AI Implementation Checklist: Every Step From Scope to Live Operation
A checklist is not a substitute for thinking, but it is a reliable defence against forgetting. AI projects have more moving parts than a website or a spreadsheet tool: a model that behaves probabilistically, data that must be prepared, systems that must be connected, staff who must trust the output, a recurring cost in US dollars, and a regulator who cares what happens to customer data. Any one of these, left unchecked, can sink a project that was otherwise well built.
This article is the working checklist for implementing AI in a Nigerian business, organised by phase with a sign-off gate at the end of each. It assumes you have already decided AI is worth trying and have a candidate use case. If you have not assessed readiness yet, start with AI Readiness Checklist for Nigerian Businesses. For a narrative walk-through of the same journey with worked examples, see How to Implement AI in a Nigerian Business.
How to use this checklist
Work through the phases in order. Do not start a phase until the previous gate is signed off by the named project owner; the gates exist because AI projects tend to run ahead of their foundations. Copy the items into a shared document, assign a name and a date to each, and keep it as the single record of the project. For a small project (a chatbot answering from a price list) some items will take minutes; for a larger one (AI connected to your CRM and accounting) they may take weeks. Skipping items is allowed only with a written reason.
Phase 1: Scope and success criteria
The first step is to define exactly what the AI will do, for whom, and how you will know it worked. Ambiguity here is the most expensive kind.
- One-sentence problem statement agreed by leadership.
- The specific process or task the AI will handle, with its boundaries (what it does not do).
- The users of the AI output identified (customers, sales staff, accounts, management).
- The level of autonomy decided: AI suggests, AI drafts for approval, or AI acts alone within rules (see How AI Can Automate Nigerian Businesses).
- Two to four success measures with current baselines (for example, average response time, hours spent per week, error rate, conversion rate).
- A target for each measure and a review date (typically 90 days after launch).
- A named project owner with time allocated, and a named business sponsor.
- A one-off budget range and a recurring monthly budget in US dollars, with a ceiling.
- The list of things the AI must never do (quote unverified prices, give medical or legal advice, promise delivery dates, contact customers without consent).
Gate 1: the sponsor signs off the scope document. If you cannot write the problem in one sentence or cannot measure the baseline, you are not ready to proceed.
Phase 2: Data, access and compliance preparation
AI implementation in Nigerian businesses fails most often here, usually because the data the AI needs does not exist in a usable form.
- Inventory of the information the AI needs (products, prices, policies, FAQs, customer records, documents) and where each currently lives.
- Each source cleaned into a single, current, structured version with a named owner and an update routine.
- Personal data in scope identified and minimised; the AI gets only what it needs.
- Legal basis and NDPA 2023 obligations reviewed for the personal data involved; current NDPC guidance verified; privacy notice updated if customer-facing.
- Sector rules checked where relevant (for example, CBN expectations for financial services, confidentiality duties in healthcare and legal work).
- System access mapped: which systems the AI must read from or write to, who administers them, and whether APIs or exports exist.
- A test dataset assembled: 50 to 200 real, anonymised examples (enquiries, documents, transactions) with the correct expected outcome for each.
- Short AI policy in place or drafted (AI Policy for Nigerian Businesses).
Gate 2: the project owner confirms the data sources are ready and the test dataset exists. Do not brief vendors on a build that depends on data you have not yet organised.
Phase 3: Build-or-buy decision and vendor selection
- Options compared: an off-the-shelf tool, a configurable platform, or a custom build (Build vs Buy AI Software for Nigerian Businesses).
- Requirements written in business language, with the scope, data sources, integrations, autonomy level and "never do" list attached.
- Two or three written proposals obtained on identical scope from vendors or developers.
- Each proposal checked for: deliverables, timeline, one-off cost, recurring cost (and its currency), who owns the code and data, hosting location, support terms, training for your staff, and exit terms.
- References or comparable work reviewed; technical questions answered clearly in plain language (How to Choose an AI Company in Nigeria).
- Contract signed with milestones, acceptance criteria tied to your success measures, and a testing phase defined.
- Model and hosting choices agreed, including where data is processed and what the vendor's own AI providers do with it.
Gate 3: the sponsor approves the vendor and contract, including the recurring USD cost under a weaker-naira scenario.
Phase 4: Build and integration
- Kick-off meeting with the vendor: scope confirmed, contacts named, communication channel agreed (weekly check-ins at minimum).
- Access granted to systems using dedicated accounts with the least privilege needed, not shared passwords.
- Integration points built and tested one by one (WhatsApp, website, CRM, database, accounting, email) rather than all at once.
- Guardrails implemented in the system: the "never do" list, escalation to a human, rate limits, logging of every AI action.
- Tone, language and Nigerian context configured (naira formatting, local place names, appropriate register, handling of Pidgin or mixed-language messages where relevant).
- Admin view or dashboard delivered so your team can see what the AI is doing and correct it.
- Documentation started: how the system works, how to update its knowledge, how to pause it.
- Recurring costs monitored during development to confirm the estimate.
Gate 4: the vendor demonstrates the complete system against the scope document, on your data, in front of the project owner and at least one end user.
Phase 5: Testing with real cases
Testing an AI system is different from testing ordinary software because the output varies. Test with volume and with adversarial cases.
- The test dataset from Phase 2 run through the system; accuracy scored against expected outcomes and compared to the agreed acceptance threshold.
- Edge cases tested: incomplete questions, angry customers, requests outside scope, attempts to get the AI to break its rules, messages in Pidgin or with typos.
- Escalation tested: does a human actually receive the handover, and how quickly?
- Integration tested end to end: an action taken by the AI appears correctly in the CRM, sheet or inbox.
- Failure modes tested: what happens when the internet drops, the API provider is down, or the knowledge source is missing a field.
- Data-protection checks: no personal data appears where it should not; logs are secured; access is limited.
- End users (for example, two sales staff and one customer service officer) use the system for a week and record problems.
- All critical issues fixed and re-tested; a known-issues list agreed for non-critical ones.
Gate 5: the project owner signs the acceptance record, stating the measured accuracy and confirming the escalation path works.
Phase 6: Controlled launch
- Launch scope limited: one channel, one branch, one customer segment, or business hours only.
- Staff who will work alongside the AI trained, with a written guide and a named person to call (see AI Training for Nigerian Employees).
- Customer-facing wording agreed, including whether and how customers are told they are interacting with AI.
- Monitoring in place: daily review of conversations or outputs for the first two weeks, weekly thereafter.
- Rollback plan agreed: who can pause the system, how, and what the manual process is if it is paused.
- Success measures tracked from day one against the Phase 1 baselines.
- Vendor support arrangement active for the launch period, with response times agreed.
Gate 6: after two weeks of limited operation, the sponsor decides whether to widen the launch, hold, or roll back, based on measured results.
Phase 7: The first 90 days and ongoing operation
- Weekly review of accuracy, escalations, complaints and recurring cost.
- Knowledge sources updated on schedule (prices, stock, policies) by the named owner.
- Improvements prioritised from real usage rather than assumptions.
- 90-day review against the Phase 1 targets, reported honestly to leadership (AI ROI: How Nigerian Businesses Should Measure It).
- Recurring USD cost compared to estimate; ceiling adjusted or usage optimised if needed.
- Documentation finalised and stored where the business, not only the vendor, can find it.
- Maintenance and support agreement in place: model updates, security patches, provider changes.
- Next use case considered only after the first is stable.
Gate 7: the sponsor confirms the system is a normal part of operations with an owner, a budget and a maintenance plan.
What changes for Nigerian businesses
- WhatsApp is usually the channel. Implementation typically means the WhatsApp Business Platform (API) from Meta, which involves business verification, a phone number dedicated to the platform and per-conversation messaging charges in addition to model costs. Build this into Phases 3 and 4.
- Data preparation takes longer. Phase 2 often takes as long as the build itself because records are spread across chats, spreadsheets and paper. Budget the time honestly.
- Two currencies, one budget. Phase 1 sets a USD ceiling; Phases 4 and 7 check it. Exchange-rate movement can change the economics of a running system, so the ceiling should trigger a review rather than be discovered on a card statement.
- Payments and confirmations are manual. Many customers pay by bank transfer and send a screenshot. An AI that handles orders must be designed around transfer confirmation, POS receipts or gateway integrations (Paystack, Flutterwave, Monnify), not assume in-chat card payments.
- Power and connectivity shape testing. Phase 5 must include dropped connections and delayed messages, because they will happen in production.
- Compliance is not optional. NDPA 2023 applies broadly; sector regulators add requirements. Verify current obligations with the NDPC and your regulator rather than relying on a vendor's assurance.
- Trust is earned slowly. Nigerian customers may distrust automated replies. Phase 6 should include honest wording, an easy path to a human and close monitoring of sentiment in the first weeks.
Example (hypothetical): an Ibadan private hospital implements an AI appointment assistant
Example (hypothetical): a 40-bed private hospital in Ibadan receives hundreds of WhatsApp messages a week asking about doctor availability, opening hours, test prices and how to book. The front desk cannot keep up, and patients complain about unanswered messages.
- Phase 1: the AI will answer availability, hours and listed test prices, and book appointments into the hospital's scheduling system for general consultations only. It will never give medical advice, discuss results or handle emergencies; those hand over to a person immediately. Success measures: response time, front-desk hours on WhatsApp, booking errors. Recurring budget: a USD ceiling for messaging and model usage.
- Phase 2: a single sheet of services, prices, hours and doctor schedules is created and owned by the admin lead. A test set of 120 anonymised past messages is assembled. The hospital reviews its patient-data obligations under the NDPA 2023 and updates its privacy notice; personal health information is excluded from the assistant's scope.
- Phase 3: three proposals compared; the chosen vendor commits to integration with the scheduling system, an admin dashboard, staff training and clear ownership of the hospital's data.
- Phase 4: WhatsApp Business Platform onboarding, integration with scheduling, guardrails for the "never do" list, and Yoruba and Pidgin-aware greeting handling.
- Phase 5: the test set scores above the agreed threshold; emergency-sounding messages are confirmed to escalate within seconds; a week of front-desk trial catches two scheduling edge cases.
- Phase 6: launch for general consultations only, weekday hours, with the front desk reviewing every conversation daily for two weeks.
- Phase 7: weekly reviews, schedule sheet updated each Monday, 90-day report to management, and a decision about extending to specialist bookings.
The details are illustrative, not a Linestech client result. How to Build an AI Appointment Assistant.
Indicative timeline and cost by phase
The table gives indicative 2026 timelines and cost shares for a medium-complexity project such as a WhatsApp assistant connected to one business system. Actual figures vary with scope, vendor and exchange rate; total one-off costs for such projects typically fall within the ranges in How Much Does AI Integration Cost in Nigeria?.
| Phase | Typical duration | Share of one-off cost | Notes |
|---|---|---|---|
| 1. Scope | 1–2 weeks | Internal time; optional facilitation ₦100,000–₦500,000 | Cheapest phase to get right, most expensive to skip |
| 2. Data and compliance | 2–6 weeks | Internal time; data clean-up help ₦100,000–₦1,000,000 | Often the longest phase for SMEs |
| 3. Vendor selection | 2–3 weeks | Internal time | Compare 2–3 written proposals on identical scope |
| 4. Build and integration | 3–10 weeks | 50–70% of build cost | Depends on integrations and channel |
| 5. Testing | 1–3 weeks | 10–20% of build cost | Do not compress this to hit a launch date |
| 6. Launch | 2 weeks limited | 5–10% of build cost plus training | Recurring USD costs start here |
| 7. First 90 days | 3 months | Support retainer; recurring model and messaging usage | Budget 15–25% of build cost per year for maintenance |
Mistakes to avoid
- Starting Phase 4 before Gate 2. Developers build against data that does not exist yet; the project stalls while you scramble to clean records.
- Accepting a demo as a test. A vendor demonstration on ten friendly questions is not Phase 5. Insist on your test set and adversarial cases.
- No autonomy decision. If nobody decides whether the AI acts alone or drafts for approval, the system either does too much or too little and staff lose trust.
- Launching everywhere at once. A wide launch multiplies the impact of every early mistake. Limit the first two weeks.
- Forgetting the USD ceiling. Recurring costs that were fine at contract time can rise with usage and currency movements; make the ceiling a monitored number.
- Leaving documentation and credentials with the vendor. The business must be able to pause, update and, if necessary, move the system.
- Treating go-live as the end. The first 90 days are part of implementation. Without weekly reviews, accuracy drifts as prices, stock and policies change.
Conclusion
AI implementation succeeds when the unglamorous phases are done properly: a scope that fits in one sentence with measurable baselines, data that is cleaned before anyone builds, a vendor chosen on identical written scope, testing on your own real cases, a limited launch, and 90 days of honest review. The gates in this checklist slow a project down only where speed would cost more later. Use it as a shared document with names and dates, and the project stays under the business's control from the first meeting to normal operation.
If you are ready to move from checklist to build, Linestech implements AI chatbots, assistants and integrations for Nigerian businesses with scoping, testing on your data and staff training built into the project.
Frequently asked questions
How long does a typical AI implementation take for a Nigerian SME?
For a medium-complexity project such as a WhatsApp assistant connected to one business system, expect roughly three to five months from scope to the end of the first 90 days, with two to four of those weeks spent on data preparation and one to three on testing. Simple assistants answering from a document can be live within four to six weeks; multi-system integrations take longer.
Can we run this checklist with an internal team instead of a vendor?
Yes, if you have builder skills in-house (see AI Skills Nigerian Businesses Need). Phase 3 then becomes a tool-selection exercise rather than vendor selection, but every other phase applies unchanged. Internal teams are more likely to skip Phase 5 and Gate 6, so keep the sign-offs formal.
What accuracy should we accept before launch?
Set it in Phase 1 based on the risk of the task. A chatbot answering opening hours can tolerate occasional misunderstandings if it escalates well; a system extracting invoice amounts needs to be near-perfect or reviewed by a person. Define the threshold with your vendor before the build and measure it on your own test set, not the vendor's.
Who should own the AI system after launch?
An operations or department lead, not IT alone and never the vendor alone. The owner maintains the knowledge sources, reviews outputs, tracks cost and coordinates with the vendor for changes. Name this person in Phase 1 and give them time; ownerless AI systems decay within months.
Do we need to tell customers they are talking to AI?
Being transparent is good practice and supports trust with Nigerian customers who are wary of automated messages; some sector rules may also require it. Decide the wording in Phase 6, make a human easy to reach, and verify any disclosure obligations that apply to your sector with the relevant regulator.
What should be in the contract with an AI vendor?
Deliverables tied to your scope document, milestones and acceptance criteria based on your test set, one-off and recurring costs with their currencies, ownership of code, prompts and data, hosting and data-processing locations, staff training, documentation, support response times, and exit terms that let you take the system elsewhere. Get two or three proposals on the same scope before signing.
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.
