Use generative AI in business with a responsible method
A useful generative AI experiment begins with a bounded task and a measurable benefit. It also defines what information may be used, who checks the result and what happens when the system is wrong.
What to remember.
- Start with a reversible, low-risk task.
- Classify data before entering it into any service.
- Evaluate quality with real examples and human review.
Choose a specific use case
Select a repetitive task where assistance could save time without making an irreversible decision. Define the current process, expected benefit, unacceptable failure and the person accountable for the final result.
Set data boundaries
Identify personal, confidential, copyrighted and strategically sensitive information. Check contractual settings, retention, model-training options and access controls before any real data is used. Provide anonymised examples when possible.
Build a representative evaluation set
Test ordinary, difficult and adversarial examples. Score factual accuracy, completeness, tone, traceability and the time required for correction. A convincing demonstration is not evidence of reliable performance.
Keep human responsibility visible
Specify who reviews outputs and which decisions cannot be delegated. Give reviewers the original sources and enough time to challenge the result rather than simply approve fluent text.
Document prompts, versions and failures
Keep a lightweight record of the task, model or service, important settings, validation rules and known limits. Monitor cost, error patterns and changes to supplier terms as part of normal operations.
Expand only after evidence
Compare the experiment with the previous method. Scale when quality, risk and total effort are understood, not because the prototype felt impressive. Stop or redesign the use case when verification removes the expected benefit.
Check the reference material.
Related guides
Put the method into practice
An illustrative application case
A team wants to summarise enquiries with an external service. Before testing, describe transmitted information, possible errors and human review. Use synthetic examples and check whether work remains possible when the service is unavailable.
Document an API before integration
Use this worksheet in a review with the person responsible for delivery. Keep the evidence alongside the decision, rather than marking a task complete on trust alone.
| Action | Evidence |
|---|---|
| Choose a bounded task and a simple baseline for comparison. | A use case with expected outcome, acceptable errors and a go/no-go decision. |
| Check provider terms before using confidential documents. | A review of access, retention, reuse and data export. |
| Review outputs against a representative test set, including difficult cases. | A log of errors, corrections and human review time. |
Download the worksheet to fill in (CSV)
A question to resolve before acting
Does a confident answer mean a reliable answer?
No. Check names, dates, calculations and sources independently of tone. For a consequential decision, appoint a competent reviewer and allow the answer to be rejected.
Resources for this project
AI tool catalogue
World AI Guide
To build an initial tool shortlist, explore the catalogue by use case and access conditions. Then check data handling, export options and contractual terms with each publisher.
Cyber and AI specialist resource
Cyber Intelligence Embassy
For monitoring or sensitive decisions, review cyber exposure and strategic intelligence topics before deciding which information to entrust to a tool.
Sites from the AR ecosystem, selected for their relevance to this topic. Check scope, terms and current offers with each publisher. Our selection method.