and invoicing tools: most mainstream cloud accounting products offer APIs for invoices, customers and payments. Connect read-only first. The article on connecting AI to accounting software goes deeper.
- POS and inventory systems: many Nigerian retail POS systems export CSV or have a database behind them. A nightly export into a database table is often enough for daily questions.
- Google Sheets: widely used for stock counts, delivery schedules and cash books. Connect via the Sheets API with a fixed range and consistent headers.
- CRM: if you have one, connect customer records, deals and notes. If your CRM is WhatsApp, the assistant cannot read it reliably; consider a real CRM first.
- Bank and payment data: payment providers such as Paystack, Flutterwave and Monnify provide transaction APIs and webhooks that can feed a cash-position summary. Bank statement access is more limited; a manual daily upload is a realistic fallback.
- Documents: contracts, price lists and SOPs go into a knowledge base rather than tools.
One rule above all: fix the data before connecting the AI. If two branches record sales in different formats, the assistant will produce confident nonsense.
Permissions, logging and trust
A business assistant sees payroll-sensitive, customer-sensitive and cash-sensitive information, so access control is not optional. Design it in three layers.
- Who can talk to it: authenticate WhatsApp numbers and web logins; never leave the assistant open to anyone who finds the number.
- What each person can ask: the operations manager may see stock and deliveries but not margins; the accountant may see receivables but not HR notes. Enforce this in the tools, not just the prompt, because prompts can be talked around.
- What it can do without confirmation: nothing that moves money, changes prices, deletes records or contacts customers. Those actions require a human confirmation step and are logged with who approved them.
Logging serves two purposes: catching wrong answers early, and satisfying the accountability expectations of the Nigeria Data Protection Act 2023 when personal data is processed. Keep logs in your own database, retain them for a defined period, and restrict who can read them.
What changes for Nigerian businesses
Four things shape an AI business assistant in Nigeria: WhatsApp as the natural interface for owners on the move, messy and partly offline data, USD-denominated running costs under naira volatility, and the NDPA's rules on personal data.
WhatsApp as the management console. Nigerian owners run their businesses from their phones. An assistant that answers "cash position?" in a WhatsApp message at 7am is used; a dashboard nobody opens is not. Use the WhatsApp Business Platform, keep replies short, and put long tables on a web link.
Data reality. Sales in a POS, expenses in a notebook, deliveries in a WhatsApp group and stock in a spreadsheet is a normal starting point. Budget time for consolidation; it is usually a third of the project.
Power and connectivity. Host the backend in the cloud, not on a laptop in the office. Design the assistant to answer from the last synced data and state the sync time ("as of 6:00pm yesterday") when live systems are unreachable.
Costs in dollars. Model usage and WhatsApp conversation fees are billed in USD. Set monthly caps and prefer smaller models for routine questions; reserve larger models for drafting and analysis.
Data protection. Customer names, phone numbers, balances and staff records are personal data. Send the model only what a task needs, choose providers with clear data-handling terms, and document your basis for processing. Verify current NDPC requirements; this is not legal advice.
Example (hypothetical): a building-materials distributor in Onitsha
Example (hypothetical): A building-materials distributor in Onitsha runs three warehouses, sells on credit to about 200 retailers across the South-East, and keeps sales in a POS system, receivables in cloud accounting software and delivery schedules in a Google Sheet. The owner spends her evenings calling managers for numbers.
Version one of her assistant handles three tasks: a daily sales and cash summary by warehouse, a list of retailers over their credit terms with a draft WhatsApp reminder for each, and stock-level questions with a reorder flag when an item falls below a threshold. The assistant runs on the WhatsApp Business Platform, with tools connected read-only to the POS database, the accounting API and the Sheet. Sending reminders requires her to reply YES per customer, and every action is logged.
The first two weeks expose that one warehouse records cement in bags while another records tonnes; the data is standardised before the assistant is trusted. After that, her evening calls drop to exceptions only. The next additions are a weekly board summary and a supplier price-list knowledge base. This is an illustrative scenario, not a Linestech client result.
How much does an AI business assistant cost in Nigeria?
The cost of an AI business assistant in Nigeria is driven by how many systems it connects to, how clean the data is, how many "do" actions need confirmation logic, and whether the interface is WhatsApp, web or both. The figures below are indicative 2026 ranges; actual quotes vary with scope, vendor and exchange rate.
| Scope | Indicative one-off cost | What it typically includes |
|---|---|---|
| Starter: one data source, "ask" only, web or WhatsApp interface | ₦1,000,000–₦3,000,000 | Tools for 3–5 questions, prompt, logging |
| Standard: 2–4 sources, ask and draft, WhatsApp plus web, permissions | ₦3,000,000–₦8,000,000 | Connectors, knowledge base, confirmation flows |
| Advanced: many sources, actions with confirmation, multi-user roles, custom dashboard | ₦8,000,000–₦15,000,000+ | Integration with custom software, audit and admin tools |
Recurring costs, mostly in USD:
- Model usage per month, which depends on question volume and model size; set a cap.
- WhatsApp Business Platform conversation fees and any provider platform fee.
- Cloud hosting: roughly ₦150,000–₦800,000 per year for a VPS or managed cloud plan.
- Maintenance and prompt tuning: ₦20,000–₦150,000 per month or a retainer.
Data consolidation work (cleaning spreadsheets, migrating from WhatsApp records into a system) is often quoted separately and is frequently the biggest surprise. Ask for it explicitly. Compare two or three written quotations on identical scope: sources, tasks, interface, permissions, logging and ownership of code.
Mistakes to avoid
- Starting with "do" instead of "ask". Automating actions before answers are trusted creates expensive errors. Earn trust with read-only tasks first.
- Giving the model direct database access. A model with raw SQL access can read anything and, if misconfigured, change it. Always go through narrow tools.
- Connecting dirty data. Inconsistent units, duplicate customers and unlabelled sheets produce confident wrong answers. Clean first.
- Skipping permissions because "only I will use it". The number gets shared, staff join, and suddenly a sales rep can see margins. Build roles from day one.
- No confirmation step on customer messages. One wrongly worded payment reminder to a big customer costs more than the confirmation tap saves.
- Measuring nothing. Track questions asked, answers corrected, and minutes saved. Without this you cannot justify version two.
- Building a giant assistant on day one. Three tasks that work beat twenty that half-work.
Conclusion
A useful AI business assistant is built on three decisions: which tasks matter most, where the data actually lives, and what the assistant is allowed to do without a human tap. Start with two or three "ask" tasks, connect clean data through narrow tools, put the interface where you already work (usually WhatsApp plus a web page), and add permissions and logging before anyone relies on it. Indicatively, a serious first version costs from around ₦1,000,000 to ₦8,000,000 in Nigeria, plus USD-denominated running costs, and data clean-up is the most common hidden cost.
If you are deciding which tasks an assistant should handle first and which of your systems can realistically be connected, Linestech can help you map the data sources, define the tools and scope a first version before committing to a build.
Frequently asked questions
What is the difference between an AI business assistant and ChatGPT?
ChatGPT is a general assistant with no access to your systems; you paste information into it. An AI business assistant is connected to your own data and tools, so it answers from your actual sales, stock and receivables and can draft or act within permissions you set. Many business assistants use the same underlying models as ChatGPT; the value is in the connections, not the model.
Can it work on WhatsApp for the owner and managers?
Yes. Using the WhatsApp Business Platform, the assistant can receive questions from authorised numbers and reply with short summaries, with links to a web page for full tables. Only verified numbers should be allowed, and WhatsApp conversation fees apply. The free WhatsApp Business App cannot be used for this.
Does the AI provider see my business data?
When a hosted model answers a question, the specific data included in that request is sent to the provider for processing. Reputable providers publish data-handling terms, including whether inputs are used for training and how long they are retained. Send only what each task needs, review the terms, and if your requirements are stricter, consider a private or self-hosted setup.
How accurate are the answers?
Accuracy of numeric answers depends almost entirely on the tools and data, not on the model. If the tool returns the correct sales figure, the model reports it correctly. Errors arise when data is inconsistent or when the model is allowed to estimate instead of calling a tool. Design the assistant to always use tools for numbers and to say when data is missing.
How long does it take to build a first version?
A starter assistant with one clean data source and a handful of questions can be piloted in two to four weeks. A standard build with several connectors, WhatsApp and web interfaces, permissions and a knowledge base typically takes six to twelve weeks, with data clean-up often the longest part.
Can it replace my operations manager or accountant?
No. It removes the time they spend answering repetitive questions and compiling summaries, and it makes their work visible to you faster. Judgement about credit, pricing, staffing and disputes stays with people. The related article on using AI without replacing staff covers this in more depth.
Should it be built as an AI agent instead?
An assistant answers and drafts on request; an agent runs multi-step processes with less supervision. Most businesses should start with an assistant, prove that data and tools are reliable, and then let selected workflows graduate to agent-style automation. Starting with an autonomous agent on messy data is the expensive route.
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.
