Using Customer Data to Build AI Products: The Legal Questions Businesses Need to Answer
Every African business with a meaningful customer base is sitting on data that looks like a competitive advantage. A digital lender in Lagos has years of repayment histories. A healthtech platform in Nairobi has millions of consultation records. A logistics company in Accra has route, delivery and customer behaviour data across several cities. A payroll provider in Johannesburg holds salary data for thousands of employers.
The commercial logic for using that data to build AI products is compelling: better credit models, smarter triage, more accurate demand forecasting and new products to sell to existing customers. Investors are asking about it, and competitors are doing it.
The legal logic is more demanding. Data collected for one purpose does not automatically become available for another, and the fact that a business holds data does not mean it is free to use it to train models. This article sets out the questions a business should answer before using customer data to build or improve AI products.
The commercial trigger: a new use for old data
The problem almost always starts with data collected for a specific purpose, such as processing a loan, delivering a parcel or providing a medical consultation, and a later decision to use it for something else: training a model, building an analytics product or improving a service sold to other customers.
That shift in purpose is the point at which legal risk crystallises. Before it, the business is using data for the reason it was collected. After it, the business needs a fresh justification.
Question 1: Is this personal data?
The first question is whether the data identifies, or could identify, an individual. If it does, data protection law applies. Under the Nigeria Data Protection Act 2023, personal data is broadly defined, and so is processing, which includes using data to train a model.
Businesses often assume that removing names makes data anonymous. It usually does not. Transaction histories, location data, device identifiers and combinations of demographic attributes can frequently be linked back to individuals. Data that has been pseudonymised, meaning direct identifiers have been replaced with codes but the business can still re-identify individuals, remains personal data.
True anonymisation, where re-identification is not reasonably possible, takes the data outside data protection law. But anonymisation is harder to achieve than most businesses assume, particularly for rich behavioural data, and a business that relies on it should be able to demonstrate how it was done.
Question 2: Is training compatible with the original purpose?
Nigerian data protection law, like most modern data protection laws including Kenya’s Data Protection Act 2019 and South Africa’s Protection of Personal Information Act, requires that personal data be collected for specified, explicit and legitimate purposes and not further processed in a way incompatible with those purposes.
Whether model training is compatible with the original purpose depends on the facts. Relevant considerations include the link between the original purpose and the new one, the context in which the data was collected and what customers would reasonably expect, the nature of the data, particularly whether it is sensitive, the consequences for individuals, and the safeguards in place, such as pseudonymisation and access controls.
Using loan repayment data to improve the same lender’s credit model for the same product may be closer to compatible. Using that data to build a scoring product sold to other lenders is a different matter. Using health consultation data to train a general-purpose model is further still.
Question 3: What is the lawful basis?
If the new use is not covered by the original purpose, the business needs a lawful basis for it. The realistic options are usually consent or legitimate interests.
Consent must be freely given, specific, informed and unambiguous, and must be as easy to withdraw as to give. Consent buried in terms and conditions, or bundled as a condition of receiving a service, is unlikely to meet the standard. Withdrawal creates a further practical problem: once data has been used to train a model, removing its influence is technically difficult.
Legitimate interests require a balancing exercise: the business must show that its interest in the processing is not overridden by the rights and interests of the individuals concerned. This is more flexible, but it requires documented assessment and is harder to justify for sensitive data or uses that customers would not expect.
Where the data includes sensitive categories, such as health data, financial data in some contexts, biometric data or data about children, the bar is higher and explicit consent or another specific condition will often be needed.
Question 4: Is the business a controller or a processor?
Business-to-business companies face a distinct issue. A software company providing services to corporate customers usually processes its customers’ end-user data as a processor, acting on the customer’s instructions. A processor cannot use that data for its own purposes, including training its own models, without the customer’s authorisation.
Many B2B contracts in African markets were signed before AI was on anyone’s mind and say nothing about this. Others contain data processing terms that expressly prohibit use for the processor’s own purposes. A business that trains models on customer data in these circumstances may be in breach of contract, and may also have become a controller in respect of that processing, with the obligations that come with that status.
The fix is contractual: customer agreements should state clearly whether the provider may use customer data, or aggregated and de-identified data derived from it, to improve its services and build new products, on what conditions, and with what opt-outs.
Question 5: Where is the training happening?
Most AI training happens on cloud infrastructure outside Africa. Transferring personal data out of Nigeria for training is a cross-border transfer subject to the Nigeria Data Protection Act’s rules, which require adequate protection in the destination country or appropriate safeguards. The same applies under Kenyan, South African, Ghanaian and other regimes, some of which are stricter. Egypt requires authorisation for cross-border transfers in many cases, and Rwanda has localisation requirements for certain data.
A business planning to train models offshore should map those transfers and put appropriate safeguards in place before data moves.
Question 6: What do sector rules say?
Sector regulation adds another layer. Banks and other financial institutions are subject to secrecy and confidentiality obligations that are independent of data protection law. Healthcare providers are subject to medical confidentiality rules and, in many jurisdictions, specific health data regulations. Telecoms operators face their own rules on subscriber data. A data protection analysis that ignores sector rules is incomplete.
Question 7: Who owns the model and what happens if customers leave?
Commercial contracts should address ownership of the model and outputs, and what happens to data and model influence when a customer terminates. Customers increasingly ask for commitments that their data will not be used to benefit competitors, or that it will be deleted on exit. A business should know what it can promise before it signs.
The enforcement context
African data protection regulators are becoming more active. Nigerian regulators have taken enforcement action against major organisations in recent years, including significant penalties against a global technology platform following a joint investigation involving the Federal Competition and Consumer Protection Commission, and fines against a commercial bank for data protection failures. Kenya’s Office of the Data Protection Commissioner has issued enforcement notices and penalties against a range of businesses, and South Africa’s Information Regulator has been increasingly assertive. Using customer data to build AI products is precisely the kind of high-visibility processing likely to attract scrutiny.
A practical framework
Before using customer data to build AI products, a business should:
- Map the data: what it is, where it came from, what customers were told and where it sits.
- Assess whether it is personal data, and whether anonymisation is realistic.
- Assess compatibility with the original purpose and identify a lawful basis for the new use.
- Carry out a data protection impact assessment, which will usually be required for this kind of processing.
- Review customer contracts, particularly B2B data processing terms, and amend them where needed.
- Address cross-border transfers and sector-specific rules.
- Update privacy notices and, where relevant, give customers a meaningful opt-out.
- Build governance around the model itself: documentation of training data, testing for bias and accuracy, and a process for handling withdrawal of consent and deletion requests.
Add a Comment
Your email address will not be published.We'd love to help you!
Let us know the needs of your business, and we will pinpoint the best-suited solution to fulfill them.
