In short: An AI pilot needs the right data, access to it, and a lawful way to use any personal data. Many SMEs keep files across systems and shared drives. Fix access and quality before you argue about models. Checked 12 Nov 2026.
The wider question of whether to start at all is the AI readiness assessment. This page is only the data.
Ten checks
1. Name the data the use case needs
Fields, document types, time period and languages. Extra years of files create personal-data and security work without improving the test. The test itself is the pilot acceptance checklist.
2. Find where it actually lives
Which system, drive, inbox or vendor. Minister Josephine Teo, at SMEs Go Digital Day on 13 October 2026, said that for many SMEs valuable business data sits across multiple systems and shared drives, which can make it harder for tools to retrieve and use. That is from the MDDI keynote. It is not a survey of your firm. Checked 12 November 2026.
3. Who can grant access
A consultant should not be given a personal login. Stale sharing links are the same problem as an over-shared Copilot tenancy. See Copilot implementation and the governance checklist.
4. Quality for this process, not a fake score
Duplicates, missing fields and inconsistent names. Write what “good enough for this pilot” means: for example, supplier name and invoice total present on the sample. Do not publish a percentage you have not measured.
5. Minimise personal data
The PDPC advisory guidelines, issued 1 March 2024 and still served when checked 12 November 2026, encourage using only the attributes required and limiting volume and time period. They are advisory. Redact or replace names before you send a sample if the test does not need them.
6. A lawful basis, checked by someone who can give advice
Meaningful consent, or a PDPA exception if you will use personal data to develop or deploy the system. This page is not that advice. If you need a specialist, use data protection officer matching. The staff rule is the customer-data policy.
7. Where the sample will be processed
Region, subprocessors, and whether the file will train someone else’s model. Those answers belong in writing. The questions are on questions to ask an AI agency.
8. Retention of prompts, logs and outputs
A pilot still creates records. What you keep, and for how long, is the data retention and audit logs guide. Do not keep a full customer file in a prompt log because it was convenient.
9. Who keeps the data current after the test
A one-off export is not a production feed. Name the person who updates the folder or the system. If that name is blank, you have an assessment finding, not a build.
10. A tidy folder versus a mixed inbox
Illustrative only. A year of supplier invoices in one format, with a named owner, can be sampled. Five years of mixed WhatsApp and email, with no owner, cannot support a confident test. This page does not claim an accuracy rate for either.
Sector versions of the matching page include F&B, retail and clinics, where the data sits in a POS, a store system or a clinic record. The brief is still one process.
WhatsApp +65 9071 3976 (chat).
Frequently asked questions
What is data readiness for AI?
Whether the data for a named use case can be found, accessed and used lawfully, and is accurate enough to test. It is not a company-wide score.
Do I need a data warehouse before a pilot?
Not necessarily. You need enough access and quality for the test you wrote down. A warehouse is a later decision, not a prerequisite this page invents.
Can I use customer data to train a model?
Only in line with the PDPA: meaningful consent or a valid exception, minimisation, and a contract with the provider. The PDPC’s 1 March 2024 guidelines are advisory. This is not legal advice.
