In brief
What should you ask before starting an AI project?
What decision or task will improve and how will we measure it? Do we have the right data? What happens when the AI is wrong? How will the output fit into real work? Who will own and maintain the system after launch?
AI discussions often start with technology: which model, which platform, which vendor. Those choices matter, but they are easier and better made once a few fundamental questions have clear answers. Five questions cover most of what leaders need to know before committing budget.
1. What decision or task will improve, and how will we know?
State the specific task, such as reading supplier invoices, prioritising maintenance jobs or answering employee policy questions. Then define how success will be measured against today: time per case, error rate, response time, forecast accuracy or satisfaction. If success cannot be described in measurable terms, the project is not yet ready.
2. Do we have the right data?
AI learns from or retrieves from data. Check that the data exists for this task, that you are allowed to use it, that its quality is sufficient, and that it reflects the conditions where the AI will operate. Look at samples early; descriptions of data are nearly always more optimistic than the data itself.
3. What happens when the AI is wrong?
Every AI system makes mistakes. The question is what those mistakes cost and who bears them. A wrong product recommendation is minor. A wrong eligibility decision or a missed safety defect is not. The answer determines how much human review is needed, how much testing is required and whether the use case is appropriate at all.
| Consequence of error | Implication |
|---|---|
| Low and easily corrected | Automation with sampling and monitoring may be acceptable |
| Moderate | Human review of uncertain cases and regular audits |
| High or affecting people's rights | Human decision with AI as support, plus rigorous validation and governance |
4. How will the output fit into real work?
An accurate model that produces results nobody sees creates no value. Identify where outputs will appear, which system or screen, who acts on them and what changes in their day. Integration with existing systems and the design of the user experience frequently take more effort than building the model.
5. Who will own it after launch?
AI systems need care after release. Data changes, models drift, knowledge sources age and users find new edge cases. Name a business owner accountable for outcomes and a technical owner responsible for monitoring, updates and incidents. Budget for that ongoing work from the start.
Using the answers
Clear answers to all five questions suggest a use case ready for design or a pilot. Uncertain answers on data or feasibility point to a feasibility study or proof of concept. Uncomfortable answers on consequences may lead to redesigning the human role, narrowing the scope or choosing a different approach. In every case, the questions turn an AI idea into a plan.
Sources and further reading
About the author: Dr. RohitCo-Founder, Director and Chief Technology Officer, AARHIT Systems. Writes on technology strategy, AI architecture, agents, evaluation and AI governance. More articles by Dr. Rohit