Question 1: What's the actual problem?

Not the AI opportunity. The business problem. Describe it without using the word "AI" or any technology term. If you can't, the project doesn't have a clear enough problem statement to build around.

The clearest version of this question is: what is the expensive, manual, error-prone thing that's happening right now that you want to stop happening? Measure it. Give it a number. "Our intake process takes 4 hours per client and we onboard 40 clients a month." That's a problem. "We want to leverage AI to streamline our operations" is not.

Question 2: Who owns the outcome?

AI projects without an internal owner fail. Not because the technology isn't good enough — because no one is accountable for making it actually work in the real world. The vendor ships something. No one drives adoption. The tool sits unused. The project gets blamed on "the AI."

Name a specific person who is responsible for the success of this project. Not a committee. One person. They should be someone who understands the workflow being automated and has the authority to change how people use it.

Question 3: What's the data situation?

Every AI system is only as good as the data it runs on. Before you commission a build, audit the data: where does it live, what format is it in, how clean is it, and who can access it?

For RAG systems, this means your documents, SOPs, and knowledge base. For CRM integrations, this means your contact data quality. For voice agents, this means your call scripts and handling procedures. If the data is a mess, plan to clean it before you build. A well-trained system on bad data produces confident wrong answers — which is worse than no AI at all.

Question 4: What does success look like?

Define it in the first meeting. Not after the build. Not as a vague goal ("we want to be more efficient") but as a specific, measurable state. "Our intake time drops from 4 hours to under 45 minutes per client within 60 days of launch." That's a success criterion. You can test against it. You can know whether you got there.

If you can't define success before you start, you'll spend the entire project arguing about whether it worked. Define it first. Hold everyone — including your vendor — to it.

Question 5: What happens when it's wrong?

AI systems make mistakes. The question isn't whether your system will be wrong — it's what happens when it is. Is there a human in the loop? Does the error surface clearly? What's the fallback?

For low-stakes errors (a chatbot gives a slightly off answer), this might just be a feedback loop. For high-stakes errors (a voice agent books the wrong appointment, a document classifier misfires), you need a real exception-handling process. Design the failure mode before you design the feature. It will save you from a serious incident post-launch.

The 5 Questions to Ask Before Starting Any AI Project

Get AI insights that matter.

Related articles

Why 80% of AI Projects Fail — And What the 20% Do Differently

AI Agents vs. AI Apps vs. AI Automation — What's the Difference, and Which Do You Need?