5 questions you must ask before deploying AI
Before you spend a single dollar on AI, answer these questions. You'll save yourself months of frustration.
Most AI projects don't fail for technical reasons. They fail because nobody asked the right questions at the start. According to McKinsey, only 39% of companies see any impact of AI on EBIT. And most of those that do report less than 5%. These five questions are the difference between success and an expensive lesson.
1. What specific problem are we solving?
Sounds obvious, but I've seen dozens of companies "deploying AI" with no clear answer to this question. "We want to be more innovative" is not a problem. "Our team manually categorizes tickets, 40 hours a week". That's a problem.
Without a specific problem, you have no success metric. Without a metric, you have no idea whether AI works better than an Excel sheet with macros.
2. Do we have the data?
AI without data is a car without fuel. Quantity is only the start. Quality, availability, and GDPR compliance matter just as much.
DATA CHECKLIST
- ✓ Is the data in one place or scattered across 15 systems?
- ✓ Do we have at least 6-12 months of history?
- ✓ Is the data compliant with GDPR/industry regulations?
- ✓ Do we have consent to use it for ML?
This is what I do hands-on — advising on AI strategy and building agents that survive the demo.
3. Who will be the business owner?
AI projects without a business owner die a natural death. IT will deploy it, but without a champion on the business side nobody will use the solution.
The business owner is the person who:
- • Has the budget and can make decisions
- • Understands the process we're automating
- • Owns whether the team actually uses it
- • Has skin in the game (their KPI depends on success)
4. What's the plan B?
What happens when the AI gets it wrong? It will get it wrong. The only question is when.
You need:
- • A process that escalates to a human
- • Monitoring that catches anomalies
- • A way to quickly switch off the AI and fall back to the old process
"AI that works in 95% of cases is a great result. But at 10,000 transactions a day, that 5% of errors is 500 cases someone has to handle by hand."
5. How will we measure success after 90 days?
Not after a year. Not "someday". After 90 days you need to know whether it works.
Concrete metrics:
- • Process time: from X hours to Y hours
- • Cost: from X to Y per transaction
- • Quality: errors dropped from X% to Y%
- • Adoption: Y% of the team uses the solution
The simplest test
Try answering each of these five questions in a single sentence. If any one of them has no answer, the project isn't ready yet. Go back to them before you write the first line of code.
