AI
How to scope an AI project before you build
The model is rarely the hard part. This is the work that decides whether an AI project becomes a system or stays a demo: the use case, the data and the evaluation set.
- Published
- 28 October 2026
- Reading time
- 6 minutes
- Written by
- PRISIUM™
Most AI projects begin with a model and a demo. The demo is convincing, everyone is enthusiastic, and then six months later there is still no system anyone uses. The problem is usually not the model. It is the work around it that nobody scoped.
Start with the decision, not the technology
The useful question is not "where can we use AI" but "which decision or process would be better if a system helped with it". Write that down in one sentence. If you cannot, you are not ready to choose a model.
A good use case has three properties: the inputs exist, the output is used by someone, and success can be measured. If any of those is missing, the project is a research exercise. That can be fine, but it should be named as such.
Look at the data honestly
AI systems are only as good as the data they can reach. Before promising anything, find out where the data lives, who owns it, how current it is and whether you are allowed to use it for this purpose.
| Question | Why it matters |
|---|---|
| Where does it live? | Determines the architecture and the cost |
| Who owns it? | Determines how quickly you can change it |
| How current is it? | Stale data produces confident wrong answers |
| Can you use it? | Permissions and purpose limitation are not details |
Most estimates drift here, not in the model. Duplicate records, conflicting rules and undocumented exceptions take time to resolve. Budget for that work separately.
Build an evaluation set first
Before you build anything, collect a set of real examples with the answer you would expect. Twenty to fifty examples is enough to start. Score the baseline approach against them.
This changes the conversation. Instead of "the answers look good", you can say "it gets 34 out of 50 right, and here are the categories it fails in". You can also see whether a simpler approach, like a good search query, already solves the problem.
Decide what happens when it is wrong
Every AI system will be wrong sometimes. Design that moment before launch. What does the user see? How do they override it? Who reviews the override? What gets logged?
A system that hides uncertainty is dangerous. A system that shows its sources and asks a person when it is unsure is useful, even when it is only partly right.
Keep the parts that matter yours
Models change. Prices change. What stays valuable is your data, your evaluation set, your rules and the interface your team uses. Keep those in your own repositories and accounts, behind interfaces that let you swap the model without rebuilding the system.
A planning shape
- Use case and data assessment: one week.
- Evaluation set and baseline test: one to two weeks.
- Build and evaluation-driven improvement: three to six weeks.
- Launch, monitoring and handover: one to two weeks.
The ranges depend on data access and how many decisions need to be made, not on which model you choose. If you want a number for your project instead of a range, the discovery sprint is where you get one.