Guide library
Getting started (3)
What is AI readiness?What does an AI readiness assessment involve?AI readiness glossaryBuilding the foundations (5)
Process mapping before automationWhen has an SME outgrown spreadsheets?How to choose business software that fitsHow to prepare your data for AIDo SMEs need a data warehouse?By market (5)
AI readiness for Singapore SMEsAI readiness for Malaysian SMEsAI readiness for Indonesian SMEse-Invoicing in Malaysia: a guide for SMEsAplikasi stok barang untuk UMKM (Bahasa Indonesia)The framework
SEL Foundation IndexAll guides →Resources · Data guide
How to prepare your data for AI: a checklist for SMEs
AI tools learn from, and act on, a business’s own data. The quality of that data therefore sets a limit on what AI can do for you. This guide explains what AI-ready data means, gives a ten-point checklist you can work through, and sets out the order in which to fix what you find.
The short answer
What AI-ready data means.
Data is ready for AI when it is accurate, complete, consistent, accessible to the tools that need it, and handled in line with privacy rules.
These five qualities appear in most data readiness guidance, although the labels vary. Accuracy means the records are correct. Completeness means the fields that matter are filled in. Consistency means the same thing is recorded and defined the same way everywhere. Accessibility means the systems and people who need the data can reach it without copying it by hand. Compliance means personal data is collected and used within the law, which in Singapore means the Personal Data Protection Act (PDPA).
You do not need all of your data to meet these standards before you start with AI. You need the data behind your first AI use to meet them. Preparing data one use at a time keeps the work manageable.
Why it matters
What happens when data is not ready.
An AI tool cannot tell good data from bad. It treats whatever it is given as correct, and it has no way of knowing otherwise.
If customer records are duplicated, an AI tool may count the same customer twice when it forecasts demand. If product codes differ between the sales and inventory systems, it cannot match what was sold with what was in stock. If key fields are often left blank, it either guesses the missing values or ignores those records.
In each case, the output looks confident but is wrong. The business may then stop trusting AI altogether, even though the real problem was the data.
This is why preparing data usually takes longer than setting up the AI tool. It is also why the work is worth doing anyway: cleaner data makes ordinary reports faster and more reliable too.
“An AI tool treats whatever data it is given as correct. It has no way of knowing otherwise.”
The checklist
Ten checks to run on your data.
Work through these checks for the data behind the first AI use you have in mind. Each check names the problem to look for and what to do about it.
| Check | What to look for | What to do |
|---|---|---|
| 1. Know where the data lives | Can you list every system and spreadsheet that holds the data this use needs? | Make a simple list of each source, who owns it, and how often it is updated. |
| 2. Name an owner | Is one person responsible for the quality of each main set of data, such as customers or products? | Assign an owner for each. Without an owner, errors are noticed but not fixed. |
| 3. Agree definitions | Do terms such as “active customer”, “order date” or “gross margin” mean the same thing in every team? | Write down one agreed definition for each key term, and use it in every report. |
| 4. Remove duplicates | Does the same customer, supplier or product appear more than once under slightly different names? | Merge the duplicates, and set a rule for how new records are created. |
| 5. Fill the gaps | Are important fields, such as product category or customer industry, often left blank? | Decide which fields are required, fill the most important gaps, and make those fields mandatory from now on. |
| 6. Standardise formats | Are dates, units, currencies and codes recorded in different formats? | Choose one format for each, and convert existing records to it. |
| 7. Connect the sources | Is data copied by hand from one system to another? | Connect the systems directly where possible, so that data moves without anyone re-entering it. |
| 8. Keep enough history | Do you have reliable records going back far enough for the use you have in mind? | Uses such as forecasting usually need at least a year of history, so that seasonal patterns show. If you do not have it, start keeping clean records now. |
| 9. Control access | Can you say who is allowed to see and change each set of data? | Set permissions by role, and remove access that is no longer needed. |
| 10. Check personal data | Do you know which records contain personal data, and whether it was collected with proper consent? | Review personal data against the PDPA before using it in any AI tool, especially one run by an outside provider. |
Run these checks for one use at a time, rather than across the whole business at once.
The order
Fix the problems in the right order.
The checks depend on one another, so the order in which you fix problems matters. In our experience, the sequence below avoids most rework. It follows a wider principle in our SEL Foundation Index: processes come before systems, and systems come before data.
- 01
Ownership and definitions first.
Before cleaning anything, agree who owns each set of data and what the key terms mean. Otherwise, the data may be cleaned to the wrong standard and then have to be cleaned again.
- 02
Then the processes that create the data.
Most data problems start at the point where data is entered. For example, if the sales process lets staff skip fields, cleaning the existing records only fixes the past. Changing the process also fixes the future.
- 03
Then clean and connect.
With owners, definitions and processes in place, remove duplicates, fill gaps, standardise formats and connect the systems.
- 04
Then keep it clean.
Set a simple monthly routine to check data quality. This catches new problems early, before they build up again.
Common questions
Questions about preparing data for AI.
Do we need a data warehouse before using AI?
Not usually at the start. Many first AI uses can work directly with data from one or two well-maintained systems. A central data store becomes useful when AI needs to combine data from several systems, such as sales, inventory and finance. At that point, it saves the manual work of bringing the data together each time.
How long does data preparation take?
For a single, focused use case, it often takes a few weeks to a few months. The time depends on how many sources are involved and how much cleaning is needed. Preparing data for the whole business at once takes much longer, which is why we recommend starting with one use case.
Can AI tools clean the data for us?
Partly. Some tools can suggest likely duplicates, spot unusual values or standardise formats. However, they cannot decide what your business terms should mean or who owns the data. Those decisions still need people.
Is it safe to put our business data into public AI tools?
It depends on the tool and its terms. Some public tools may store what you enter or use it to improve their models. Before using any AI tool with customer, employee or commercially sensitive data, check the provider’s data handling terms and your obligations under the PDPA.
Where should we start?
Pick one AI use you would like to try, list the data it needs, and run the ten checks on that data only. Our free Readiness Snapshot can also show whether data is the foundation most likely to hold your business back.
Find out where you stand
See how ready your foundations are.
The free Readiness Snapshot takes about five minutes. It emails you a PDF report showing where your processes, systems and data stand on the path to AI, and which gap to close first.