AI GuidePractical Applications of Generative AI in Business
Beyond the hype: the everyday business tasks where generative AI genuinely helps, and how to introduce it without creating new risks.
6 min read
AI Guides
Practical Guides for Putting AI to Work
Concise, actionable guides for leaders and teams planning AI initiatives, covering readiness, technique selection, evaluation and responsible use.
Guide 1
AI readiness checklist
Use this checklist to assess whether a proposed AI initiative has the foundations it needs. Gaps are not reasons to stop; they show what to address first.
Strategy and use case
- The task to improve is specific and described in plain language
- A business owner is accountable for the outcome
- Success measures and a current baseline are agreed
- The cost of an AI mistake is understood and acceptable with review
Data
- The data or documents the AI needs are identified and accessible
- Data quality and freshness are known, including gaps
- Sensitive and personal data are classified
- Representative examples exist for testing
Technology and integration
- It is clear where AI output will appear in existing tools
- Integration options (APIs, exports) for source and target systems are known
- Hosting and data-residency requirements are defined
- Monitoring and logging responsibilities are assigned
People and governance
- Users have been involved in defining the problem
- Review responsibilities for AI output are defined
- An AI usage policy exists or is being drafted
- Training and change communication are planned
Guide 2
RAG or fine-tuning? Choosing how to customise a language model
Retrieval-augmented generation and fine-tuning solve different problems. Many teams reach for fine-tuning when retrieval would be simpler, cheaper and easier to keep current.
| Consideration | Retrieval-augmented generation (RAG) | Fine-tuning |
|---|---|---|
| Primary purpose | Give the model access to specific, current information | Teach the model a style, format or specialised behaviour |
| Keeping knowledge current | Update documents and re-index; no retraining | Requires retraining when knowledge changes |
| Citations and traceability | Natural: answers can link to source passages | Difficult: knowledge is absorbed into model weights |
| Data needed | Well-organised source documents | Many high-quality example input-output pairs |
| Access control | Can filter retrieved content per user | Cannot easily restrict what the model has learned |
| Typical use | Policy assistants, product Q&A, research support | Consistent tone, structured outputs, domain phrasing at scale |
Our advice
- Start with good prompting and retrieval; measure quality on real questions
- Consider fine-tuning only when retrieval and prompting cannot reach the quality bar, or when volume makes a smaller tuned model more economical
- The two techniques can be combined: a tuned model for style, with retrieval for facts
Guide 3
Evaluating an AI feature before launch
Evaluation turns opinions about AI quality into evidence. A simple, repeatable process is far better than ad-hoc testing.
- 1
Collect real examples
Gather representative inputs from the actual process, including difficult and unusual cases.
- 2
Define good output
For each example, record the expected answer or the criteria a reviewer would use.
- 3
Choose measures
Accuracy of extracted fields, factual correctness, citation quality, tone, refusal behaviour when information is missing.
- 4
Run and score
Score every prompt, model or retrieval change against the full set, combining automated checks with expert review.
- 5
Test safety
Include attempts to misuse the system, such as prompt injection or requests outside scope, and confirm safe behaviour.
- 6
Set launch criteria
Agree the minimum scores required to launch and the monitoring that will continue in production.
Guide 4
Writing an AI usage policy for your team
A short, clear policy helps employees use AI tools productively while protecting the organisation. It should enable, not just restrict.
Approved tools
List which AI tools are approved for work use and how to request new ones.
Data rules
Specify which categories of information may and may not be entered into AI tools.
Human responsibility
Make clear that people remain responsible for work they produce with AI assistance.
Disclosure
State when AI use should be disclosed to customers or colleagues.
Quality checks
Require verification of facts, figures and citations produced by AI.
Reporting
Explain how to report problems, errors or concerns about AI behaviour.
Go deeper
Related reading
AI Guide
AutomationUsing AI to Improve Business Workflow Efficiency
A step-by-step approach to finding where AI removes friction from everyday workflows, and where simple rules work better.
4 min read
Ready to apply these guides?
We can facilitate an AI readiness workshop with your team and help prioritise use cases.
