Most organisations do not have an AI problem. They have a work problem: too many manual steps, information scattered across systems, and decisions that wait for someone to find the right file. AI can help with all of that, but only when the starting point is the work itself.
1. List the work, not the tools
Ask each team which tasks are repetitive, slow or error-prone. Capture volume (how often), effort (how long) and pain (what goes wrong). You will usually find candidates in document handling, service requests, reporting and data entry.
2. Score each candidate
Rate value (time or cost saved, risk reduced), feasibility (are the inputs digital and reasonably consistent?) and risk (what happens if the output is wrong?). The best first project is valuable, feasible and low-risk, not the most impressive.
3. Name an owner and a baseline
Every pilot needs a business owner who will use the result and a measurement taken before anything changes: average handling time, error rate, backlog. Without a baseline, success becomes a matter of opinion.
4. Pilot with people in the loop
Run the first version alongside the existing process. Let the system suggest and a person decide. Track how often people accept, correct or reject its output. Those numbers tell you whether to trust it with more.
5. Decide deliberately
After a fixed period, compare results with the baseline. Scale what works, fix what nearly works and stop what does not. Stopping is a valid outcome; it protects budget and credibility for the next idea.
The foundations matter as much as the model. AI depends on accessible data, secure access and systems that can talk to each other. If those are weak, the first project may be to strengthen them.
Start small, measure honestly and keep people responsible for important decisions. That is how AI moves from an experiment to part of the way an organisation works.