Questions to Ask an AI Development Company Before You Commit

AI projects fail differently from ordinary software projects. Software either works or throws an error. An AI system produces a confident answer that may be wrong, costs money every time it runs, depends on data quality you may not have assessed, and behaves differently in month six than it did in the demonstration.

That means the due diligence questions are different too. A firm can be an excellent web developer and still have no framework for evaluating whether a language model is answering your customers correctly.

These questions are written for Nigerian businesses buying an AI chatbot, an AI agent, a document-processing system or an AI feature inside existing software. Ask the ones that apply, in writing, and compare answers across your shortlist.

Start with the business question, not the technology

Before speaking to any vendor, write down three things. Without them, every proposal you receive will be a demonstration rather than a plan.

1. The task you want handled, described as work a person currently does. "Answering the same forty delivery questions on WhatsApp" is a task. "AI for customer service" is not.
2. The current baseline. How many of those enquiries arrive daily, how long each takes, what it costs in staff time, and how often the answer is wrong today. Without a baseline you cannot tell whether the system helped.
3. What a wrong answer costs. A wrong delivery estimate is a complaint. A wrong medical or financial answer is a serious matter. This single judgement determines how much accuracy work the project needs and therefore most of the budget.

A capable AI firm will ask for all three in the first conversation. If they do not, they are selling a technology rather than solving a problem.

Questions about the use case and success measurement

The cheaper-alternative question is the most revealing one you can ask. A great many business problems described as AI problems are solved better by a decision tree, a form, a search function or a properly configured autoresponder — faster, cheaper and with no running cost per message. A firm that volunteers this is thinking about your outcome.

Questions about your data

Most AI disappointments in Nigerian businesses are data problems wearing a technology costume.

1. What data does this system need to work, and do we have it? Product information, policies, past conversations, pricing, records.
2. What condition is our data in? Ask them to assess a sample before quoting. Contradictory or out-of-date source material produces confidently wrong answers.
3. Who prepares the data, and is that in the quote? Cleaning, structuring and writing knowledge base content is often the largest hidden task in the project.
4. Will our data be used to train anyone's model? You generally want a clear no, with the contractual position written down.
5. Where will our data be processed and stored? Which country, which provider, and what that means under the Nigeria Data Protection Act 2023.
6. What happens to data we send during testing? Test data is real customer data surprisingly often.
7. How do we update the knowledge the system uses? Can your staff do it, or does every price change require a developer?
8. What happens when our information changes? Stale knowledge is the most common cause of AI systems quietly becoming unhelpful.

The practical test: ask the vendor to take one real policy document from your business and show what the system would answer from it. Fifteen minutes of that is worth more than an hour of slides.

Questions about models and lock-in

1. Which model or provider will you use, and why that one? Expect reasoning about cost, accuracy on your task, language handling and latency.
2. How hard is it to switch models later? Systems built with a separation between application and model are far easier to move. Ask directly how it is structured.
3. Does the system handle Nigerian English, Pidgin and local name spellings? Customers write how they speak. Ask to see it tested on real messages from your business.
4. Will it work over WhatsApp, and through which route? The WhatsApp Business Platform has its own approval process and message costs, which belong in the budget.
5. Are you reselling another company's product or building this? Both are legitimate, but you should know which, and what the underlying licence costs.
6. What happens if the model provider changes prices or retires a model? This happens regularly. Ask what the migration would involve and who pays for it.
7. Is any part of this dependent on a service that could disappear? Smaller AI tooling companies change direction frequently.
8. Can any of this run without an internet connection? For field operations in Nigeria the answer is usually no, which affects where the system is useful.

Questions about accuracy and being wrong

This is the section most buyers skip and most regret skipping.

On hallucination. Language models can produce fluent, incorrect answers. The mitigation is engineering, not optimism: restrict the system to answering from your approved documents, instruct it to decline rather than guess, show sources where appropriate, and route anything uncertain to a person. A vendor who does not raise this unprompted has not run an AI system in production.

Insist on a human escalation route in every customer-facing deployment. In Nigerian customer service, the fastest way to lose a customer is an automated loop with no exit.

Questions about running costs in dollars

AI is the first category of business software where the bill grows with usage in a currency you do not earn. Treat this as a central commercial question.

1. What is the estimated monthly model or API cost at our expected volume? Ask for a figure in US dollars with the assumptions stated.
2. What happens to that cost if volume doubles? Ask for the cost at half, current and double volume.
3. Which costs are fixed and which vary with usage? Hosting, vector storage, message fees, model calls and monitoring each behave differently.
4. Who pays the provider — us directly or you? Accounts in your own name give you visibility and control; billing through a vendor adds a margin you should know about.
5. What controls exist to cap spending? Usage limits, alerts and rate limiting should be built in from day one.
6. How will exchange-rate movement affect us? Naira costs for a dollar-denominated service change with the rate, and your budget should reflect a range.
7. Can cost be reduced with a smaller model or caching? Competent firms have several levers and should explain them.
8. What is the total first-year cost including build, running costs and support? Ask for one number with the assumptions listed.

A useful framing: compare the monthly running cost against the staff cost of the work being handled. If an AI assistant costs the equivalent of a junior staff member's salary and handles a third of the enquiries, the case is weak. If it costs a fraction of that and handles the repetitive majority, the case is strong. Insist on doing this arithmetic before signing.

Questions about integration and operations

1. What systems will this connect to? WhatsApp, your website, CRM, order system, inventory, email. Each connection is work.
2. Will it read data or also change it? A system that can issue refunds or update orders needs far stricter controls than one that only answers questions.
3. What permissions will it have, and who approves actions? For anything financial, require human approval steps.
4. How fast will responses be? Customers abandon slow chats. Ask for expected response times under load.
5. What happens when a connected system is down? The AI should degrade gracefully, not fail silently or invent an answer.
6. How will we know if it stops working? Monitoring and alerting, with a named recipient.
7. Who maintains it, and what does maintenance include? Prompt updates, knowledge refreshes, model changes, monitoring and reporting.
8. How often will it need attention? AI systems drift as your business, your data and the underlying models change. Quarterly review is a reasonable minimum.

Questions about privacy, security and NDPA compliance

1. What personal data will the system process? Names, phone numbers, addresses, order history, health or financial details.
2. How does this comply with the Nigeria Data Protection Act 2023? Obligations sit with your business as the data controller. Ask what they will implement, and verify requirements with the Nigeria Data Protection Commission.
3. Where is the data processed, and is it transferred outside Nigeria? Cross-border processing needs to be understood and documented.
4. What are customers told? A clear notice that they are interacting with an automated assistant, and how to reach a person.
5. How long are conversations retained, and who can read them? Retention periods and access controls should be explicit.
6. How is the system protected against prompt manipulation? Users will try to make a public assistant say things it should not. Ask what guardrails exist.
7. Can it be made to reveal internal information? Ask how the system is restricted to the intended scope.
8. What happens in a data breach? An incident response process, notification obligations and who does what.

For regulated sectors — financial services, healthcare, education — ask specifically what sector requirements the firm has worked with, and confirm current obligations with the relevant regulator rather than relying on the vendor's reading.

Questions about ownership and exit

• Who owns the prompts, configuration and knowledge base we paid to build? It should be you, in writing.
• Who owns any fine-tuned model or embeddings created from our data? Establish this before the work starts.
• Are the provider accounts in our name? Model provider, hosting, messaging platform and monitoring.
• What do we receive if we end the relationship? Source code, prompts, documentation, data exports, credentials.
• Can another firm take this over? Ask what documentation makes that possible and request a sample.
• Is anything built on a proprietary platform only they can operate? Disclosure now, not discovery later.
• Is our data deleted on exit, and how is that evidenced?

What AI work costs in Nigeria

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

Recurring costs to budget for: model or API usage, messaging fees where WhatsApp is involved, hosting, monitoring, and periodic prompt and knowledge maintenance. Most of these are US dollar-denominated.

Example (hypothetical): a logistics company buying a support agent

This is an illustrative scenario, not a Linestech client.

A Lagos logistics company receives a high volume of daily WhatsApp enquiries, most of them asking where a package is, what delivery costs to a given area, and how to book a pickup. Three staff answer them. Two vendors are approached.

The written questions produce very different answers:

• Baseline. Vendor A does not ask. Vendor B asks for a week of message exports, categorises them, and reports that a large share are tracking questions that could be answered by connecting to the existing tracking system without any language model at all.
• Architecture. Vendor B proposes a hybrid: a direct system lookup for tracking and pricing, with a language model only for unstructured questions. This cuts the expected monthly model cost substantially.
• Accuracy. Vendor A promises "highly accurate responses". Vendor B proposes building a test set of 200 real questions with correct answers, measuring performance before launch, and defining an escalation rule for anything unmatched.
• Running cost. Vendor A does not provide one. Vendor B gives estimated monthly USD figures at half, current and double volume, plus WhatsApp message costs, and recommends spending caps with alerts.
• Data protection. Vendor B raises the Nigeria Data Protection Act 2023 unprompted, proposes a retention period and an automated notice that customers are speaking to an assistant with a route to a person.
• Ownership. Vendor B confirms in writing that the prompts, knowledge base and accounts belong to the company.

Vendor B's quote is higher. It is also the only one that answers the question the business actually asked.

Answers that should stop the conversation

• "It will handle everything." No AI system does. A vendor who will not define the boundary has not built one in production.
• No estimate of monthly running costs. This is the cost that continues forever. It cannot be an afterthought.
• "We will use your data to improve our product." Unless you have explicitly agreed it, this should be a no.
• No accuracy testing method. Without a test set, "accurate" is an opinion.
• No human escalation route. A customer-facing assistant with no exit will damage your service reputation.
• Silence on the Nigeria Data Protection Act 2023. Any firm handling customer data in Nigeria should raise it unprompted.
• Refusal to run a small paid pilot. Pilots protect both sides; reluctance suggests confidence the demonstration will not survive real data.
• Guaranteed percentage improvements. Nobody can promise a specific business outcome before seeing your data.
• Evasiveness about who owns prompts and configuration. A one-sentence answer, or a problem.

Conclusion

The questions that matter when buying AI are not about the technology. They are about your baseline, your data, what happens when the system is wrong, what it will cost every month in dollars, and who owns what when the relationship ends. A firm that answers those specifically, and is willing to tell you that a simpler tool would do the job, is worth more than one with a polished demonstration.

Start with a narrow pilot, measure against a real baseline, insist on accuracy testing and a human escalation route, and get running-cost estimates in writing at three volume levels before you commit.

If you are evaluating AI proposals and want one built on a measured baseline, tested accuracy and transparent monthly running costs, Linestech works with Nigerian businesses on AI chatbots, agents and integrations from pilot through to ongoing support.

Frequently asked questions

How do I know whether we need AI at all?

Look at whether the work is repetitive, language-based and high volume, and whether your information is written down somewhere reliable. If the task is rule-based and predictable, conventional automation is cheaper and more dependable. If it involves interpreting varied human language at volume, AI earns its place.

Should we run a pilot first?

Almost always. A pilot with a narrow scope, a fixed period and a defined decision point lets both sides learn on a small budget. Handling one category of enquiry well for six weeks tells you more than any proposal, and the cost is a fraction of a full deployment.

What is a realistic accuracy expectation?

It depends entirely on the task and your data quality, which is why any figure quoted before seeing your material is meaningless. Insist that accuracy is measured on a test set built from your own real questions, and that the escalation rule catches what the system cannot answer.

Who is responsible if the AI gives a customer wrong information?

Commercially and legally, the business deploying it generally carries the responsibility, which is why escalation rules, grounding in approved documents and monitoring matter. Your contract should address vendor obligations, and for regulated sectors take qualified legal advice.

Can we use AI without sending data outside Nigeria?

Options exist, including self-hosted models, but they carry higher infrastructure cost and often lower capability. Most Nigerian deployments use international providers, which makes documenting cross-border processing and reviewing obligations under the NDPA 2023 important.

How much should the first AI project cost?

Start small enough that failure is affordable. For most Nigerian SMEs a first project in the ₦1,000,000–₦5,000,000 range with a narrow scope is sensible, with running costs understood before launch. Large first projects rarely survive contact with real data.

How often will the system need maintenance?

Expect quarterly attention at minimum: refreshing the knowledge base, reviewing sampled conversations, updating prompts and responding to model changes from the provider. Budget for it from the start rather than discovering it in month four.

What if our data is disorganised?

Then that is the first project. Cleaning, structuring and writing an accurate knowledge base often accounts for a substantial share of effort, and doing it improves your business whether or not the AI proceeds. A vendor who ignores data condition is quoting for a system that will underperform.

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)
• WhatsApp Business Platform — Meta for Developers
• OpenAI platform documentation
• Anthropic developer documentation