What data do I need before building an AI product?

Eighty-seven percent of AI projects never reach production (Gartner, 2022). The usual suspect is the model. The actual culprit, almost every time, is the data feeding it.

Founders pour months into model selection, architecture debates, and vendor evaluations. Then the product launches and the predictions are wrong, the recommendations are irrelevant, or the chatbot hallucinates answers that sound confident and mean nothing. The root cause traces back to the same place: the training data was incomplete, mislabeled, stale, or simply not the right data for the problem.

This article walks through what data you actually need before you write a single line of code, how much of it, how clean it has to be, and where to get it if you are starting from zero.

Every AI product consumes one of four broad data types: text, numbers, images, or user behavior logs. The product you want to build determines which type you need, and mixing that up is the fastest way to waste six figures on a project that never works.

A customer support chatbot needs thousands of real conversations between your support team and your customers. Not FAQ pages. Not marketing copy. Actual back-and-forth exchanges where a customer describes a problem and a human resolves it. The model learns from that resolution pattern.

A demand forecasting tool needs historical transaction records: what sold, when, how much, and alongside what external factors (weather, holidays, promotions). McKinsey's 2022 analysis found that companies using demand forecasting with clean historical data reduced inventory costs by 20-30%.

A fraud detection system needs labeled examples of both legitimate and fraudulent transactions. The tricky part: fraud is rare. In a typical financial dataset, fewer than 2% of transactions are fraudulent (Nilson Report, 2022). Your model needs enough examples of both to learn the difference.

The pattern: your data must mirror the decision your product will make. If the product recommends items, you need records of what people chose. If the product classifies documents, you need documents that are already sorted. If the product predicts churn, you need records of customers who stayed and customers who left, with every data point you had about them before they made that choice.

A machine learning model is a pattern-matching engine. You feed it thousands of examples where you already know the right answer, and it figures out which patterns in the input data predict that answer. Then, when it sees new data without an answer, it applies those patterns.

Think of it like training a new employee. You sit them next to an experienced rep and let them watch 5,000 customer calls. They start noticing patterns: angry tone plus billing keyword usually means refund request. Short call plus tracking number usually means delivery check. After enough examples, they can handle new calls on their own because they have internalized the patterns.

That is what a model does with numbers. It finds correlations between inputs and outputs, weighted by how reliably each correlation predicts the correct answer. Stanford's 2022 AI Index report found that modern models can identify patterns across hundreds of variables simultaneously, something no human analyst could do manually.

Why does this matter for your data strategy? Because the model can only learn patterns that exist in your data. If your sales records do not include weather data, the model will never learn that umbrella sales spike when it rains. If your customer data does not include how long someone has been a customer, the model cannot learn that tenure predicts loyalty. The model is only as smart as the data you train it on.

A 2022 IBM survey found that poor data quality costs organizations an average of $12.9 million per year. For AI products specifically, the cost is higher because bad data does not just produce bad reports. It produces confidently wrong predictions that people act on.

The honest answer: it depends on the complexity of the task. But "it depends" is not useful, so here are concrete ranges based on what works in production.

For classification tasks (sorting things into categories), you need a minimum of 500-1,000 labeled examples per category. A spam filter needs at least 1,000 spam emails and 1,000 legitimate emails. A sentiment analyzer needs at least 1,000 positive reviews and 1,000 negative reviews. Research from Google Brain (2022) showed that model accuracy improves sharply up to about 5,000 examples per category, then gains slow down.

For prediction tasks (forecasting numbers), you need 12-24 months of historical data at minimum. A demand forecasting model trained on three months of data will miss seasonal patterns entirely. Training on two years gives the model a chance to see every season twice and learn which patterns repeat.

For text generation (chatbots, content tools), the situation changed dramatically in late 2022 with the release of large language models. Instead of training from scratch, you can now fine-tune a pre-built model on a much smaller dataset. Fine-tuning a large language model for a specific domain requires as few as 500-2,000 high-quality examples (OpenAI, 2022).

Here is the practical takeaway: you probably have less data than you think you need and more than you think you have. Most companies sit on years of transaction logs, support tickets, and user behavior data that has never been organized for training purposes. A good data audit often reveals that the raw material is already there. It just needs cleaning and labeling.

Bad data does not just make models less accurate. It makes them confidently wrong. And a product that gives wrong answers with high confidence is worse than one that admits it does not know.

Missing values are the most common problem. If 30% of your customer records are missing the "industry" field, the model either ignores industry entirely or learns the wrong relationship between industry and whatever you are predicting. MIT researchers found in 2022 that datasets with more than 20% missing values in any single column produced models with 15-40% lower accuracy than datasets with complete records.

Label errors are more dangerous because they are harder to spot. If a human reviewer tagged 5% of your training examples incorrectly, the model learns those mistakes as truth. A 2021 study from MIT found systematic label errors in 10 widely used AI benchmarks, including datasets used to train commercial products. When they corrected the labels, model performance improved by up to 6 percentage points without changing anything else.

Class imbalance trips up nearly every first-time AI builder. If your fraud detection dataset contains 98% legitimate transactions and 2% fraud, a model can achieve 98% accuracy by simply labeling everything as legitimate. It looks great on paper and catches zero fraud. The fix is either oversampling the minority class, undersampling the majority, or using evaluation metrics that account for imbalance.

A Gartner study from 2022 found that organizations that invested in data quality before building AI products were 2.5 times more likely to reach production deployment. The founders who skip this step end up rebuilding six months later with twice the budget.

Starting from zero is more common than most founders admit. If your company is pre-launch or has been operating without systematic data collection, you have three practical options.

Public datasets are the fastest starting point. Kaggle hosts over 50,000 datasets across every industry. The US government's data.gov publishes economic and demographic data. The World Bank, EU Open Data Portal, and dozens of academic institutions publish freely usable datasets. A 2022 survey by Databricks found that 40% of production AI models incorporated at least one public dataset during initial development.

Synthetic data generation has become a credible alternative for companies that cannot share real customer information. Instead of using real records, you create artificial records that share the statistical properties of real data without containing any actual personal information. Gartner predicted in 2022 that by 2024, 60% of the data used for AI would be synthetically generated. That prediction tracked close to reality. Synthetic data works especially well for training fraud detection models, where real fraud examples are scarce and sensitive.

Data partnerships with complementary businesses can fill gaps that public data cannot. A logistics startup might partner with a weather data provider to improve delivery time predictions. A retail AI company might license anonymized transaction data from payment processors. These partnerships cost $5,000-$50,000 per year depending on the dataset, which is a fraction of the cost of collecting equivalent data yourself.

A practical starting path for most founders: begin with whatever internal data you have, supplement it with one relevant public dataset, build a prototype, and use the prototype's performance to identify exactly which data gaps to fill next. Trying to assemble the perfect dataset before writing any code is the planning equivalent of analysis paralysis.

Raw data is never ready for training. Even well-maintained databases contain inconsistencies that will confuse a model. The cleaning and labeling process typically consumes 60-80% of the total time spent on an AI project (CrowdFlower/Appen, 2022). Founders who budget two weeks for data preparation and two weeks for model building should reverse those numbers.

Cleaning means standardizing formats, handling missing values, removing duplicates, and fixing obvious errors. Dates stored as "12/9/2022" in one column and "2022-12-09" in another will confuse any model that tries to learn time-based patterns. Customer names with inconsistent capitalization, addresses with varying abbreviations, currencies mixed between dollars and euros: all of these need standardization before training.

Labeling means telling the model what the correct answer is for each example in your training data. If you are building a support ticket classifier, someone has to read each ticket and tag it: billing issue, technical problem, feature request, complaint. If you are building a sentiment analyzer, someone has to mark each review as positive, negative, or neutral.

The cost of labeling depends on complexity. Simple binary labels (spam/not spam) cost $0.05-$0.10 per example through services like Amazon Mechanical Turk or Scale AI. Complex labels that require domain expertise (medical image annotation, legal document classification) cost $1-$5 per example because you need trained specialists. At 10,000 examples, that is the difference between $500 and $50,000.

A 2022 study from Stanford's HAI found that label quality had 3-5 times more impact on model performance than label quantity. Two thousand perfectly labeled examples consistently outperformed ten thousand examples with 10% label noise. The practical lesson: invest in labeling quality (clear instructions, multiple reviewers, disagreement resolution) rather than racing to label as many examples as possible.

Timespade builds data preparation pipelines alongside AI products, handling the cleaning, labeling workflow setup, and validation before any model training begins. For a mid-size dataset, that preparation costs $3,000-$8,000, while a Western data consultancy charges $15,000-$30,000 for equivalent work. The preparation phase takes 2-4 weeks depending on data complexity.

Data privacy law caught up with AI faster than most startups expected. If your training data contains personal information, you are subject to privacy regulations whether you intended to build a "data product" or not.

GDPR (covering the European Union) requires explicit consent for collecting personal data and gives individuals the right to request deletion of their data from your systems, including your training datasets. A model trained on data that someone later requests deleted creates a technical and legal headache. The GDPR fines are not theoretical: regulators issued over 1.6 billion euros in fines during 2022 alone (DLA Piper GDPR Fines Report, 2023).

CCPA (covering California) gives consumers the right to know what personal data a company collects and to opt out of its sale. If your AI product uses California consumer data for training, you need disclosure mechanisms and opt-out functionality built into your product from day one.

Copyright is a less obvious but equally serious concern. Training a model on copyrighted content without permission is legally contested territory as of late 2022. Several lawsuits are pending against companies that trained models on scraped web content. The safest approach: use data you own, data you have licensed, or data that is explicitly public domain.

Practical steps that keep you out of legal trouble: anonymize personal data before training (strip names, emails, phone numbers, and any combination of fields that could identify an individual). Document the source of every dataset and its license terms. Build a deletion pipeline so you can retrain without specific records if someone exercises their right to be forgotten. These steps cost $2,000-$5,000 upfront with a team like Timespade, compared to $10,000-$25,000 at a Western compliance consultancy. The alternative, a GDPR fine, starts at 2% of global annual revenue.

An AI product is not a one-time build. The model degrades over time as the real world changes and the training data grows stale. Proper data storage and versioning are what separate a product that improves month over month from one that slowly breaks.

Version your datasets the same way software teams version code. Every time you add new training data, retrain the model, or correct labels, save that version with a timestamp and a description of what changed. When a new model version performs worse than the previous one, you need the ability to roll back to the exact dataset that produced the last good version. Without versioning, debugging becomes guesswork.

Storage costs are often overestimated. A 2022 Cloudflare analysis found that cloud storage prices dropped 90% over the previous decade. Storing one terabyte of training data costs $20-$25 per month on Amazon S3 or Google Cloud Storage. Even large datasets with millions of records rarely exceed a few terabytes. The expensive part is not storing the data. It is organizing it so your team can find and reproduce any version on demand.

Retrain frequency depends on how fast your domain changes. E-commerce recommendations should retrain weekly because product catalogs and user preferences shift constantly. Fraud detection models should retrain monthly because fraud patterns evolve. Demand forecasting models need retraining at least quarterly to absorb seasonal patterns.

Timespade sets up automated retraining pipelines as part of every AI product build. The system monitors model accuracy in production, flags when performance drops below a threshold you set, and kicks off retraining with the latest data. Setup costs $4,000-$10,000 depending on complexity. A Western consultancy charges $20,000-$40,000 for equivalent monitoring infrastructure. Without this system, most teams discover their model has degraded only when customers complain.

Yes, and for most startups this is the only practical approach.

Waiting until you have the "perfect" dataset before building anything is the number one reason AI projects stall before they start. A 2022 Harvard Business Review analysis found that companies taking an iterative approach to AI (starting with limited data and expanding) were 3 times more likely to reach production than companies that spent months assembling data before writing any code.

The practical path looks like this. Start with whatever data you have, even if it is only a few hundred examples. Build a prototype. Test it against real scenarios and measure where it fails. Those failure patterns tell you exactly what data you need next. Collect that specific data and retrain.

Pre-trained models have made this approach much more viable. In late 2022, large language models became available that already understand language and common knowledge from training on massive public datasets. Fine-tuning one of these models on 500-2,000 of your own examples can produce a product-ready tool for many use cases. You are not starting from zero. You are starting from a model that already knows how to read and reason, then teaching it the specifics of your business.

The cost of iteration is low. Each cycle of collecting 500-1,000 new labeled examples, cleaning them, and retraining costs $1,500-$4,000 with a team like Timespade. A Western agency charges $8,000-$15,000 per iteration. Over four iterations, that gap compounds: $6,000-$16,000 total versus $32,000-$60,000.

A staged approach also limits financial risk. If your first prototype shows the product concept does not work, you have spent $5,000-$10,000 learning that lesson, not $50,000. If it does work, each iteration makes it better with evidence guiding every improvement.

The founders who ship successful AI products in 2023 are not the ones with the most data. They are the ones who started building with imperfect data, learned from the gaps, and improved systematically. The data you collect after launch, from real users interacting with a real product, is always more useful than the data you imagine you need before launch.

Timespade builds AI products across all four stages: prototype through production. The team handles data preparation, model selection, training, deployment, and ongoing monitoring as a single engagement. Book a free discovery call to walk through your data situation and get a concrete plan for what to collect, what to clean, and what to build first.