Outsourcing AI Development in Nigeria: How to Contract It

Nigerian businesses are increasingly commissioning AI work, and the contracts they use are usually copied from website or software projects. That mismatch causes most of the disappointment. "The chatbot must answer customer questions" is not an acceptance criterion — it has no pass mark, no test set and no definition of a wrong answer.

This guide covers what AI outsourcing actually involves, how to write criteria that can be tested, what the ongoing costs look like in naira and dollars, what you must own, and how to structure a pilot before committing serious budget.

What outsourcing AI development actually covers

"AI project" covers very different kinds of work, with very different risk profiles. Be specific about which you are buying.

• Rule-based assistant. Decision trees and keyword matching. Predictable, cheap, limited — often the right answer for a narrow enquiry flow.
• Language-model assistant with a knowledge base. Answers customer questions using your own documents, prices and policies. The most common Nigerian business AI project.
• AI agent with system integrations. Reads and writes to your systems: checks order status, books an appointment, raises a ticket. Higher value and higher risk, because it acts rather than only answers.
• Document and data processing. Extracting fields from invoices, receipts, statements or KYC documents, then feeding them into your systems.
• Prediction and scoring. Demand forecasting, churn risk and similar work, which requires substantial historical data and careful governance in regulated contexts.
• AI embedded into existing software. Adding drafting, summarising, search or classification to a system you already run.

The engineering effort differs, but so does the governance. An assistant that answers questions can be wrong and embarrassing. An agent that takes actions can be wrong and expensive. Scope the oversight to match.

Why AI projects need different acceptance criteria

Traditional software either works or does not: the form submits or it does not. AI systems produce outputs that are correct most of the time, with a distribution of errors. A contract that ignores this cannot be enforced by either side.

Replace feature-based acceptance with measured performance:

The essential elements are: a test set drawn from real data, a pass mark agreed in advance, a definition of failure (especially confidently wrong answers, which are worse than admitting uncertainty), and who judges. Your team should judge, not the vendor.

Build the test set before the vendor starts, and keep part of it private so the system is not tuned to the exact questions it will be graded on.

Data readiness: the work you must do first

AI quality is bounded by the information you can give it. This is client-side work and it cannot be outsourced meaningfully, because only you know which version of a document is current.

A readiness checklist:

• Source documents identified — price lists, service descriptions, policies, FAQs, manuals, terms.
• A single current version of each, with outdated copies removed. Conflicting documents produce conflicting answers.
• A named owner for each document, responsible for keeping it current after launch.
• Historical examples where relevant — past enquiries and the correct answers, or sample documents with correct extractions.
• Data cleaned of what should not be exposed — staff personal details, internal margins, customer records that need not be included.
• Access to the systems the AI must read or write, with someone who can authorise integration.
• A test set of real questions or documents, with correct answers written by your team.
• An update process for keeping the knowledge base current — weekly, monthly, or triggered by price changes.

A realistic rule: if your policies live in several conflicting WhatsApp messages and an outdated PDF, expect to spend two to four weeks organising them before any vendor can do useful work. Pay for that time in your own plan rather than discovering it as a project delay. How to Prepare for an AI Integration Project.

Indicative costs, and the running cost nobody budgets for

Indicative 2026 naira ranges. Actual quotes vary with scope, vendor and exchange rate.

The item most often missing from Nigerian AI budgets is model usage. Most language-model services charge per unit of text processed, priced in US dollars and billed monthly. Your cost therefore depends on how many conversations or documents you process, how much context each one carries, and the exchange rate on the day you are billed.

What that means practically:

• Ask the vendor to estimate monthly usage cost at three volumes: expected, double and five times expected, and require a spending cap with alerts from day one.
• Understand that a successful launch increases your running cost. Growth is billed.
• Add platform costs beyond the model — hosting, search infrastructure, monitoring — and budget maintenance: knowledge-base updates, tuning, evaluation runs and model version changes.

How to Budget for AI Integration in Nigeriaia cover the ongoing picture. Treat all figures as indicative and confirm current model pricing with the provider directly.

What you must own: keys, prompts, knowledge base and evaluation set

AI projects create a distinctive form of lock-in. Guard against it in the contract.

Two practical clauses to insist on. First, portability: the vendor states which model provider is used and confirms that the prompts, knowledge base and integration code are transferable to another provider, with an estimate of the effort. Second, no hidden markup: if the vendor bills you for model usage rather than you paying the provider directly, require the underlying usage report each month.

Contract it as a pilot before a build

The highest-return decision in AI outsourcing is to buy a small pilot first. AI feasibility depends on your specific data, and nobody — including honest vendors — can promise results before seeing it.

A workable structure:

Phase 1 — Pilot (2–4 weeks, indicatively ₦400,000–₦1,500,000). A narrow scope: one use case, a subset of documents, a test set of real questions, a measured accuracy result and a written recommendation. Deliverable is evidence, not a product.

Phase 2 — Build (if the pilot passes). Full scope, integrations, oversight design, monitoring, accuracy targets in the contract, staged payments.

Phase 3 — Operate. Monthly support covering knowledge-base updates, evaluation runs, prompt tuning, model version changes and usage monitoring.

The pilot does three things at once: it tells you whether the idea works on your data, it shows you how the vendor actually works, and it produces the test set and baseline you will use for the rest of the project. A vendor who resists a paid pilot in favour of a large fixed contract is asking you to carry all the uncertainty.

Measuring accuracy and designing human oversight

Decide oversight by consequence, not by enthusiasm.

Build these into the contract as behaviour requirements: what the system must refuse to do, when it must hand over to a person, what it says when it does not know, and how the handover reaches a real human — in Nigeria, usually WhatsApp or a phone call rather than an email ticket.

Then require ongoing measurement: a monthly evaluation run against the test set, a sample of real conversations reviewed by your team, a log of escalations, and a record of what changed and why. Without measurement, quality drifts silently as your prices, policies and product range change.

Data protection and sending Nigerian data to foreign models

If your AI system processes the personal data of people in Nigeria, the Nigeria Data Protection Act 2023 applies to your business as the data controller, regardless of who built the system or where the model runs.

Settle these points in writing before signing:

• What personal data reaches the model provider. Names, phone numbers, addresses, account details, health or financial information. Minimise what is sent.
• Where processing happens and what the model provider's terms say about retention and use of your inputs for training. Providers differ; read the specific terms of the one proposed.
• Whether customers are told they are interacting with an AI system, and how they reach a human.
• Retention of conversation logs — how long, where stored, who can access them.
• Deletion on request and what happens to data at contract termination.
• Sector rules. Financial services, health and education carry additional expectations from bodies such as the Central Bank of Nigeria and relevant regulators.

Verify current requirements with the Nigeria Data Protection Commission (https://ndpc.gov.ng/) and take qualified advice. This article describes the issues; it is not legal advice.

What changes for Nigerian businesses

WhatsApp is where the assistant must live. For most Nigerian consumer-facing businesses, an AI assistant that only exists on the website misses the channel customers actually use. Scope explicitly for the WhatsApp Business Platform, and note that message delivery through it carries its own fees on top of model usage. Confirm current terms in Meta's WhatsApp Business Platform documentation.

Language and expression vary. Customers write in English, Nigerian Pidgin and a mix of both, with local shorthand, abbreviations and spelling variation. Build your test set from real messages, not invented ones, and make handling of Pidgin an explicit requirement if your customers use it.

Currency exposure is structural. Model usage is USD-priced and scales with success. A doubling of customer volume doubles that cost while your revenue may not follow immediately. Model the sensitivity before launch.

Data is often unstructured and scattered. Prices in WhatsApp threads, policies in someone's head, stock in a spreadsheet. The preparation phase is usually the real project.

Trust matters, and so does speed. Responses must arrive quickly on mobile data, and long generated replies cost the customer data. Nigerian customers are also alert to being fobbed off by a bot: an assistant that admits uncertainty and hands over quickly builds more confidence than one that answers everything with false authority.

Verify the vendor's actual experience. AI is a crowded label. Ask to see a working system they built, what its measured accuracy was, and what they would do differently. Questions to Ask an AI Development Company.

Example (hypothetical): an insurance broker outsources a support assistant

The following is a hypothetical illustration, not a Linestech client result.

A Lagos insurance broker handles a high volume of repetitive enquiries on WhatsApp: policy coverage questions, renewal dates, claim procedures and document requirements. Four staff spend much of the day on them.

Preparation (three weeks, internal). The operations lead consolidates product documents into one current version each and assembles 180 real enquiries from WhatsApp with correct answers written by a senior colleague. Half are kept private as a holdout set.

Pilot (three weeks, ₦900,000). The vendor builds an assistant over the consolidated documents. Against the public half of the test set it answers 82% correctly; against the private holdout, 74% — a gap that reveals how much tuning was fitted to the visible questions. Errors cluster around claims procedures, where the source documents contradict each other.

Decision. The broker fixes the claims documentation, narrows scope so that anything mentioning a claim routes straight to a human, and commissions the build.

Build (₦3,400,000). WhatsApp integration, a knowledge-base pipeline with a monthly update routine, confidence-based handover to staff, logging and a weekly review dashboard. Acceptance requires 88% correct on a fresh 200-question set with under 2% confidently wrong answers.

Running costs. Model usage plus messaging fees and hosting, with a hard spending cap and alerts, and a ₦250,000 monthly support agreement covering knowledge updates, evaluation runs and tuning.

The pilot cost ₦900,000 and changed the scope of a ₦3,400,000 build. That is what pilots are for.

Implementation: a nine-step process

1. Define the business problem and the measure of success — hours saved, response time, enquiries resolved without a human.
2. Assess data readiness using the checklist above, and fix conflicts before contacting vendors.
3. Build a test set of real questions or documents with correct answers, keeping a private holdout.
4. Shortlist vendors on systems they have actually built, with measured results and references.
5. Buy a paid pilot with a written accuracy report and recommendation as the deliverable.
6. Decide on the evidence: proceed, narrow the scope, or stop. Stopping after a ₦900,000 pilot is a good outcome, not a failure.
7. Contract the build with accuracy acceptance criteria, oversight rules, ownership of keys, prompts and knowledge base, and staged payments.
8. Launch with limits — a narrow scope, clear escalation to humans, spending caps and logging.
9. Operate and measure — monthly evaluation runs, sampled conversation reviews, knowledge updates and a quarterly cost review.

How to Manage an AI Implementation Projecthat Is the ROI of AI Integration? helps you judge whether it is paying for itself.

Mistakes to avoid

• Contracting AI like a website. Without a test set and a pass mark, neither side can say whether the work succeeded.
• Skipping data preparation. Conflicting source documents guarantee conflicting answers, and the vendor will be blamed for your filing.
• Letting the vendor hold the model provider account. You lose visibility of usage, cost and portability.
• Ignoring monthly usage cost. A successful assistant costs more than a neglected one, in dollars, every month.
• Automating high-consequence actions from day one. Start with drafting and suggestion; add autonomy once measured accuracy justifies it.
• Buying a large build without a pilot. Feasibility depends on your data, and nobody knows until they look at it.
• Forgetting the knowledge base ages. Prices, policies and products change; without an update routine the assistant becomes confidently wrong.
• Choosing a vendor on terminology. Ask for a working system, its measured accuracy and a reference you can call.

Conclusion

AI outsourcing succeeds when it is treated as a measurement problem rather than a build problem. Prepare your data, write a test set from real enquiries, buy a paid pilot, and only then commit to a build with an accuracy target the vendor accepts and your team judges.

Keep the model provider account, the prompts, the knowledge base and the evaluation set in your own hands. Budget for USD usage that grows with success. Design the handover to a human before you design anything else. Those decisions, made before signing, determine whether you end up with a system that saves real hours or a demonstration that quietly stops being used.

If you are scoping an AI project and want help defining the pilot, the test set or the acceptance criteria, Linestech can work through it with you before you commit to a full build.

Frequently asked questions

How much does it cost to outsource an AI project in Nigeria?

Indicatively ₦300,000–₦1,500,000 for a basic rule-based assistant, ₦1,000,000–₦5,000,000 for a language-model assistant with a business knowledge base, and ₦3,000,000–₦15,000,000+ for an AI agent that integrates with your systems. Add monthly USD-denominated model usage, hosting and a support retainer. These are 2026 indicative figures; obtain written quotes.

What is a realistic accuracy target for a business AI assistant?

It depends on the scope and the quality of your source material. For a narrow, well-documented domain, 85–95% correct answers with a low rate of confidently wrong responses is a reasonable target. Broad, poorly documented domains perform worse. Agree the target and the test set with your vendor rather than accepting a general promise.

Can we outsource AI if our data is disorganised?

You can start, but expect the first phase to be data work rather than AI work, and expect poor results until it is done. Consolidating documents, removing outdated versions and assigning owners typically takes two to four weeks of internal effort. Doing it before the vendor starts is cheaper than paying them to wait.

Who should pay the model provider directly?

Your business should hold the account and pay the provider, with the vendor granted access. That gives you visibility of actual usage, control over spending caps, and the ability to change vendor without changing accounts. If the vendor bills you instead, require the underlying usage report each month.

How do we stop an AI assistant giving wrong answers to customers?

You reduce the risk rather than eliminating it: narrow the scope, ground answers in approved documents, set a confidence threshold below which it hands over to a person, forbid it from discussing defined sensitive topics, require it to say when it does not know, and review sampled conversations weekly. Design the escalation route before launch.

What ongoing work does an outsourced AI system need?

Knowledge-base updates whenever prices, policies or products change; monthly evaluation runs against the test set; prompt and behaviour tuning based on real conversations; model version changes when the provider updates; usage and cost monitoring; and periodic review of what is being escalated to humans and why.

Should we use a ready-made AI tool instead of commissioning one?

Often yes, particularly for standard needs such as call summaries, drafting assistance or basic website chat, where subscription tools are cheaper and faster to adopt. Commission custom work when the value depends on your own documents, your own systems or workflows no off-the-shelf product supports. When Should You Add AI to Your Business?.

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.

• Nigeria Data Protection Commission (NDPC)
• Central Bank of Nigeria (CBN)
• WhatsApp Business Platform — Meta for Developers
• Cloudflare Learning Center
• OpenAI platform documentation
• Anthropic developer documentation