How to Measure AI ROI: A Guide for Enterprises in African Markets

AI return on investment (AI ROI) is the financial value an AI system produces in production, net of everything it costs to build, run and supervise. Most enterprises can now point to AI in use. Far fewer can point to AI that pays. In McKinsey's August 2026 State of AI survey, nearly nine in ten organisations reported regular AI use, while 6 percent attributed 5 percent or more of EBIT to it.

The distance between those two figures is decided early. It is set by which use cases an enterprise chooses to build, and whether each one can reach production under the conditions it will operate in. For enterprises operating in African markets, those conditions include data spread across legacy systems, national data protection law, infrastructure the business must control, and customers who transact in local languages.

This article sets out how AdwumaTech measures AI ROI before a build begins: what makes it hard to forecast, the two axes every candidate use case is scored on, and how that score becomes a decision.

In brief

• AI ROI is decided by which use cases an enterprise builds and whether each can reach production.
• Score every candidate on value (business impact, unit economics) and on readiness (data, infrastructure, governance, language).
• A readiness factor scored 1 gates the build, whatever the value.
• Unit economics—the full cost per completed task after review—is the most reliable predictor of payback.

Why is AI ROI difficult to forecast?

AI ROI is difficult to forecast because the costs that decide it sit outside the model. A pilot measures capability on curated data. Production adds the conditions that determine return: integration, supervision, regulation and scale.

Pilot results do not transfer to production

A pilot runs on a clean sample, with engineers watching every output. In production the same system meets incomplete records, unfamiliar inputs and real customers. Accuracy measured in the pilot overstates accuracy in operation, and the forecast built on it overstates the return.

The cost of correction is carried in headcount

Every wrong output that reaches a reviewer costs time. That cost rarely appears under an AI budget line. It sits in the operations team, and it grows with volume. On AdwumaTech's internal Akan and Ewe evaluation set, a general-purpose frontier model returned a 38.2 percent task error rate, equivalent to roughly 74 hours of qualified human review per 1,000 outputs. A model priced low per token can still be the most expensive option once review is counted.

Integration determines the timeline

A system that returns a recommendation to one screen an operator already uses can be live in weeks. A system that writes to a core banking platform, an ERP or a national register carries months of integration, testing and approval. Two use cases with identical value can differ by months in time to return.

Task savings are not enterprise savings

A system that halves the time spent on a task saves money only when that time is redeployed or removed. Measured at the task level, the gain looks immediate. Measured at the enterprise level, it depends on workflow and staffing decisions the model does not make.

How should enterprises score AI use cases?

Every candidate use case is scored on two axes: the value it would create, and the enterprise's readiness to run it. Value without readiness produces a stalled build. Readiness without value produces a working system nobody needs. AdwumaTech's AI Readiness Assessment scores six factors on a 1-to-5 scale.

The two axes are scored separately and never averaged together. A high value score cannot compensate for an infrastructure gap, and a high readiness score cannot justify a use case with no return.

What are AI unit economics?

AI unit economics is the full cost of one completed task, measured after quality review, compared with the cost of the same task today. It is the most reliable single predictor of whether a use case will pay back, because it captures the costs a pilot hides.

Cost per completed task = (inference + infrastructure + amortised integration + human review) ÷ tasks completed to standard

Human review is usually the largest term. It equals the error rate on the enterprise's own inputs, multiplied by volume, multiplied by the time a qualified reviewer takes to correct one output. A model with a lower price per token and a higher error rate often costs more per completed task.

Three practices keep the figure honest:

• Measure error on production-representative data. Public benchmarks describe other populations and other languages.
• Count only tasks completed to standard. An output that needs rework belongs in the cost line.
• Re-measure after launch. Inputs drift, volumes change and review time falls as confidence thresholds are tuned. Unit economics is a live figure, reviewed quarterly.

A worked example

The table compares two models on the same customer service task, against a manual cost of USD 2.00 per task. Review is priced at 6 minutes per correction and USD 12 per reviewer hour. Figures are illustrative.

The adapted model costs three times more per call and 44 percent less per completed task. Error rate, measured on the enterprise's own inputs, decides the result.

What changes when AI is deployed in African markets?

The readiness factors are universal. The form they take in African markets is specific, and it moves the ROI calculation in ways a global playbook does not anticipate.

Data. Operational records in banks, telecoms and public institutions often span core systems, scanned documents and call-centre logs. Consolidation and labelling are frequently the first workstream, and they belong in the cost line before any model is chosen.

Infrastructure. Regional cloud and data-centre capacity is limited, and many regulated enterprises require systems to run on private cloud or on-premise hardware they control. Deployment architecture is therefore a scoring factor from the first conversation.

Governance. National data protection statutes set the obligations for personal data: Ghana's Data Protection Act, 2012 (Act 843), Nigeria's Data Protection Act 2023, Kenya's Data Protection Act 2019 and South Africa's Protection of Personal Information Act. Sector supervisors add security review before production. Both shape the timeline and the architecture.

How these obligations shape architecture is covered in AdwumaTech's companion article on sovereign AI deployment: running enterprise AI inside your own perimeter.

Language. Customers transact in Twi, Hausa, Swahili and hundreds of other languages that general-purpose models handle poorly. Error rates on these inputs feed directly into the review cost term of unit economics, so language performance is measured on the enterprise's own data before a use case is ranked.

How does a score become a decision?

Each use case leaves the Assessment with one of three outcomes, set by where it lands on value and readiness.

1. Build now. High value, and every readiness factor at 3 or above. The use case enters engineering.
2. Remediate first. High value, with at least one readiness factor at 1. The engagement opens with the data, infrastructure or governance work that closes the gap, then moves to build.
3. Deferred. Value too low to prioritise at current cost. Revisited when volumes, prices or the process change.

A single readiness score of 1 gates the build, whatever the value. This rule matters most for the most valuable use cases, because they tend to touch the systems of record that carry the deepest integration and the strictest review. A credit decisioning model can score 10 out of 10 on value and still be the wrong first build if it must write to a core banking platform the enterprise cannot yet connect to safely.

The output is a ranked register of every candidate, the score and evidence behind each factor, and a sequenced roadmap that shows which remediation unlocks which build. The engineers who produce it then deliver the highest-value use case that is ready for production, held to the success criteria the register defines.

Frequently asked questions

What is the formula for AI ROI?

AI ROI equals the financial gain an AI system produces minus its total cost, divided by its total cost, expressed as a percentage. Total cost includes inference, infrastructure, integration, governance and human review, measured over the system's operating life.

What is an AI readiness assessment?

An AI readiness assessment is a structured review that scores an enterprise's candidate AI use cases on business value and on the data, infrastructure, governance and language conditions each one requires. It returns a ranked list and a decision for each use case before engineering budget is committed.

Why do AI pilots fail to reach production?

Pilots are measured on curated data with close supervision. Production adds integration with core systems, regulatory review and real input at volume. Use cases that were never scored on those conditions stall at the point they first meet them.

How long does an AI Readiness Assessment take?

AdwumaTech's AI Readiness Assessment is completed in ten business days by engineers embedded with the client team.

Does data have to leave the country?

No. AdwumaTech deploys systems on the client's private cloud or on-premise hardware, so processing and records remain within the client's environment.

About AdwumaTech

AdwumaTech is an AI services firm headquartered in Accra. Engineers and product managers embed inside client enterprises to decide what to build, then build and deploy it inside the client's security perimeter. AdwumaTech is ISO/IEC 27001 certified and aligned to ISO/IEC 42001 and GDPR. Its open-source speech translation dataset, mGhana-ST, was named Best African Dataset at Deep Learning Indaba 2026, and it holds research MOUs with the University of Ghana and Valley View University.

Request an AI Readiness Assessment at adwumatech.ai.

Sources

• McKinsey, The State of AI (August 2026)
• AdwumaTech internal evaluation set, Akan and Ewe task outputs