Data is the foundation of AI Agents - what do SMBs need to prepare?
Diagram: data from 4 core sources determines the quality of everything an AI Agent produces.
Many businesses rush to deploy an AI Agent while skipping the most important question: is our data ready? An AI Agent is only as smart as the data you feed it - and most SMBs have not yet built even the minimum data foundation.
An AI Agent runs on the data your business provides - it does not "just know" everything. Small and medium businesses need to prepare 4 data types: customers, products or services, internal processes (SOPs), and interaction history. Two to four weeks of basic cleaning and structuring is enough to pilot your first AI Agent.
Why does data decide whether an AI Agent succeeds or fails?
There is a common misconception: many business owners assume an AI Agent is "smart out of the box" - switch it on and it knows your products, prices and company policies. The reality is exactly the opposite.
An AI Agent (whether it runs on GPT-4, Claude, or any other model) is just a "brain" that can reason. It needs your business's specific data to operate accurately. Without data, the Agent will:
- Make things up (hallucinate) - quote wrong prices, invent promotions that do not exist
- Give generic answers, no different from plain ChatGPT
- Have no idea who the customer is, what they bought before, or what they need next
The core principle: Garbage in, garbage out. Junk data produces junk answers. Good data produces an Agent that is smart, accurate, and creates real value.
The 4 data types every AI Agent needs
Whatever industry you are in - restaurants, spas, dental clinics, or online shops - an AI Agent needs the same 4 basic data types:
1. Customer data
Names, phone numbers, emails, purchase history, special notes (allergies, preferences, birthdays). This lets the Agent personalize and remember context.
2. Product / service data
Price lists, specifications, inventory, promotions and their conditions. This lets the Agent quote correctly and recommend the right options.
3. Process data (SOPs)
Return policies, complaint handling procedures, internal FAQ, escalation rules. This lets the Agent handle situations by the book.
4. Interaction data
Chat history, emails, calls, feedback. This teaches the Agent the right tone and stops it from re-asking questions customers already answered.
Important note: you do not need perfect data on day one. An AI Agent can start with the basics (a price list + FAQ + customer list) and improve gradually as it collects more data from every conversation. But without a minimum foundation, the Agent will do more harm than good.
How much damage does bad data cause?
Here are 3 real examples we have seen while advising Vietnamese businesses:
- Duplicate data: a spa had 2 records for the same customer (name spelled differently). The AI Agent sent the appointment reminder twice on the same day - the customer got annoyed and cancelled.
- Stale, unmaintained data: the price list in the system was 6 months old. The Agent quoted the old price, the customer agreed, then had to pay extra on arrival - trust lost, sometimes a 1-star review too.
- Unstructured data: the internal FAQ lived in Word files, each written differently. The Agent could not parse them and answered "I don't have that information" to questions it absolutely should have known.
According to a Gartner survey (2024), 40% of AI projects fail because of poor data quality - not because of the technology. It is one of the most common mistakes businesses make when deploying AI Agents.
Data preparation checklist for SMBs (2-4 weeks)
You do not need a complex system. A small business can start with a Google Sheet or Airtable. What matters is that the data exists, is clean, and is structured - no "big data" or data warehouse required.
- Week 1: consolidate customer data from every source (Zalo, Facebook, Excel files, notebooks) into ONE place. Standardize: capitalized names, 10-digit phone format, lowercase emails.
- Week 1: build a master price list - one row per product or service, with columns for name, price, short description, conditions, and status (in stock / out).
- Week 2: write an internal FAQ of 30-50 questions and answers. Source them from: what customers actually ask on Zalo/Messenger, return policies, opening hours, how to book.
- Week 2: document situation-handling SOPs: what happens when a customer complains? When stock runs out? When to hand over to a human? Write up the 10-15 most common scenarios.
- Week 3: export the last 3 months of chat history (Zalo OA, Facebook Messenger). This is the best "training data" to teach the Agent your tone and real-world handling.
- Weeks 3-4: remove duplicates, fix bad records, tag customers (new / returning / VIP). Check: does every record have a name plus at least one contact channel?
Field-tested tip: advising Vietnamese businesses, we consistently see 80% of the value come from the 20% of data that matters most: a correct price list + a 30-question FAQ + a properly formatted customer list. Those three are enough to pilot your first AI Agent within 2 weeks. Do not wait for "perfect" to start.
Data is ready - what comes next?
Once your data is in shape, the AI Agent deployment process moves much faster:
- Small pilot: start with 1 Agent for 1 specific job (for example, answering FAQ on Zalo). Use the FAQ + price list. Monitor for 1-2 weeks.
- Evaluate and expand: measure how often the Agent answers correctly and how customers respond. Fill data gaps, correct wrong answers.
- Connect more data: once the pilot is stable, connect customer data so the Agent personalizes, and interaction data so it remembers context.
- Add new Agents: a sales Agent, a marketing Agent, an internal Agent - each uses different data on the same foundation.
One thing matters most: data is not a one-time cost. It needs continuous upkeep. Many businesses invest in deploying an AI Agent but forget to maintain the data - three months later the Agent starts answering incorrectly because the data went stale. Treat data as "fuel": the Agent needs regular refills. See our AI Agent services to understand what a maintained deployment looks like.
Frequently asked questions
What types of data does an AI Agent need?
An AI Agent needs 4 core data types: customer data (names, contacts, purchase history), product or service data (price lists, specs, inventory), process data (SOPs, policies, internal FAQ), and interaction data (chat, email and call history). Missing any of these makes the Agent answer incorrectly or incompletely.
How does bad data hurt an AI Agent?
Garbage in, garbage out: duplicate records make the Agent message the same customer twice, outdated data makes it quote wrong prices, and unstructured data means it cannot find the information it needs. The result is lost customers, costly clean-up work, and lost trust in AI.
Where should a small business start with data preparation?
Start with 3 tasks: consolidate customer data from scattered sources into one place (a Google Sheet is fine), standardize formats for names, phone numbers and emails, and write an internal FAQ of 30-50 common questions. Two to four weeks is enough foundation to pilot your first AI Agent.