Building a Data Foundation Before You Build AI
What a data platform for AI needs: inventory, owners, definitions, trusted pipelines, quality checks and PDPA controls, built in stages around one use case.
A data platform for AI is the set of pipelines, storage, quality checks and access controls that make your company's data reliable enough for models to use. Build the foundation first: know what data you have, who owns it, whether it is correct, and who may use it. Without that, an AI project spends most of its time cleaning data and still produces answers nobody trusts.
This article is for CIOs, data leads and business owners in Thailand who have an AI idea and a pile of spreadsheets, databases and file shares. It explains what a foundation contains, how to build it in stages, and when you can safely skip steps.
Why AI projects stall on data
The model is rarely the hard part. Teams discover during a pilot that the customer list exists in four systems with four spellings, that dates are stored three ways, or that nobody can say which file is current. The pilot then becomes a data cleanup project, and the schedule slips.
Common symptoms include:
- The same customer, product or branch has different IDs in different systems.
- Reports from different teams give different totals for the same measure.
- Key data lives in email attachments, scanned PDFs and personal spreadsheets.
- Nobody owns a dataset, so nobody fixes its errors.
- Access is either wide open or locked so tightly that no one can use it.
None of these problems are caused by AI, but AI exposes them quickly. A model trained on inconsistent data learns the inconsistency.
What a data foundation contains
You do not need a large platform on day one. You do need these layers, even if each starts small.
| Layer | What it does | Typical question it answers |
|---|---|---|
| Sources | Systems that create data: ERP, CRM, LINE, sensors, files | Where does the data start? |
| Ingestion | Pipelines that move data on a schedule or in real time | Does it arrive reliably? |
| Storage | A warehouse or lake with raw and cleaned copies | Where is the trusted version? |
| Modeling | Shared definitions such as "active customer" | Does everyone count the same way? |
| Quality | Automated checks for gaps, duplicates and outliers | Can we trust it today? |
| Governance | Ownership, access rules, retention, audit logs | Who may see and use it? |
| Serving | Dashboards, APIs and feature tables for models | How do people and models consume it? |
Our data and analytics practice covers these layers, from data platforms and warehouses to dashboards and model monitoring.
Step 1: Take stock before you build
Start with an inventory, not a technology choice. For each important dataset, record what it contains, which system holds it, who owns it, how often it changes and how it is used today. A shared spreadsheet is enough.
Then rank datasets by value to the AI use cases you actually plan. If your first project is a customer service assistant, your product, policy and ticket data matter more than your finance ledger. Focus effort where it pays off, and leave the rest.
Step 2: Agree on definitions and owners
Most data disputes are really definition disputes. Is a customer someone who signed up, someone who bought, or someone who bought this year? Write each key term down once, give it a business owner, and have the data team implement exactly that.
Ownership means a named person who answers questions about a dataset and approves changes. Without owners, quality decays quietly. In Thai organizations with several departments or subsidiaries, a short working group across business and IT usually settles these questions faster than a long policy document.
Step 3: Build pipelines you can trust
Pipelines should be repeatable, monitored and boring. Avoid one-off scripts that only one person understands.
- Land raw data unchanged. Keep an untouched copy so you can always reprocess.
- Clean in a separate step. Fix formats, remove duplicates, standardize codes and handle Thai and English text consistently.
- Add automated checks. Flag missing fields, impossible values and sudden volume changes before the data reaches users.
- Alert someone. A failed pipeline that nobody sees is worse than no pipeline.
- Version the logic. Keep transformation code in version control, so changes can be reviewed and reversed.
Open-source tools are often enough, and choosing them helps avoid lock-in. Decide between cloud and on-premise based on your data rules, not on fashion.
Step 4: Handle unstructured data deliberately
Much of the knowledge AI needs is not in tables. It lives in contracts, manuals, emails, scanned forms and chat logs. Treat it as data too.
Scanned Thai documents need OCR, and the output needs spot checks because errors pass silently into later steps. Documents need metadata such as owner, version and date. For assistants that answer from documents, this groundwork decides answer quality. We explain the effect in how to stop an AI assistant making things up.
Step 5: Set governance and PDPA controls early
Personal data needs rules from the start, because retrofitting them is difficult. Thailand's Personal Data Protection Act (PDPA) applies to much of what AI uses: names, phone numbers, ID numbers, health and financial details.
In practice:
- Classify datasets by sensitivity and record the lawful basis for personal data.
- Limit access by role, and log who reads what.
- Mask or remove fields the AI does not need.
- Define retention periods and a deletion process.
- Decide where data may be processed and stored. Some sectors, such as banking and government, may require in-country hosting.
Check current guidance from Thailand's PDPC and sector regulators, and confirm your plan with legal counsel. This article is not legal advice.
Step 6: Connect the first AI use case
Once one slice of data is trusted, build the first use case on it. This tests the foundation with a real demand. You will find gaps, and that is the point.
Good early candidates are forecasting, reporting and search over documents. Our AI demand forecasting demo, a working demo with sample data, shows how a clean time series of load and weather becomes a 48-hour forecast with a likely range. The model is a small part. The trustworthy inputs do the heavy lifting.
Measure whether the foundation is working
Track a few plain indicators: how many key datasets have an owner, how many pass their quality checks each day, how long it takes to give a new team access, and how often two reports disagree. Review them monthly with business owners. If the numbers improve, the foundation is doing its job. If not, you will see where it is failing.
What can go wrong
- Boiling the ocean. A program to clean all data first never ends. Work by use case.
- Tool-first thinking. Buying a platform before agreeing on definitions and owners moves the mess to a new place.
- No business ownership. If IT alone owns data, business meaning drifts.
- Quality checked once. Data degrades. Checks must run every time the pipeline runs.
- Shadow copies. Teams export data to personal files, and the official version falls behind.
- Skipping security. Centralizing data creates a larger target. Apply identity controls, encryption and monitoring, and consider alignment with ISO/IEC 27001.
When you can skip parts of this
Not every project needs a full platform. If your first use case reads a single clean system, such as a document library for an internal assistant, you can start with a light version of steps 1, 2 and 5. If the data is small and changes slowly, a well-kept warehouse is plenty. Build more only when a use case demands it.
How to start
- List your top ten datasets and name an owner for each.
- Choose one AI use case and trace its data from source to answer.
- Write definitions for the five terms that matter most.
- Set up one monitored pipeline with automated quality checks.
- Review gaps, then expand to the next use case.
If you would like an outside view of your data readiness, our team can help you assess it and plan the order of work. Contact us.
Frequently asked questions
What is a data foundation for AI?
The pipelines, storage, shared definitions, quality checks and access controls that make company data reliable for models. It answers where data comes from, who owns it, whether it is correct and who may use it.
Do I need a data lake or warehouse before starting AI?
Not always. If your first use case reads one clean system, a light version is enough. Build a warehouse or lake when several sources must be combined or reused across projects.
Who should own the data?
A named business owner for each key dataset, supported by IT. The owner answers questions, approves changes and decides on quality issues. Without owners, quality decays and definitions drift.
How does PDPA affect the data platform?
Personal data needs a lawful basis, limited use, controlled access, retention rules and logs. Set these early, check current PDPC guidance, and confirm your plan with legal counsel.
Ready to put this to work?
Turn this playbook into a working system
From a first agent to a national platform, the same team plans it, builds it and runs it with you.
