Planning for an AI solution? After all everybody does, right? Whether aiming for AI-first or embedded, be sure first that AI is necessary to solve your problem and is feasible. If yes, start small and iterate often.
There are dozens of reasons why most of AI projects (95%) underdeliver or fail. Spend sufficient time with the following questions; you may end up with dropping AI solution altogether, scaling down the AI part, aiming at incremental but real benefits and/or significantly accelerate time to ROI.
1. Solving the right problem with the right tool?
Is there a clear, measurable goal that AI can help you to achieve?
Who needs to be satisfied and how will success be perceived?
Can you achieve the goal with deterministic, automated solution or does the problem require probabilistic, cognitive one?
Have you explored rule-based, statistical, or human-driven solutions? Why are they insufficient?
Which of the seven AI patterns does this project align with?
Can the project be split into smaller iterations aligned with a distinct patterns, deliverables, or level of augmentation?
Do project dimensions and constraints allow for experimentation and extensive data related activities?
2. Is AI solution operationally viable?
Have you done evaluation of data availability, data V’s and model fit?
Are you AI-first or AI-embedded? Can AI transform or be embedded into your existing workflows, tools, or environments?
Is your target operationalization environment (cloud, edge, hybrid) compatible with the AI architecture?
Are there latency, privacy, explainability, or compliance requirements that constrain the solution?
Do you have access to necessary AI, data science, and engineering expertise?
Is your infrastructure capable of supporting model training, inference, and monitoring?
Can you shortcut using pre-trained models, synthetic data, or prompt engineering?
3. Building something responsible, measurable, and resilient?
What metrics (accuracy, precision, ROI, user satisfaction) define a successful outcome?
How will you track solution performance, quantitatively and qualitatively?
If the project fails, what are the consequences, financial, reputational, strategic? Can you pivot?
If it succeeds, how will it scale, evolve, or influence future initiatives?
Can you break the project into smaller, faster experiments to validate assumptions early?
What level of model performance is acceptable? What trade-offs are tolerable? Which measures are critical?
What could go wrong (data drift, bias, model overfitting, misclassification), and how will you mitigate?
Are you meeting standards and regulations for fairness, privacy, transparency, accountability, and human oversight?
AI solution is not a software development project, not even close. Consider the data-centric-and-Agile approach that will maximize the probability of success.
– Greg