It's March 2025. You've just secured $2M in seed funding. Your board wants AI. Your investors want AI. Your potential customers are asking if you have AI. Everyone assumes that bolting machine learning onto an existing product is the path to unicorn status. Then reality hits. I've watched dozens of startup founders make the same mistake: they chase the AI opportunity instead of solving a real problem with AI. They spend six months building a system that does everything, optimize for nothing, and end up with a product that's slower, more expensive to run, and less differentiated than their non-AI competitors. with This article isn't about why AI is transformative—that's been written a thousand times. Instead, I'm sharing a framework that early-stage founders and technical leads can use to evaluate whether AI actually belongs in their product, and if it does, how to build it without burning cash or credibility. The Three Questions Every Startup Should Ask Before Building AI The Three Questions Every Startup Should Ask Before Building AI Before writing a single line of machine learning code, ask yourself these three questions. If you can't give a confident "yes" to all three, you're probably optimizing for hype instead of business value. Question 1: Does your problem actually require adaptive learning? Question 1: Does your problem actually require adaptive learning? This is the filter most startups miss. Not every hard problem is an ML problem. A pricing optimization tool might be better served by rule-based logic. A document classification task might be 80% solved by keyword matching and regex patterns. A recommendation engine might work fine with collaborative filtering before you jump to neural networks. Here's the real talk: machine learning is useful when you have a problem where the optimal solution changes based on input patterns you can't pre-define. If you can hand-code the rules or the solution is deterministic, you're adding complexity and latency for no reason. Question 2: Do you have the data infrastructure to support it? Question 2: Do you have the data infrastructure to support it? I worked with a B2B SaaS startup that wanted to build predictive analytics for enterprise clients. Their analytics engine would help customers anticipate demand fluctuations. Great idea. But they'd been collecting data for only four months, and they had no data pipeline, no real-time logging system, and no way to backfill historical data in a structured format. Building the ML model took three weeks. Building the data infrastructure took four months. Your AI product is only as good as the data feeding it. Before you commit to machine learning, ask: Where's the data coming from?How reliable is that source?Can you access it in real-time or batch?Do you have historical data to train on?What's your plan for handling bad data, missing values, or shifting distributions? Where's the data coming from? How reliable is that source? Can you access it in real-time or batch? Do you have historical data to train on? What's your plan for handling bad data, missing values, or shifting distributions? If you don't have confident answers, plan for data infrastructure before you plan for models. Question 3: Can your business model support the cost? Question 3: Can your business model support the cost? This is where I see startups get burned most often. GPT-4 API calls cost money. Running inference at scale costs money. Retraining models costs money. GPU instances for fine-tuning cost money. Real-time model serving costs money. A few months ago, I reviewed a startup that built an AI-powered copywriting tool. Their unit economics looked great—customers paid $29/month. But each customer generated roughly $0.15 in API costs per month just for inference, and they were burning another $0.20 per customer in retraining and infrastructure. Add in salaries, and they were losing money at scale. The founder hadn't done the math. He assumed that because LLMs were cheap, the business would work. It didn't. Do the honest math: What are your inference costs per unit? What are your training costs? At what customer acquisition cost and lifetime value does this pencil out? If the answer is "we'll figure it out later," you're not ready. The Startup AI Framework: Problem → Data → Model → Shipping Once you've decided AI makes sense, here's a practical framework for building it without overcomplicating things. Phase 1: Problem Clarity (Week 1-2) Phase 1: Problem Clarity (Week 1-2) Define the exact business problem you're solving. Not "we want to use AI." Not "we want better predictions." Specific. Examples: "Reduce the time our support team spends categorizing customer tickets from 3 hours/day to 30 minutes/day""Identify which leads are most likely to convert, so our sales team can prioritize outreach""Recommend relevant products to customers, increasing average order value by 15%" "Reduce the time our support team spends categorizing customer tickets from 3 hours/day to 30 minutes/day" "Identify which leads are most likely to convert, so our sales team can prioritize outreach" "Recommend relevant products to customers, increasing average order value by 15%" Write this down. Share it with your team. Make sure you all agree on what success looks like in business terms, not technical terms. Phase 2: Data Strategy (Week 2-4) Phase 2: Data Strategy (Week 2-4) Before touching any ML frameworks, understand your data. Where will you source it?How much training data do you actually need? (Usually less than startups think)How will you label it? (Manually, programmatically, or using weak labels?)What's your validation strategy? Where will you source it? How much training data do you actually need? (Usually less than startups think) How will you label it? (Manually, programmatically, or using weak labels?) What's your validation strategy? Here's a template I use: Data SourceVolumeLatencyQualityLabeling StrategyUser database50K recordsReal-timeHighRules-based + manual auditEvent logs2M events/day24h batchMediumProxy metricsThird-party API10K recordsSync nightlyMediumDepends on provider Data SourceVolumeLatencyQualityLabeling StrategyUser database50K recordsReal-timeHighRules-based + manual auditEvent logs2M events/day24h batchMediumProxy metricsThird-party API10K recordsSync nightlyMediumDepends on provider Data SourceVolumeLatencyQualityLabeling Strategy Data SourceVolumeLatencyQualityLabeling Strategy Data Source Volume Latency Quality Labeling Strategy User database50K recordsReal-timeHighRules-based + manual auditEvent logs2M events/day24h batchMediumProxy metricsThird-party API10K recordsSync nightlyMediumDepends on provider User database50K recordsReal-timeHighRules-based + manual audit User database 50K records Real-time High Rules-based + manual audit Event logs2M events/day24h batchMediumProxy metrics Event logs 2M events/day 24h batch Medium Proxy metrics Third-party API10K recordsSync nightlyMediumDepends on provider Third-party API 10K records Sync nightly Medium Depends on provider Even a simple spreadsheet like this forces clarity. And clarity beats sophisticated data engineering when you're bootstrapping. Phase 3: Start with the Simplest Model That Works (Week 4-6) Phase 3: Start with the Simplest Model That Works (Week 4-6) Here's where startups' eyes glaze over. They want to build a transformer. They want to use cutting-edge architecture. They want to fine-tune GPT-4. Don't. Start with the simplest baseline that solves the problem: Logistic regression for binary classificationNaive Bayes for text classificationLinear regression for predictionsOff-the-shelf models (using Hugging Face, OpenAI API, or Anthropic Claude) Logistic regression for binary classification Naive Bayes for text classification Linear regression for predictions Off-the-shelf models (using Hugging Face, OpenAI API, or Anthropic Claude) Yes, a fancy model might give you 95% accuracy instead of 87%. But if the simple model is already better than the human baseline and ships in two weeks, that's the right choice. I worked with a marketplace startup trying to detect fraud. The founding engineer wanted to build a custom LSTM. I convinced him to start with XGBoost on engineered features. It took three days to build, achieved 92% precision/recall, and worked in production for six months before they even needed to iterate. Phase 4: Build the Feedback Loop (Week 6-8) Phase 4: Build the Feedback Loop (Week 6-8) The model is just one piece. You need a system to understand how it's performing in production, catch failures, and collect data for retraining. This is boring infrastructure work. It's also the difference between a product that gets better over time and one that decays into uselessness. At minimum: # Pseudocode for logging predictions + actuals
def make_prediction(user_id, features):
    prediction = model.predict(features)
    
    # Log everything
    log_event({
        "user_id": user_id,
        "prediction": prediction,
        "confidence": prediction.confidence,
        "features": features,
        "timestamp": now()
    })
    
    return prediction

# Later, when you observe the true outcome:
def record_actual(user_id, actual_outcome):
    log_event({
        "user_id": user_id,
        "actual": actual_outcome,
        "timestamp": now()
    }) # Pseudocode for logging predictions + actuals
def make_prediction(user_id, features):
    prediction = model.predict(features)
    
    # Log everything
    log_event({
        "user_id": user_id,
        "prediction": prediction,
        "confidence": prediction.confidence,
        "features": features,
        "timestamp": now()
    })
    
    return prediction

# Later, when you observe the true outcome:
def record_actual(user_id, actual_outcome):
    log_event({
        "user_id": user_id,
        "actual": actual_outcome,
        "timestamp": now()
    }) This simple logging infrastructure lets you compute metrics, detect drift, and identify failure patterns. A Real Example: The Ticket Routing Startup Let me walk through a real scenario. A startup called TicketFlow is building a customer support platform. They want to automatically route tickets to the right team (Sales, Support, Technical, Billing). Problem clarity: Right now, tickets sit in a queue until someone manually categorizes them. This causes 2-hour delays. They want to route 90% of tickets correctly in under 1 second. Problem clarity Data strategy: They have 6 months of historical tickets (~12K total), all manually categorized. They can label new tickets as they come in. Data strategy Model choice: Start with a text classifier. Don't overthink it. Model choice from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import MultinomialNB
from sklearn.pipeline import Pipeline

# Simple, fast, interpretable
classifier = Pipeline([
    ('tfidf', TfidfVectorizer(max_features=1000)),
    ('nb', MultinomialNB())
])

classifier.fit(tickets, categories) from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import MultinomialNB
from sklearn.pipeline import Pipeline

# Simple, fast, interpretable
classifier = Pipeline([
    ('tfidf', TfidfVectorizer(max_features=1000)),
    ('nb', MultinomialNB())
])

classifier.fit(tickets, categories) This trains in seconds, explains itself (you can see which words matter), and achieves 82% accuracy. Is that perfect? No. But it's better than humans, it's much faster, and it works in production without GPU costs. Shipping: Integrate the prediction into their ticket system. Log the prediction, the actual category assigned by humans, and measure real-world accuracy over time. After two weeks, they have enough data to know if it's helping. Shipping This isn't flashy. It's not a transformer. But it works, and it's defensible: they've built a product that saves their customers time, with clear ROI. The Mistakes I See Startups Make (And How to Avoid Them) Mistake 1: Premature Sophistication Startups often over-engineer their first model. They add complexity, latency, and costs before they've proven the basic idea works. Mistake 1: Premature Sophistication Fix: Bias toward simplicity. Use off-the-shelf models or frameworks. Prove value first, optimize later. Fix Mistake 2: Ignoring Data Quality A brilliant model with garbage data produces garbage predictions. Startups often rush to the "sexy" ML part and skip data cleaning. Mistake 2: Ignoring Data Quality Fix: Spend 30% of your time on data. Run exploratory analysis. Look at outliers. Check for class imbalance. Manually review a sample of predictions. Fix Mistake 3: Building Without a Metric You can't improve what you don't measure. Many startups build a model, ship it, and have no idea if it's actually helping users. Mistake 3: Building Without a Metric Fix: Define your metric before you build. For a classifier: precision, recall, F1-score. For a recommender: click-through rate, conversion rate. For a predictor: MAE, RMSE. Track it obsessively. Fix before Mistake 4: Assuming Your First Model is Final It's not. Models degrade. User behavior changes. Competitors iterate. You'll need to retrain, adjust, and improve. Mistake 4: Assuming Your First Model is Final Fix: Plan for iteration from day one. Build infrastructure that makes retraining easy. Version your models. A/B test changes. Fix Key Takeaways Key Takeaways Not every hard problem is an ML problem. Ask three questions before committing to AI: Does it require adaptive learning? Do you have data infrastructure? Can your business support the costs?Start simple. Logistic regression, XGBoost, or off-the-shelf APIs beat cutting-edge architectures for early-stage startups. Prove the concept works before adding complexity.Data infrastructure is as important as the model. Spend time on data sourcing, validation, and labeling. A 80% accurate model with clean data beats a 95% accurate model with garbage data.Ship with a feedback loop. Log predictions, track real-world performance, and plan for retraining. The model on day one isn't the model on day 100.Do the unit economics math. Calculate your inference costs, training costs, and infrastructure costs. If it doesn't pencil out, fix the business model, don't hope the costs disappear.Focus on customer value, not technical cleverness. Your investors don't care that you trained a transformer. Your customers care that you saved them three hours a week. Not every hard problem is an ML problem. Ask three questions before committing to AI: Does it require adaptive learning? Do you have data infrastructure? Can your business support the costs? Not every hard problem is an ML problem. Start simple. Logistic regression, XGBoost, or off-the-shelf APIs beat cutting-edge architectures for early-stage startups. Prove the concept works before adding complexity. Start simple. Data infrastructure is as important as the model. Spend time on data sourcing, validation, and labeling. A 80% accurate model with clean data beats a 95% accurate model with garbage data. Data infrastructure is as important as the model. Ship with a feedback loop. Log predictions, track real-world performance, and plan for retraining. The model on day one isn't the model on day 100. Ship with a feedback loop. Do the unit economics math. Calculate your inference costs, training costs, and infrastructure costs. If it doesn't pencil out, fix the business model, don't hope the costs disappear. Do the unit economics math. Focus on customer value, not technical cleverness. Your investors don't care that you trained a transformer. Your customers care that you saved them three hours a week. Focus on customer value, not technical cleverness. The Road Ahead The Road Ahead We're in a moment where AI is genuinely useful and genuinely hyped at the same time. That's dangerous for startups. It's easy to build cool technology that nobody needs. It's hard to build simple technology that solves a real problem. The startups that win in this cycle won't be the ones with the fanciest models. They'll be the ones who stayed disciplined, proved their metrics, and iterated faster than competitors. If you're building an AI product, ask yourself: Would I rather ship a simple solution that works in two weeks, or build the perfect solution in three months? Most early-stage startups should pick the first option. The opportunities are real. The execution standards are higher than ever. That's exactly the kind of environment where boring, disciplined work beats hype.

This story contains new, firsthand information uncovered by the writer.

OpenAI

The Three Questions Every Startup Should Ask Before Building AI

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

How to Win in AI Search: The New Rules of Visibility in 2026

How to Win in AI Search: The New Rules of Visibility in 2026

Optimizing Local LLM Inference for 8GB VRAM GPUs

Interface Singularity

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

How to Win in AI Search: The New Rules of Visibility in 2026

How to Win in AI Search: The New Rules of Visibility in 2026

Optimizing Local LLM Inference for 8GB VRAM GPUs

Interface Singularity

The Noonification: How Often Do NFTs Pass The Howey Test? (1/13/2023)

Darwin's Hybrid Intelligence to Align AI & Human Goals for Startups & VCs

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps