Five questions before approving an AI initiative
If I were approving AI work inside an established SaaS product, I would not start with the model. I would start with these five questions.
If I were approving an AI initiative inside an established SaaS product, I would not start by asking which model we should use. I would start with five questions.
1. What customer problem are we actually trying to solve?
Not “Where can we add AI?” The better question is what users are repeatedly trying to understand, decide, or do that the product could make substantially easier.
2. What does the AI need access to?
The answer is usually more than a database. It may need workflow state, user context, permissions, historical activity, integrations, and business rules.
3. What should the AI be allowed to do?
Answering a question is one thing. Updating a record, sending something to a customer, changing a workflow, or triggering an external system is very different.
4. Where does a human need to stay in control?
Some actions can be automated. Others need review, approval, or a clear escalation path.
5. If this works, does it make the next AI capability easier to build?
This one matters more than it seems. If every new AI feature needs its own data access, permissions, orchestration, and evaluation framework, the product will accumulate AI experiments rather than become AI-native.
Capability, not features
I think this is where many AI roadmaps go wrong. They start with “What AI feature should we build?”
I would rather start with “What capability should the product gain, and what has to become true underneath it?”
That usually leads to a very different roadmap.