The short answer
It answers one question: is this worth building, and what will it cost to run? Choose one contained use case with a named owner, write thresholds down first, test on your own data, safely, against an evaluation set of real examples, and keep it to a few weeks. End with results per threshold, a go or no-go and a production cost estimate, not "everyone liked the demo".
Choose a use case that can lead somewhere
- Frequent, so savings add up
- Measurable today, for a baseline
- Easy for a person to check
- Backed by data you have
- Owned by someone who decides and runs it
- Connected to known systems: a path to production
Typical: answering staff questions from approved policies, extracting supplier invoice fields, drafting replies for staff approval, summarising case history. Leave decisions about individuals, payments and anything that must be right without review for later. Our AI readiness checklist helps you shortlist.
Define success before you start
| Measure | What to set | How to measure |
|---|---|---|
| Accuracy | Share that must be right; errors that fail automatically | Scored against the evaluation set |
| Time saved | Minutes per item versus today | Time today's process first |
| Cost per transaction | Maximum per question, document or task | Actual charges in testing |
| User acceptance | Would pilot users choose it? | Feedback; drafts accepted with light edits |
| Safety | No answers from content the user cannot open; resists prompt injection | Permission and security tests |
Agree what "go", "no-go" and "go with changes" mean, and have the business owner sign off. Thresholds set after the results are not thresholds.
Example thresholds
A hypothetical HR policy assistant. The evaluation set holds 120 real staff questions, 20 with no answer in the policies. Go ahead only if at least 90% of answerable questions are answered correctly with a citation, at least 18 of the 20 unanswerable ones get "I could not find that", no answer draws on a document the user cannot open, and the measured cost per question fits the budget finance approved. Your thresholds depend on the task and the cost of an error.
Use your own data, safely
Demo data hides the hard cases, so test on your own records, carefully:
Minimise
Anonymised, synthetic or de-identified data where personal data is not needed, prompts included.
Enterprise services
Terms that exclude your content from training; deployment type, region and transfer basis chosen deliberately.
Limited access
Project team only, in an approved environment, with deletion planned for the end.
Or fully private
An open-weight model on a temporary GPU server in your own cloud account.
Early data protection review
The NDPA applies to personal data used in AI systems; plan a DPIA early.
Data minimisation is a principle of the Nigeria Data Protection Act 2023 (NDPA), and NITDA has advised against putting personal, classified or confidential data into public AI tools. The NDPC's GAID 2025 makes a data protection impact assessment mandatory for, among other things, sensitive data, new technologies and transfers outside Nigeria, which most hosted AI services involve: there is no major public cloud region in Nigeria, and Microsoft's "Global" deployments may process prompts in any geography where the model is deployed (Microsoft Learn). See AI and the NDPA; for data that cannot leave even for a test, private AI and self-hosted models. General information, not legal advice.
Build an evaluation set
It is the most valuable thing a proof of concept produces: you reuse it every time the model, prompts or sources change.
- Collect real examplesQuestions staff actually ask, documents you actually receive.
- Expert answersSubject-matter experts write the expected answer or correct fields.
- Awkward casesNo answer, conflicts, poor scans, off-topic, restricted content, hidden instructions.
- Hold some backFor the final test, so you are not tuning to seen questions.
- Score consistentlyPeople, automated evaluators or both.
Microsoft's guidance: evaluate with your own datasets before deployment, then monitor quality in production (Microsoft Learn). Document-answering assistants: see retrieval-augmented generation.
Keep it small and time-boxed
A few weeks, one use case, a small user group, weekly results. Skip the full interface and every source; do sketch the production integration and measure running costs (Microsoft's cost guidance: test a small, representative workload, compare charges with your estimate). Clearly short of thresholds halfway? Stop: change approach or drop it.
What you should receive at the end
- Results per threshold, with failure examples
- Go, no-go or go with changes, and why
- Production estimate in naira (or US dollars where a licence or service is priced in dollars): one-time build apart from usage, hosting, licences, support (what an AI project costs)
- Risks, controls and human review points for production
- Integration outline: systems, access, owner
- What was not tested
Why many pilots never reach production
| Cause | How to avoid it |
|---|---|
| No owner | A business owner who decides and will run it |
| Demo data | Your own data, including edge cases |
| No integration plan | Systems, APIs and permissions identified during the proof of concept |
| No budget for running costs | Cost per transaction measured and presented with the build estimate |
| Success never defined | Thresholds written down before starting |
Rather talk it through? If you have an AI idea and want evidence before committing a budget, we can scope a time-boxed proof of concept on your own data, with thresholds agreed up front. Talk to a Promatics specialist
How Promatics runs a proof of concept
- Agree in writingUse case, owner, thresholds, data handling, timebox, estimate in naira (or US dollars where priced in dollars).
- Prepare safelyDe-identified approved sample, enterprise environment, evaluation set with your experts.
- Build and testPrototype, weekly tuning, held-back final run, permission and prompt injection tests.
- ReportResults against thresholds, recommendation, production design, cost estimate, risks.
- Your decisionIf the answer is no, we say so.
If it is yes, we build it through generative AI development or AI agents, and support it after launch.
When to bring in help
Copilot or a built-in AI feature on non-sensitive content with a few users: trial it yourself. Bring in help when the test needs personal or confidential data, production would connect to your ERP or CRM, leadership needs reliable evidence, or nobody in-house has built an evaluation set.
Plan your proof of concept
- One use case and a named business owner.
- Today's baseline measured: volume, time per item, error rate.
- Signed-off thresholds: accuracy, time saved, cost per transaction, user acceptance.
- Automatic-failure conditions listed.
- Data sample approved and de-identified where possible.
- Enterprise service, region, retention and transfer basis confirmed.
- Evaluation set with edge cases and a held-back portion.
- Timebox and weekly check-ins agreed.
- Production systems and integration path identified.
- Decision meeting booked for the end.
Sources and further reading
Product capabilities and guidance change. These are the primary sources this article relies on, checked on the review date above.
- Nigeria Data Protection Act 2023, Nigeria Data Protection Commission
- Data, privacy, and security for Foundry Models sold by Azure in Microsoft Foundry, Microsoft Learn
- Observability in generative AI, Microsoft Learn
- Plan and manage costs for Microsoft Foundry, Microsoft Learn
This article is general information, not legal, accounting or security advice for your specific situation. Examples are hypothetical unless stated otherwise.