AI for Business Decision Making in Nigeria: Where to Trust It and Where Not To

A Nigerian business owner makes dozens of consequential choices a week: whether to take an order on credit, which branch gets the last container of stock, whether to raise prices after a currency move, whether to replace a departing staff member, whether a customer complaint warrants a refund.

Most of those decisions are made from experience, under time pressure, with incomplete information. That is not a failure of discipline; it is what running a business in a volatile market looks like. The useful question about AI is narrow and practical: which of these decisions can be made better, and by what mechanism?

Reporting and dashboards are covered elsewhere in this library. This article is about the decisions themselves — how to classify them, where to delegate, where to keep a human firmly in charge, and how to tell afterwards whether the decisions actually improved.

Start with a decision inventory, not a tool

A decision inventory is a simple list of the recurring decisions your business makes, with four columns: how often it happens, who makes it, what information it currently uses, and what happens when it is wrong.

Building it takes an afternoon and immediately reshapes priorities. Most businesses discover two things. First, a handful of decisions repeat hundreds of times a month — approving a credit order, routing a delivery, responding to an enquiry, approving a discount — and each is made inconsistently by whoever is available. Second, the decisions that consume the most management attention are often rare and hard to improve with data at all.

Prioritise decisions that are frequent, consistent in structure, and measurably wrong some of the time. Those are where AI pays. A decision that happens twice a year and depends on relationships and judgement is not an AI project, however important it feels.

Three tiers of decision and how AI fits each

The commonest and most expensive error is misclassification — automating a decision that belongs in the recommend tier because the volume looked attractive. Reversibility is the test that matters most: if a wrong decision cannot be undone cheaply, a person approves it.

What AI is genuinely good at in decision support

• Consistency. A model applies the same criteria at 9am on Monday and 6pm on Friday. For decisions where the main failure is inconsistency between staff, this alone is the benefit.
• Reading everything. Two years of invoices, thousands of WhatsApp conversations, all supplier contracts, every delivery record. No manager can hold that; a properly connected system can query it.
• Probability over anecdote. People generalise from the last vivid case. A model weighs all cases, which corrects the natural tendency to over-react to the most recent disaster.
• Speed on structured questions. "Which twenty accounts have both a widening order gap and a recent complaint?" takes seconds rather than a day of spreadsheet work.
• Drafting and summarising. Preparing a board pack, summarising a long contract, extracting the terms from a tender document, turning a dataset into a written briefing.
• Surfacing what was not asked. Anomaly detection finds the branch with unusual write-offs or the supplier whose lead time quietly doubled.

Where AI should not be trusted alone

• Anything where the data is thin or biased. A model trained on past hiring decisions learns past preferences, including unfair ones. A credit model trained only on customers you already extended credit to cannot tell you about the ones you refused.
• Decisions about individuals with significant effects. Employment, credit refusal, access to services. These attract obligations under the Nigeria Data Protection Act 2023 regarding automated decision-making, and a human review route should exist. Confirm current requirements with the Nigeria Data Protection Commission or a qualified adviser.
• Novel situations. Models extrapolate from history. A policy change, a new competitor, a sudden fuel or FX move creates conditions no model has seen.
• Anything a language model asserts as fact without a source. Language models produce fluent, confident text that can be wrong. For decisions, require citations back to your own data or documents.
• Relationship and negotiation calls. In Nigerian B2B, supply continuity, long-standing credit history and personal trust carry weight that no transaction table records.
• Ethical and reputational judgements. A model optimises the objective it is given, which is rarely the whole objective.

Scenario planning in a volatile economy

Single-point forecasts are of limited use when the exchange rate, fuel price and policy environment can all move within a quarter. Scenarios are the better instrument, and AI makes building them cheap.

The method:

1. Identify the two or three variables that genuinely drive your economics — usually exchange rate, input or landing cost, and demand volume.
2. Define three coherent states: a base case, a stress case and an upside, each with specific values rather than adjectives.
3. Model the effect on the numbers that matter: gross margin, cash position by week, stock cover, break-even volume.
4. Identify decisions that differ between scenarios. If you would do the same thing in all three, the scenario work is complete and you can act now.
5. Set trigger points. "If the landed cost of the 20-litre line exceeds X, we raise price by Y or switch supplier." Deciding the trigger in advance removes panic from the moment.
6. Review monthly, and update the scenarios when a driver moves outside its assumed range.

AI's contribution is speed: it builds the model, runs the variants, writes the summary and flags which assumptions the result is most sensitive to. The judgement about which scenarios are plausible remains yours.

The decision log: how to know if decisions improved

Businesses that adopt AI for decisions rarely find out whether it helped, because nobody recorded what was decided or why. A decision log fixes that, and it costs almost nothing.

For each significant decision, record: the date, the decision, the options considered, what the AI recommended, what was actually decided, the reasoning, and the expected outcome with a review date. Then review.

Three things emerge within a quarter:

• Where overrides cluster. If staff override the model on a particular type of case every time, either the model is missing information or the policy is wrong. Both are fixable.
• Whether recommendations were better. Comparing outcomes of followed and overridden recommendations is the only honest evaluation available.
• Which decisions can move tier. A recommendation accepted 95% of the time over six months is a candidate for automation, with audit sampling.

Guardrails: keeping automated decisions safe and lawful

• Write the policy before the model. Credit limits, discount ceilings, refund thresholds and allocation rules should be documented decisions of the business, with the model applying them.
• Set hard limits the model cannot exceed. Maximum discount, maximum credit, maximum refund, excluded categories.
• Require a reason with every recommendation. If the system cannot explain the basis, it should not be used for decisions affecting people or money.
• Provide an appeal or review path for any decision affecting a customer, applicant or employee, and make it easy to find.
• Log every automated decision with its inputs and outputs, retained for audit.
• Sample and audit regularly. Review a random selection of automated decisions monthly against what a competent person would have decided.
• Monitor drift. Model performance decays as conditions change; schedule review and retraining rather than waiting for complaints.
• Keep personal data minimal. Feed the model only what the decision requires, and check your lawful basis under the NDPA 2023.

What changes for Nigerian businesses

Volatility shortens the useful horizon. Decisions built on annual assumptions age badly. Rebuild key assumptions quarterly, and hold pricing, purchasing and credit decisions on a shorter review cycle than you would in a stable market.

Data on the decision environment is thinner. Reliable external data on competitor prices, market size and category share is harder to obtain than in many markets. That raises the value of your own operational data and lowers the value of imported benchmarks.

Cash position drives more decisions than profit. Many decisions that look like margin decisions are really timing decisions — whether to take a discounted bulk purchase, whether to extend credit to a good customer, whether to prepay a supplier. Decision support that ignores the weekly cash position will give confident but unusable advice.

Informal information matters. Market intelligence arrives through sales reps, drivers, market associations and WhatsApp groups. Capturing structured notes from these sources gives models context that formal systems never see.

Regulatory decisions need professionals, not models. Tax positions, employment terminations, licensing and compliance questions should go to a qualified adviser, with reference to CAC, FIRS, NDPC, NITDA or the relevant sector regulator. Use AI to prepare the question and gather documents, not to answer it.

Staff trust has to be earned deliberately. A team that believes a model exists to catch them out will work around it. Introduce decision support as an assistant that saves them work and explain how overrides are used to improve it.

Example (hypothetical): a building-materials wholesaler deciding on credit

This is an illustrative scenario, not a Linestech client.

A wholesaler supplying contractors and retailers across Lagos and Ogun extends credit on about 60% of its orders. Credit decisions are made by the managing director, and on his travel days by whoever is senior in the office. Bad debt is a recurring drain, and good customers are sometimes refused while risky ones are approved because they called at the right moment.

The approach:

1. Decision inventory identifies credit approval as high volume, structured and measurably wrong some of the time — the strongest candidate in the business.
2. Policy first. The board documents credit tiers, maximum exposure per customer type, and the conditions for exceptions. This is the decision; the model only applies it.
3. Data assembled: three years of invoices and payment dates, disputes, order patterns, customer tenure, sector, and whether the customer's cheques or transfers ever failed.
4. Model built to estimate probability of payment within terms and expected days to pay, with explanations attached to every score.
5. Tier assignment: orders below a defined value with a strong score are auto-approved; anything above it, or with a weak score, is recommended with reasons for a human to approve; new customers above a threshold always go to a person.
6. Guardrails: hard exposure caps, automatic hold when total outstanding exceeds a limit, and a documented appeal route for refused customers.
7. Decision log records every recommendation, the decision taken, the override reason and the eventual payment outcome.
8. Monthly review compares outcomes, tightens the policy, and moves categories between tiers as evidence accumulates.

Within two quarters the business can answer a question it previously could not: which credit decisions were wrong, in which direction, and why.

How much does AI decision support cost in Nigeria?

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

Recurring costs separate from build: cloud hosting, language-model and analytics usage billed in US dollars, BI or database subscriptions per user per month, and internal time for audit and review. Exchange-rate movement affects the USD lines, so review quarterly.

Step-by-step: introducing AI into your decision process

1. Build the decision inventory with the management team in a single session.
2. Classify each decision as automate, recommend or inform, using volume, reversibility and consequence.
3. Pick one recommend-tier decision with good data and a clear cost of error.
4. Write the policy that governs it, in plain language, approved by whoever owns the risk.
5. Assemble the data, including the outcomes of past decisions — especially the bad ones.
6. Build the model to recommend with reasons, never to decide silently.
7. Run it alongside the current process for four to eight weeks without changing authority.
8. Start the decision log on day one of the pilot.
9. Review overrides and fix the model, the data or the policy accordingly.
10. Move to automation only where evidence supports it, with caps, sampling and an appeal route.
11. Expand to the next decision, reusing the data plumbing.
12. Review quarterly for drift, changed conditions and decisions that should move tier.

Mistakes to avoid

• Buying a dashboard and calling it decision support. A chart informs; it does not decide. Name the decision first.
• Automating an irreversible decision. Reversibility is the test, not volume.
• Letting a language model answer commercial questions from memory. Require it to cite your own data or documents.
• No policy behind the model. Without documented rules, the model becomes the policy by accident, and nobody can defend it.
• Hiding the reason. Staff and customers will not accept, and should not accept, unexplained decisions.
• No decision log. Without it, you cannot tell whether anything improved.
• Ignoring override patterns. Systematic overrides are the most valuable feedback you will receive.
• Using AI for legal, tax or regulatory conclusions. Prepare the question with AI; get the answer from a qualified professional.
• Rolling out to the whole business at once. One decision, well proven, teaches more than five half-implemented ones.

Conclusion

Better decisions come from classification before technology. List the decisions your business repeats, sort them by volume, reversibility and consequence, automate only what is frequent and reversible, keep a person on anything that affects an individual or cannot be undone, and use AI on high-stakes decisions to prepare evidence and scenarios rather than to conclude. Write the policy before the model, demand reasons with every recommendation, keep a decision log, and let override patterns drive improvement. Indicatively, a decision-support build in Nigeria runs from around ₦1,000,000 for a focused model to ₦10,000,000 or more for automated decisioning with full guardrails.

If decisions in your business depend on whoever happens to be available, Linestech can help you map your decision inventory, build models that recommend with reasons, and put the guardrails, logging and review in place before anything is automated.

Frequently asked questions

Can AI make decisions for a small business with little data?

Yes, but in a different mode. With limited data, the value comes from structure rather than prediction: documenting the decision rule, applying it consistently, and using a language model to assemble the relevant information and check that nothing was missed. Prediction models need history; consistency does not.

Is it safe to paste company figures into a public AI tool to get advice?

Not without care. Business data pasted into a consumer AI tool leaves your control, and customer or staff data carries obligations under the NDPA 2023. Use business-tier services with clear data-handling terms, remove identifying details, or deploy a private assistant connected to your own systems.

How do I stop AI recommendations from simply repeating past mistakes?

Train on outcomes, not on past decisions. A model that learns which credit approvals were repaid is useful; a model that learns to imitate which applications were approved just reproduces your existing bias. Where possible, include the cases you refused and what happened to them.

Should the AI recommendation be visible to staff, or only to managers?

Visible to whoever makes the decision, with the reasoning attached. Hidden scores create suspicion and prevent the override feedback that improves the system. Publish the policy and the reason codes so decisions can be discussed rather than merely accepted.

What decisions should never be automated in a Nigerian business?

Anything that significantly affects an individual without review — dismissal, credit refusal to a person, denial of a service — plus anything irreversible, anything with legal or regulatory consequence, and anything where the data is thin or known to be biased. Keep a documented human review path for all of these.

How long before decision support shows a return?

For a well-chosen, high-frequency decision, a pilot usually shows measurable differences within one to two quarters, because the decision repeats often enough to accumulate evidence. Rare decisions take far longer and are generally better served by scenario work than by models.

Do I need a data warehouse before doing any of this?

Not for a first project. A single decision usually needs data from one or two systems, which can be extracted on a schedule. Build the warehouse when three or four decision-support projects keep needing the same joined data, not as a prerequisite.

How do I handle disagreement between the model and an experienced manager?

Record both positions and the outcome. Over a quarter, the log will show where the manager's context beats the model and where the model's consistency beats intuition. That evidence, rather than seniority or novelty, should determine how much authority the system is given.

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)
• National Information Technology Development Agency (NITDA)
• OpenAI platform documentation
• Anthropic developer documentation