Don’t Start with AI. Start with the Problem.

The most common AI project failure mode: “We should use AI for something.”

Technology-first thinking produces solutions looking for problems. Problem-first thinking produces solutions that actually matter.

The Wrong Way

1. “AI is hot, we need to do something with it”
2. Find any problem that might involve AI
3. Build AI solution
4. Discover it doesn’t solve a real pain point
5. Project languishes, unused

The Right Way

1. Identify a painful, expensive, or slow process
2. Understand why it’s painful
3. Evaluate if AI is the right solution (it might not be)
4. Build only if AI uniquely enables something valuable
5. Measure impact on the original problem

Questions to Ask Before Building

Is this a real problem?

• Who experiences this problem?
• How often does it occur?
• What does it cost (time, money, frustration)?
• How do they solve it today?

Is AI the right solution?

• Could this be solved with traditional software?
• Does the problem require reasoning, generation, or understanding?
• Is “good enough” output acceptable, or do you need perfection?
• What happens when AI makes mistakes?

Can you measure success?

• What metric improves if this works?
• How much improvement would be meaningful?
• How will you know if AI is helping or hurting?

AI Is Not Magic

AI excels at:

• Pattern recognition at scale
• Language understanding and generation
• Synthesizing information from multiple sources
• Augmenting human decision-making
• Automating cognitive tasks with tolerance for imperfection

AI struggles with:

• Perfect accuracy (hallucinations are real)
• Tasks requiring current information (knowledge cutoffs)
• Deterministic outputs (same input → same output)
• Accountability (who’s responsible for AI decisions?)
• Tasks with zero error tolerance

Problem-Solution Fit

Good AI problems:

“Our support team spends 40% of their time answering the same 20 questions.” → AI can draft responses, reduce time to first reply, improve consistency

“Analysts spend 3 days preparing monthly reports from multiple data sources.” → AI can synthesize data, generate narrative, compress timeline

“Engineers waste time searching for relevant code examples in our codebase.” → AI can understand code context, surface relevant patterns

Bad AI problems (usually):

“We need AI to automate hiring decisions.” → High-stakes, requires accountability, error-intolerant

“We want AI to replace our customer service team.” → Replacement mindset vs. augmentation, misses human value

“Our competitors are using AI so we need to.” → Technology-first, no problem definition

The Build vs. Buy Decision

Before building custom AI:

1. Can you use an existing AI product? (Chat Studio, AI Lab, etc.)
2. Can you accomplish this with good prompting?
3. Do you have the data, expertise, and runway to build custom?
4. Is this a differentiating capability or commodity?

Most AI value comes from applying existing tools well, not building from scratch.

Minimum Viable AI

Start small:

• Prototype with existing models (no training)
• Test with real users on real tasks
• Measure against the baseline
• Iterate based on feedback

Don’t:

• Build elaborate infrastructure before validating value
• Train custom models before prompting fails
• Automate completely before proving augmentation works

Success Metrics

Define success before building:

Efficiency metrics:

• Time saved per task
• Tasks completed per day
• Cost per operation

Quality metrics:

• Accuracy compared to human baseline
• User satisfaction scores
• Error rates

Adoption metrics:

• Users actively engaging
• Repeat usage
• Feature utilization

If you can’t define the metric, you can’t claim success.

The Problem-First Checklist

Before starting any AI project:

• Have I clearly defined the problem?
• Do I understand who experiences it and how often?
• Do I know the current cost (time/money/frustration)?
• Is AI the right solution vs. traditional software?
• Can I tolerate AI imperfection for this use case?
• Have I defined measurable success criteria?
• Have I validated the problem with actual users?

Start with the problem. End with impact.

Explore AI solutions that solve real problems →