You know you should use AI. Here's where to start.
Most businesses feel pressure to 'do something with AI' without a clear idea of what. That's how companies end up paying for tools nobody uses or chatbots that annoy customers. The good news: finding a worthwhile first AI project is a fairly simple exercise.
By Usama bin Mazhar · 5 min read · Updated
Start from a problem, not from AI
Don't ask 'where can we use AI?'. Ask 'what work do we repeat all day that a smart assistant could do?'. List the tasks that take the most staff time each week. AI is a candidate for some of them; plain automation is the better answer for many others.
The four-question test
A good first AI project usually passes all four:
- Frequency — does the task happen many times a day or week?
- Messy input — does it involve reading free text, emails, documents or conversations? (If the input is already structured, normal automation is cheaper and more reliable.)
- Available knowledge — does the information needed to do it already exist in your documents or systems?
- Safe failure — if the AI gets it wrong, can a person catch it before it causes real harm?
Common first projects that work
- Answering repeat customer questions from your own policies and data
- Replying to new leads instantly and booking them in with a salesperson
- Reading invoices, forms or applications and filling in your system
- Helping staff find answers in internal documents and procedures
- Drafting routine replies, summaries and reports for a person to check
Decide the hand-off before you build
Every AI system should know exactly when to stop and pass work to a person. Define it up front: which questions it may answer, which actions it may take, and what happens when it isn't sure. Customers trust AI more when a human is clearly one step away.
Measure one thing
Pick one measure before you start — response time, hours saved, documents processed, leads booked — and record today's number. Without a 'before', you can't tell whether the project worked.
Questions worth answering before you ask them.
Almost never for a first project. Existing models, connected to your own data with the right controls, cover most business needs.
It depends on the provider and setup. Use business-grade services that don't train on your data, limit what the AI can access, and check the data protection rules in your country before sending personal data to any AI service.