Most organisations now feel pressure to "do something with AI". The risk is starting with a broad ambition, such as an AI strategy for the whole company, that produces slides rather than results. A better approach is one narrow project that solves a real problem, proves value within months and teaches the organisation what it needs for the next one.
Pick a problem, not a technology
Good first projects share three traits: the task is repetitive and time-consuming, the data needed already exists in your systems, and success can be measured. Examples include answering common customer questions, matching payments to invoices, extracting data from documents and forecasting demand for a product range.
Check the data before the model
- Is the data available, and do you have the right to use it for this purpose?
- Is it accurate and complete enough? AI amplifies data problems rather than fixing them.
- Does it contain personal data? If so, the Data Protection Act, 2019 applies, including purpose limitation and security safeguards.
Keep a person in the loop
For a first project, design AI to assist people rather than replace their judgement. A customer-service assistant can draft answers from your approved knowledge base and hand complex cases to an agent. A finance assistant can propose payment matches for a person to approve. This keeps quality high while the organisation builds confidence.
Measure from day one
Agree the baseline before you start: how long the task takes today, how many errors occur and what it costs. Measure the same things after launch. A clear before-and-after comparison is what turns a pilot into a funded programme.
Build responsibly
- Ground answers in your own verified content to reduce made-up responses.
- Log interactions so issues can be reviewed and the system improved.
- Restrict access to sensitive systems and data to what the use case needs.
- Tell customers and staff when they are interacting with an AI system.
