An AI procurement memo for a business that cannot afford a demo

A vendor demonstrates an AI system on prepared data, under good connectivity, with staff who know which prompts work. The buyer sees a fast answer. Procurement begins with a less attractive screen: export, termination and failure.

GSMA's African AI work identifies local-language data, compute, skills, connectivity and policy as practical constraints. Its 2026 mobile report places AI inside a market where 63% of people live within mobile-broadband coverage but do not use mobile internet. A buyer cannot assume the demo environment represents employees or customers.

Exit is a practical test.

Write the exit paragraph first

List every asset and record needed to change supplier or return to manual operation: source data, cleaned data, prompts, configurations, evaluation sets, logs, output history, workflow rules and documentation. State the format and transfer time.

Identify what the vendor owns. Fine-tuning, embeddings or annotations created from buyer data may have restricted portability. A contractual export right has little value if the format cannot be used elsewhere.

Set deletion and certification obligations after termination, including backups and subprocessors. Legal requirements can compel retention; the contract should explain exceptions.

Estimate switching work and service interruption. A low subscription with a six-month rebuild is not month-to-month flexibility.

Name the data and permission

Create a data-flow diagram from collection to model, output, logs and support. Mark personal, confidential and regulated information. Identify countries and providers involved.

The vendor should disclose whether buyer inputs or outputs train shared models, by default and under optional settings. “We do not sell data” does not answer training, retention or human review.

Confirm the buyer has rights to provide the data. Customer records, employee communications and licensed documents can carry restrictions. A procurement contract cannot grant rights the buyer lacks.

Minimise before transfer. Remove fields the use case does not need. Synthetic or masked data can support testing, subject to whether it preserves the relevant failure modes.

Test the difficult slice

Build an evaluation set from real permitted work. Include local languages, code-switching, older devices, weak audio, incomplete records and cases that should be escalated. Keep a final set away from vendor tuning.

Define the correct result and consequence of error. A draft marketing summary can tolerate different wording. A wrong payment instruction cannot. Use separate thresholds.

Ask the vendor to run the set without hand-editing. Record model and product version, settings, response time and cost. A result with vendor engineers in the room is a supported test, not normal operation.

Measure refusals and unsupported confidence. The system should expose uncertainty or route to a person where policy requires it.

Price human work

Identify who reviews outputs, handles exceptions, corrects records and answers complaints. Estimate minutes and volume from the pilot. Add training and management.

If employees must copy information between systems, the integration is unfinished and the labour remains. If reviewers accept nearly every answer without checking, the control may be ceremonial.

Changes to roles need consultation, training and employment-law review. A business case cannot book full payroll savings while keeping the same staff and assigning them new work without valuing that output.

Vendor support also has a human boundary. Which incidents receive a person, in what time zone and under which severity definition?

Put connectivity in acceptance

Measure application load, data per task, offline behaviour and retries. Test on customer or employee networks. A system that loses a form after a dropped session adds cost and distrust.

Cloud availability is one component. The local network, identity provider, integration and vendor can each fail. Error messages should identify the action users can take without exposing security details.

A mobile channel must fit device memory and operating-system versions. Voice channels need call cost and speech evaluation. Do not approve “mobile support” from a recent flagship phone test.

Set an availability and recovery requirement around the business process, then price redundant options only if consequence justifies them.

Inspect the bill under success

Pricing can grow with users, tokens, documents, calls, storage or compute. Model expected interaction mix and a high-usage month. Add implementation, support, data movement, taxes and currency.

A model that becomes more useful can become more expensive. Set budget alerts, rate limits and a process for approving new use. Ask how model or pricing changes are communicated.

Minimum commitments transfer adoption risk to the buyer. Discounts should be compared with a slower ramp and the cost of unused capacity.

Benchmark total cost per completed business task, not per model request. Failed requests and human correction still belong to the task.

Verify the company behind the product

Review the contracting entity, security evidence, incident history available under diligence, insurance, financial position and subprocessors. Certifications cover a defined scope and date; read the report and exceptions.

Ask which model providers the vendor depends on and whether it can switch them without buyer approval. A reseller can inherit an upstream outage and price change.

Contractual limits of liability should be compared with plausible harm and insurance. Procurement may accept a limit, but the board should see the gap.

References need comparable use and market conditions. A testimonial from another sector does not validate local language or regulation.

Return to the use case

Only now write the problem in one sentence, current process, baseline cost and desired outcome. If the product does not beat that baseline on the evaluation set after human work and operating cost, the demo has not earned procurement.

Approve a bounded deployment with owner, users, data, thresholds, budget, review date and stop condition. Expansion requires fresh evidence.

A business that cannot afford a failed demo can still test AI. It has to buy the exit, evidence and operating controls before it buys the performance story.

Sources

• GSMA, July 2024: AI for Africa: Use cases delivering impact
• GSMA, 16 June 2026: The Mobile Economy Africa 2026