Most AI projects that stall were pointed at the wrong work. The technology was fine; the task was a poor fit. Before choosing a tool, run the task itself through five questions. They work the same for a claims department and a two-person shop.
1. Does it happen often?
AI earns its keep on volume. A task done fifty times a day is a candidate. A task done twice a year is cheaper to keep doing by hand, however tedious it feels.
2. Is the input mostly words, documents or conversations?
Current AI is strongest at reading, summarizing, classifying and drafting language. Emails, forms, call notes, contracts, product questions. If the work is mostly physical or depends on judgment nobody can write down, look elsewhere.
3. Is there a source of truth it can check?
An assistant that answers from your price list, policy manual or product catalogue can be held to those sources. An assistant asked to answer from general knowledge will sound confident and sometimes be wrong. If the source does not exist in writing, creating it is step one.
4. What happens when it is wrong?
This question decides the design. If a mistake costs a few seconds, such as a draft a person edits, automate freely. If a mistake costs money, safety or trust, keep a person approving the step. Very few useful systems need to run with nobody watching.
5. Can you tell whether it worked?
Pick the number before you start: minutes per case, enquiries answered after hours, tickets resolved without escalation. If you cannot name a measure, you will not be able to defend the project in six months.
Reading the result
- Five yeses: a strong first project. Scope it small and build it.
- Three or four: usually workable once the missing piece is fixed, most often the source of truth or the measure.
- Two or fewer: leave it. There is a better candidate somewhere else in the business.
This is a screening test, not a business case. It tells you where to look, and it takes ten minutes per workflow.
