zedechGetting Your Business Data Ready for AI (The Unglamorous Prerequisite)
The Boring Reason AI Projects Stall
When an AI project underdelivers, the post-mortem usually blames the model or the vendor. In our experience the cause is almost always further upstream: the data the system depends on was never in a usable state, and nobody checked before the build started.
This is not a glamorous problem. It is also the one that most reliably determines whether a project works.
The good news is that data readiness is a finite, knowable amount of work. You can assess it in an afternoon.
What "Ready" Actually Means
Forget the enterprise vocabulary for a moment. For a Malaysian SME putting AI into a real workflow, ready means four things.
Findable. The system can reach the data programmatically. Not "it's in the shared drive somewhere." An API, a database, an export that runs on a schedule.
Consistent. The same thing is recorded the same way. One customer is one record. A date is a date, not sometimes a date and sometimes free text.
Current. The data reflects reality closely enough for the decision being made. A stock figure updated nightly is fine for reporting and useless for telling a customer whether something is available right now.
Governed. Somebody knows what is in there, where it came from, and who is allowed to see it — which is also a PDPA question, and we covered that in PDPA and AI.
Most SMEs are strong on one or two of these and weak on the rest.
The Failure Patterns We See Most
The same customer, five times
Chan Wei Ming, CHAN WEI MING, and W.M. Chan are the same person with three phone number formats across two systems. To a human, obvious. To an automated system, three customers with fragmented histories.
Duplicate and inconsistent records are the most common blocker, and they compound: every system you add makes reconciliation harder.
Knowledge that only exists in someone's head
You want AI to answer staff questions about policy or process. You go looking for the documentation and discover the real process lives in the operations manager's memory, and the written SOP has been wrong since 2023.
AI cannot retrieve what was never written down. Frequently the highest-value part of an AI project is the documentation exercise it forces.
Data locked in the wrong format
Scanned PDFs of forms. Photos of delivery notes. A WhatsApp thread that constitutes the actual order history for a major account. All of it is real information, none of it is queryable without a conversion step that needs to be budgeted for.
The spreadsheet that runs the business
Nearly every SME has one — a workbook that has grown into critical infrastructure, with formulas nobody wants to touch and a structure only its author understands. It works, right up until you try to connect anything to it.
This does not have to be replaced before you start. But it needs to be understood, and its structure needs to be stable.
Systems that don't talk
Your accounting software, your inventory system, and your order platform each hold part of the truth. Nobody has ever needed them joined up because a human has always done the joining. Automation makes that gap visible immediately.
What to Actually Do
You do not need a data warehouse. You need enough order to make one workflow work.
Pick the workflow first, then look only at the data it touches. Trying to clean everything is how this turns into a two-year project that never ships. Automating order status enquiries? Then you care about orders, customers, and stock — and nothing else, yet.
Trace it end to end by hand. Follow one real transaction through every system it touches. Write down where the data lives at each step, what format it is in, and what breaks. This single exercise surfaces most problems.
Fix identity before anything else. If you cannot reliably say that two records refer to the same customer, product, or order, everything downstream stays unreliable. Deduplication is unexciting and pays for itself immediately.
Write down the rules that live in heads. What counts as an urgent order. When a discount is allowed. Who approves a refund. If a human currently makes the judgement, that judgement needs to be explicit before anything can automate around it.
Decide what "good enough" is. Perfect data is not the target and never arrives. Nightly-accurate is fine for some decisions and unacceptable for others. Name the standard for this workflow, hit it, move on.
Get the access sorted early. API credentials, export permissions, vendor cooperation. This is administrative rather than technical and routinely takes longer than the engineering.
The Reframe Worth Making
Clients often experience the data readiness phase as a delay before the real work. It is more useful to see it as the part of the project that produces value regardless of what happens next.
Deduplicated customer records make your marketing better whether or not you ever deploy a model. Written-down process rules make onboarding faster. A traced end-to-end workflow usually reveals a manual step that should just be deleted.
We have had projects where the discovery phase found enough waste that the client fixed the process and postponed the automation. That is a good outcome, not a failed engagement.
Where This Leaves You
If you are considering an AI project and want a quick self-assessment: pick the process you most want automated, and try to answer these three questions.
Can a program reach every piece of data this process needs, without a human exporting anything? Is each entity in that data recorded consistently enough to match reliably? Are the decision rules written down anywhere other than in a person's head?
Three yeses and you are in good shape. Three noes and the first phase of your AI project is a data project — which is worth knowing before you start, rather than three months in.
Zedech starts every engagement by mapping what the process actually does and what state the data is in — before anything gets built. Book a free discovery call for an honest read on whether your data is ready.