AdAI

What Data Do I Need for AI?

By AdAI Research Team||6 min read

For most SMB AI use cases, the answer is: less than you think, and probably data you already have. Off-the-shelf AI (ChatGPT, HubSpot Breeze, Intercom Fin) needs none of your data to start. Custom fine-tuning needs a few hundred labelled examples. The bottleneck is rarely volume; it is consistency and cleanliness. What matters is clean, connected data that reflects your business, not massive data lakes.

Key Takeaways

  • Off-the-shelf AI needs zero of your own data. The models are pre-trained. Turn them on and they work.
  • Custom fine-tuning needs a few hundred to a few thousand labelled examples per task. Cheaper and easier than most SMBs assume.
  • Data quality beats data quantity. 500 clean records beat 50,000 messy ones for most AI purposes.
  • A data warehouse (Snowflake, BigQuery) is not needed for most SMB AI. Your existing tools + middleware cover most cases.
  • Customer PII rules apply. Check vendor terms (reputable business AI does not train on your data), apply GDPR/CCPA properly, and handle deletion requests across the whole stack.
200-500
labelled training examples typically sufficient for fine-tuning a foundation model on a specific business use case, per OpenAI and Anthropic developer guidance
Source: OpenAI Fine-tuning Documentation, 2025; Anthropic Fine-tuning Documentation, 2025
~$5-50
typical cost to fine-tune GPT-4o-mini for an SMB-scale use case (a few hundred to a few thousand examples), placing custom AI within reach without enterprise budgets
Source: OpenAI Fine-tuning Pricing, 2025

What Different AI Approaches Actually Need

Off-the-shelf AI (zero of your data required). ChatGPT, Claude, Gemini, HubSpot Breeze, Intercom Fin, Microsoft Copilot. The model is already trained. Turn it on and it works. This covers the majority of SMB use cases.

Retrieval-augmented generation, aka RAG (your knowledge base + off-the-shelf model). You give the AI access to your help centre, product docs, or knowledge base. The model uses that context to answer questions specific to your business. Requires: content that already exists (support articles, product documentation, wiki). Setup: hours, not months.

Fine-tuning (a few hundred to a few thousand examples). You supply examples of the kind of output you want (customer emails in your tone, sales replies matching your style, classifications you want the model to learn). The model learns the pattern. Cost is typically $5-50 for SMB scale. Time: hours to train.

Custom classifier (a few hundred examples per category). For tabular prediction tasks (churn scoring, lead scoring, fraud detection), a custom classifier trained on your historical outcomes. Tools: Google Vertex AI AutoML, Hugging Face AutoTrain, Amazon SageMaker AutoPilot. Cost: tens of dollars.

Training a foundation model from scratch (hundreds of billions of tokens). This is a frontier-lab project. OpenAI, Anthropic, Google, Meta. No SMB does this and no SMB needs to.

Getting Your Existing Data Ready

Most SMBs already have the data they need. It just is not in a shape AI can use easily. A short cleanup usually gets you there.

Consistency of fields. The same information stored the same way. "Illinois," "IL," "Ill." should all be one value. Dates formatted uniformly. Money in one currency and to one decimal place. Fix this once and every downstream AI use benefits.

Categorical values from a fixed list. Deal stages ("prospect," "qualified," "closed-won," "closed-lost") drawn from a controlled list, not from free text. Support ticket categories the same. Lead sources the same. Free-text where structured values belong is one of the biggest sources of AI performance loss.

Outcomes labelled. For any predictive AI, the outcome you are trying to predict has to be recorded in the historical data. Deals that closed, tickets that were resolved, customers who churned. If you cannot tell which past cases were "yes" and which were "no," the AI cannot learn from them.

These are usually days of cleanup work, not months. And the improvement is not just to AI performance. The reporting your team already does gets more useful too. Clean data pays back on multiple axes at once.

Frequently Asked Questions

How much data do I need for AI to work?
Depends on the use case. For off-the-shelf AI (ChatGPT, Claude, HubSpot Breeze, Intercom Fin), zero of your own data is required to start; the model comes pre-trained. For custom fine-tuning, a few hundred to a few thousand labelled examples produce meaningful results. For training a foundation model from scratch, hundreds of billions of tokens, which is why almost no SMB does that.
What kind of data works best?
Consistent, well-labelled, and connected to real business outcomes. A CRM where every deal has a clear status. A support tool where tickets are categorised. An accounting system where transactions are correctly coded. Volume matters less than consistency. 500 clean records beat 50,000 messy ones for most AI training purposes.
What is a "clean" data set?
Records with the same field format (dates formatted the same way, categorical values from a fixed list, no free-text where a structured value belongs). Missing values kept to a minimum. Duplicates removed. Categorical labels used consistently ("USA" vs "United States" vs "US" all collapsed to one). Most SMB data cleanup projects take a few days, not months.
Do I need a data warehouse?
For most SMB AI use cases: no. The AI features built into your existing tools (HubSpot, Microsoft 365, Salesforce, Google Workspace) work directly against the data already in those tools. Middleware platforms (Zapier, Make) plumb data between tools without a warehouse. Warehouses (Snowflake, BigQuery, Redshift) become useful mainly at higher revenue tiers or when doing serious cross-tool analytics.
What if my data has customer PII?
Almost every AI use case in an SMB involves customer data. Two rules matter. First, verify your AI vendors' data-processing terms: reputable vendors (OpenAI, Anthropic, Microsoft, Google, HubSpot) do not train on your data by default when you use their business plans. Second, apply the same principles you would to any handling of PII: minimum-necessary access, retention policies, deletion request handling, GDPR/CCPA compliance across the whole chain.

Related Resources

AI Governance

The compliance and privacy layer around all AI use in a business, including data handling.

Data Enrichment

How to fill in the data gaps when your existing records are too sparse.

Custom Model

What you need when moving beyond off-the-shelf AI to something trained on your data.

AI Automation Statistics 2026

Adoption benchmarks across AI and data readiness.

Join 5,000+ SMB owners getting weekly AI agent insights

Subscribe Free