Back to Blog
Nov 17, 2026Startup guide

AI Procurement: What Companies Should Negotiate Before Signing With an AI Vendor

The AI vendor market in Africa has moved quickly from experimentation to enterprise deployment. Banks are signing with providers of AI-driven fraud detection and customer service. Insurers are licensing claims automation. Telecom operators are deploying AI for network optimisation and churn prediction. Professional services firms, including law firms, are rolling out generative AI assistants across their teams. Many mid-sized companies are simply upgrading their existing software and discovering that it now includes AI features.

Most of these contracts are signed on the vendor’s standard terms, often through click-through acceptance or online order forms. That is understandable for small purchases, but for AI tools that will touch sensitive data, inform material decisions or sit inside customer-facing processes, the standard terms rarely allocate risk in a way that works for the buyer.

AI contracts raise questions that ordinary software contracts do not. The vendor may want to use the buyer’s data to improve its models. The outputs may be inaccurate in ways that are hard to predict. The underlying model may change without notice. Intellectual property in the outputs may be uncertain. This article sets out what companies should negotiate before signing.

Start with the use case

The terms that matter depend on what the tool will do. A writing assistant used internally for first drafts carries different risks from a model that approves or declines loan applications, or a chatbot that speaks to customers on the company’s behalf. Before negotiating, the buyer should define the use case, the data involved, the decisions the tool will inform and the consequences if it gets things wrong. That analysis determines where to push.

1. Use of your data for training

This is the first question to settle. Many AI vendors reserve the right to use customer inputs and outputs to train and improve their models. For enterprise customers, that is often unacceptable: inputs may include confidential business information, client data, personal data or trade secrets.

The buyer should negotiate a clear commitment that its inputs and outputs will not be used to train or improve the vendor’s models without express written consent, and that any use for service improvement is limited to aggregated, de-identified data. The commitment should cover the vendor’s subcontractors and the providers of any underlying foundation models the vendor uses.

2. Confidentiality and data protection

The contract should include robust confidentiality obligations that extend to inputs, outputs and any information the vendor learns about the buyer’s business.

Where personal data is involved, the vendor will usually act as a processor, and the contract must include data processing terms that satisfy the Nigeria Data Protection Act 2023 and the Nigeria Data Protection Commission’s General Application and Implementation Directive, as well as the laws of any other jurisdictions where the buyer operates. Those terms should cover processing only on the buyer’s instructions, security measures, sub-processor approval, assistance with data subject requests and impact assessments, breach notification within defined timeframes, and deletion or return of data on termination.

3. Data location and cross-border transfers

Most AI services process data outside Africa. The buyer should know where its data will be processed and stored, including by sub-processors and foundation model providers, and should ensure that cross-border transfers comply with applicable law. For regulated entities, data location may also be subject to sector rules, and some public sector and sovereign data may need to remain in Nigeria under NITDA guidelines. Where local or regional hosting is available, it is worth negotiating for.

4. Ownership of outputs

The buyer should own, or at least have unrestricted rights to use, the outputs generated for it. Most enterprise AI terms now assign outputs to the customer, but some reserve rights for the vendor, and the position on similar outputs generated for other customers needs to be understood.

The buyer should also recognise a legal uncertainty: copyright protection for AI-generated material is unsettled in many jurisdictions, because copyright has traditionally depended on human authorship. A contractual assignment of outputs does not create rights that the law does not recognise. For outputs that matter commercially, human contribution and review strengthen the buyer’s position.

5. Intellectual property indemnity

AI models are trained on vast amounts of material, some of which may be subject to third-party rights, and outputs may occasionally reproduce protected material. Several major vendors now offer indemnities against third-party intellectual property claims arising from outputs, usually subject to conditions such as using the service as intended and not disabling safety filters.

The buyer should seek an indemnity covering claims that the service or its outputs infringe third-party rights, understand the conditions attached, and check that the indemnity is not undermined by a low liability cap.

6. Accuracy, testing and human oversight

AI vendors invariably disclaim accuracy. Their terms state that outputs may be incorrect and that the customer is responsible for evaluating them. For low-risk uses this may be acceptable. For uses where errors have real consequences, the buyer should negotiate:

  • Performance commitments tied to agreed metrics, where the vendor is prepared to give them.
  • Rights to test the tool before deployment and on an ongoing basis.
  • Information about the model’s known limitations, testing for bias and appropriate use cases.
  • Cooperation with the buyer’s own validation and audit processes.

Internally, the buyer should design processes that keep a human accountable for material decisions, both because it reduces risk and because data protection law in Nigeria, Kenya and South Africa gives individuals rights in relation to solely automated decisions with significant effects.

7. Model changes and continuity

AI services change constantly. Vendors update models, retire older versions and alter behaviour, sometimes in ways that affect the buyer’s processes. A workflow tuned to one model version may perform differently after an update.

The buyer should negotiate advance notice of material changes, the ability to remain on a version for a defined period, deprecation timelines, and a right to terminate without penalty if a change materially degrades the service.

8. Security and incident response

The contract should require security measures appropriate to the data, including protection against risks specific to AI systems, such as prompt injection and data leakage between customers. It should require prompt notification of security incidents, cooperation with the buyer’s response, and access to independent security certifications or audit reports.

9. Liability and remedies

Standard AI terms typically cap the vendor’s liability at fees paid over a short period, often twelve months or less, and exclude indirect and consequential losses, data loss and lost profits. For tools that process sensitive data or support material decisions, the buyer should push for higher caps, or uncapped liability, for breaches of confidentiality and data protection obligations, the intellectual property indemnity, and gross negligence or wilful misconduct.

10. Regulatory cooperation

Regulated buyers need contractual rights that allow them to meet their regulators’ expectations. Banks and payment service providers in Nigeria should consider Central Bank of Nigeria expectations on outsourcing and technology risk, which may require audit and access rights for the institution and the regulator, business continuity commitments and notification of material changes. Similar considerations apply to insurers, pension operators and capital market operators.

11. Commercial terms that bite in African markets

Two commercial issues deserve particular attention in Nigeria and other markets with foreign exchange constraints.

First, pricing for AI services is often usage-based, measured in tokens, queries or seats, and priced in US dollars. Usage can grow quickly once a tool is adopted, and so can costs. The buyer should negotiate usage caps, alerts and pricing protections, and should consider currency and payment mechanics carefully.

Second, payments to foreign vendors for software and technical services may engage National Office for Technology Acquisition and Promotion (NOTAP) registration requirements, withholding tax and VAT. The contract should allocate these costs and set realistic timelines for any registration.

12. Exit

Finally, the buyer should plan for the end of the relationship: return or deletion of data, certification of deletion, export of any fine-tuned models, prompts or configurations the buyer has paid to develop, and reasonable transition assistance.

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.

Services


Copyright 2026 Vazi Legal Business Attorneys. All Rights Reserved. | Privacy Policy

LinkedInXInstagram