Everyone Has an AI Problem (Most Are Solving the Wrong One)
Andres Max
Your startup doesn’t have an AI problem. It has a product problem that might be solved with AI. This distinction is everything. Get it wrong, and you’ll spend six months building AI features that don’t move your business forward. Get it right, and AI becomes a genuine competitive advantage.
Every founder I talk to now has “AI” somewhere in their plans. Add AI to the product. Use AI for operations. Become an AI company. The pressure to “do something with AI” is enormous.
But here’s what I’ve learned from building AI-powered products over the past few years: most founders are solving the wrong problem with AI. They start with the technology and work backwards to find a use case. That’s exactly backwards.
The Wrong Way to Think About AI
Let me describe what I see most founders doing.
Pattern 1: Feature-First AI
How it sounds: “We should add AI to our product. What can AI do for us?”
The process:
- See competitor add AI feature
- Ask engineering what AI features are possible
- Build something that sounds impressive
- Launch it, hope users care
Why it fails: You’re adding technology, not solving problems. Users don’t want AI. They want outcomes. If AI doesn’t make the outcome better, faster, or cheaper, they don’t care.
Pattern 2: Hype-Driven AI
How it sounds: “Investors want to see AI. We need an AI story.”
The process:
- Fundraising pressure to be an “AI company”
- Retrofit AI narrative onto existing product
- Add token AI features to support the story
- Pitch AI capabilities that don’t exist yet
Why it fails: You’re building for investors, not customers. The AI features become demo-ware that nobody actually uses. And sophisticated investors see through this immediately.
Pattern 3: Solution-Looking-for-Problem AI
How it sounds: “GPT-4 can do X, Y, and Z. Let’s figure out where to use it.”
The process:
- Get excited about new AI capability
- Brainstorm applications across the product
- Build features that showcase the capability
- Wonder why adoption is low
Why it fails: Capabilities don’t equal value. Just because AI can do something doesn’t mean customers need it done.
The Right Way to Think About AI
Here’s the frame shift that changes everything:
Start with the customer problem. Then ask if AI is the best solution.
Not “what can AI do?” but “what problem are we solving, and is AI the best way to solve it?”
The AI Decision Framework
Before building any AI feature, answer these questions in order:
1. What’s the customer problem?
Be specific. Not “users want AI.” What actual pain point, in their actual workflow, are you addressing?
Good: “Users spend 2 hours per week manually categorizing support tickets.” Bad: “Users want AI-powered features.”
2. What’s the current solution?
How do customers solve this today? What’s wrong with the current approach?
Good: “They manually read each ticket and assign a category. It’s slow and inconsistent.” Bad: “They don’t have AI, so they need it.”
3. Is this problem worth solving?
Would customers pay more, stay longer, or recommend you more if you solved this?
Good: “Our churned customers cited ticket management as their #2 pain point.” Bad: “It would be cool to have AI do this.”
4. Is AI the best solution?
Could you solve this with simpler approaches? Rules? Better UX? More features?
Good: “We tried rules-based categorization. Accuracy was 60%. We need semantic understanding.” Bad: “AI is the future, so we should use it.”
5. Is the AI solution viable?
Can current AI actually solve this reliably? At what cost? With what error rate?
Good: “GPT-4 achieves 95% accuracy on our test set at $0.02 per ticket.” Bad: “AI can probably do this. Let’s try.”
If you can’t clearly answer all five questions, you’re not ready to build AI features.
Where AI Actually Creates Value
After working on multiple AI products, I’ve identified four patterns where AI genuinely helps. If your use case doesn’t fit one of these, think twice.
Pattern 1: Automation of Repetitive Cognitive Work
What it looks like: Tasks that require judgment but are repetitive enough to be predictable.
Examples:
- Categorizing support tickets
- Summarizing meeting notes
- Extracting data from documents
- Initial screening of applications
Why AI works here: Humans can do this but it’s tedious and expensive. AI can do it faster and cheaper at acceptable quality.
Key requirement: You need clear quality standards and enough volume to justify the investment.
Pattern 2: Personalization at Scale
What it looks like: Customizing experiences based on individual user behavior or preferences.
Examples:
- Product recommendations
- Dynamic content selection
- Personalized onboarding flows
- Adaptive learning paths
Why AI works here: The number of variations is too large for rules. AI can identify patterns humans can’t see.
Key requirement: You need significant user data and clear metrics for personalization success.
Pattern 3: Making Expertise Accessible
What it looks like: Taking domain expertise and making it available to non-experts.
Examples:
- Legal document review for non-lawyers
- Medical symptom triage
- Financial planning suggestions
- Code review for junior developers
Why AI works here: Expertise is scarce and expensive. AI can democratize access to good-enough answers.
Key requirement: You need to be clear about limitations and when to escalate to human experts.
Pattern 4: Enabling New Interactions
What it looks like: Making interfaces more natural or enabling previously impossible workflows.
Examples:
- Conversational interfaces to complex data
- Natural language search over documents
- Voice control for hands-free contexts
- Real-time translation enabling global collaboration
Why AI works here: The interface improvement creates genuine new value.
Key requirement: The new interaction must be significantly better, not just different.
Common AI Mistakes (And How to Avoid Them)
Mistake 1: Building AI Before Understanding the Problem
The pattern: Excited about AI capabilities, the team builds features without deeply understanding user needs.
The result: Impressive technology that nobody uses.
The fix: Validate the problem before building the solution. Show users mockups of the AI feature. Ask if they would use it. Ask what they would pay for it.
Mistake 2: Overestimating AI Reliability
The pattern: Assuming AI will work 99% of the time because it works in demos.
The result: Edge cases, errors, and hallucinations that destroy user trust.
The fix: Test on real data, not demos. Build for the failure case. Design graceful degradation. Add human review for high-stakes outputs.
Mistake 3: Underestimating AI Costs
The pattern: Building features without calculating per-query costs at scale.
The result: Unit economics that don’t work. Losing money on every customer.
The fix: Model costs before building. Calculate per-user and per-action costs. Ensure pricing supports AI expenses with margin.
Mistake 4: Ignoring the User Experience
The pattern: Treating AI output as the final product.
The result: Raw AI output is often too long, poorly formatted, or lacking context.
The fix: Design the UX around AI output. Edit, format, contextualize. Make it feel like part of your product, not a chatbot bolted on.
Mistake 5: Building Custom When APIs Exist
The pattern: Training custom models for problems that API calls solve.
The result: Months of ML engineering for marginal improvement over off-the-shelf solutions.
The fix: Start with APIs (OpenAI, Claude, etc.). Only build custom models when you’ve proven the use case AND the APIs are insufficient.
The AI Evaluation Process
Here’s my process for evaluating any AI feature idea.
Step 1: Problem Definition
Write a one-paragraph description of the problem you’re solving. Include:
- Who has this problem
- What they currently do
- Why that’s painful
- What success looks like
If you can’t write this clearly, you’re not ready to build.
Step 2: Solution Options
List at least three ways to solve the problem:
- Non-AI solution (better UX, more features, simpler workflow)
- AI-assisted solution (AI helps humans work faster)
- AI-automated solution (AI does the task independently)
Consider trade-offs for each: cost, reliability, time to build, maintenance burden.
Step 3: Prototype Testing
Before building anything real, test the concept:
- For AI automation: Manually do the task and show users the output
- For AI assistance: Create mockups of the interface
- For AI personalization: Use rules-based approximations
Get feedback on the outcome, not the technology.
Step 4: Technical Validation
Can AI actually solve this?
- Test with real data (not cherry-picked examples)
- Measure accuracy, latency, and cost
- Identify failure modes
- Design error handling
Step 5: Economics Validation
Will this make money?
- Calculate cost per user per month
- Determine if customers will pay more
- Model unit economics at scale
- Factor in error-handling costs (human review, support)
Step 6: Build Decision
Only if:
- Problem is real and valuable
- AI is the best solution (not just a solution)
- Technical validation is positive
- Economics work
Then build.
Case Study: AI Done Right
A B2B analytics company was considering adding “AI insights” to their dashboard. The typical approach would be: “Let’s have AI generate insights from the data.”
Instead, they followed the framework.
Problem definition: “Marketing managers spend 3 hours weekly digging through our dashboards to find what changed and why. They often miss important shifts because there’s too much data.”
Solution options:
- Better dashboard UX with change highlighting (no AI)
- AI that flags significant changes for human review (AI-assisted)
- AI that generates full reports automatically (AI-automated)
Prototype testing: They showed mockups of options 1 and 2 to customers. Option 1 got “nice to have” responses. Option 2 got “when can I have this?” responses.
Technical validation: They tested GPT-4 on 100 real data sets. 85% of flagged changes were genuinely significant. False positives were annoying but not harmful.
Economics validation: Feature could justify a $50/month price increase. AI costs would be $2-5 per user per month. Economics worked.
Result: They built option 2. Usage was high. Upgrade rates increased. Customers specifically cited the feature in reviews.
What they didn’t do: Build an “AI insights” feature that sounded impressive but nobody needed.
When NOT to Use AI
Sometimes the answer is “don’t use AI.” Here’s when:
The problem is too simple: If rules or basic logic solve it, AI adds complexity without value.
The stakes are too high: If errors could cause real harm, and you can’t build adequate safeguards, avoid AI automation.
The data doesn’t exist: AI needs data to work. If you don’t have enough data or the right data, AI will fail.
The costs don’t work: If AI costs make your unit economics negative, wait for costs to drop or find a different solution.
Users don’t trust AI: Some domains (legal, medical, financial) have users who deeply distrust AI output. Forcing AI on them will backfire.
A simpler solution works: Better UX, more features, or clearer workflows often solve problems more reliably than AI.
FAQ: AI Product Strategy
How do I know if my product needs AI features?
Start with customer problems, not AI capabilities. If customers have pain points that fit the four patterns (automation, personalization, accessible expertise, new interactions) AND can’t be solved more simply, AI might help. Otherwise, focus on core product improvement.
Should I build my own models or use APIs?
Start with APIs. Always. Only consider custom models when you’ve validated the use case, APIs are insufficient (accuracy, cost, latency, or data privacy), AND you have the ML expertise to maintain custom models long-term.
How do I compete with AI startups as a non-AI company?
You compete on product, not technology. AI is an ingredient, not the meal. Deep customer understanding, great UX, and domain expertise beat generic AI features. See my take on AI in product management for more on this.
What’s the minimum viable AI feature?
The smallest AI integration that solves a real problem. Often this is much simpler than founders imagine. A single API call that saves users 10 minutes is more valuable than a sophisticated AI system that impresses but doesn’t help.
Key Takeaways
- You don’t have an AI problem. You have a product problem that might be solved with AI. Start with the problem, not the technology.
- AI creates value in four patterns: Automating repetitive cognitive work, personalization at scale, making expertise accessible, and enabling new interactions.
- Before building AI, ask five questions: What’s the problem? What’s the current solution? Is it worth solving? Is AI the best solution? Is the AI solution viable?
- Common mistakes: Building before understanding, overestimating reliability, underestimating costs, ignoring UX, and building custom when APIs exist.
- Sometimes the answer is “no AI.” Simpler solutions often work better. Don’t add AI complexity to impress investors or follow trends.
What’s Next
If you’re considering AI features for your product, start by listing your top customer problems. For each one, run through the framework:
- Define the problem precisely
- List solution options (including non-AI)
- Validate with prototype testing
- Check technical feasibility
- Confirm economics work
Only then decide to build.
AI is a tool. Like any tool, it’s powerful when used for the right job and useless when used for the wrong one. The founders who win with AI aren’t the ones with the most impressive technology. They’re the ones who solve real problems effectively.
Related Reading:
- How to Fix Failing AI Initiatives - When AI projects go wrong
- Why AI Broke Product Management - The new rules for AI products
- Product Strategy for Startups - Strategy before technology
- How to Validate an AI Product - Testing AI ideas before building