Most AI conversations start with the model. In practice, most AI projects succeed or stall on something less exciting: the systems the model has to read from and write back to. An agent that can't look up an order, or a forecast that can't reach the purchasing screen, stays a demo. Before you can put AI to work, the software you already run has to be ready for it.
We've spent twenty years building and maintaining enterprise software in Malaysia, and much of our AI work now starts with this groundwork. Here's how we approach it.
What "AI-ready" actually means
It doesn't mean rewriting everything. A system is ready for AI when three things are true:
- Its data can be reached. There's a reliable, secure way to read the records a model needs, without someone exporting spreadsheets every Friday.
- Its actions can be called. The things a person does in the system, such as approving a claim or creating an order, are available as APIs, with the same permissions and audit trail.
- Its events can be heard. When something important happens, other systems find out quickly, so AI can react in time to matter.
Many core systems meet none of these. That's normal, and it's fixable.
Start with an API layer, not a rewrite
The fastest route is usually to wrap what you have. We put a thin, well-documented API layer in front of the legacy system. It handles authentication, maps old data structures to clean resources and logs every call. The core system keeps running untouched underneath.
This one step unlocks a lot. Mobile apps, partner portals and AI services can all use the same layer, and you can replace parts of the old system behind it later, one at a time, without breaking anything in front.
A rewrite promises everything in two years. An API layer delivers something useful in two months, and keeps your options open.
Fix the data where it's created
AI exposes data problems quickly: free-text fields where there should be codes, duplicate customers, dates in three formats. It's tempting to clean data in a separate pipeline and leave the source alone, but the mess just comes back.
Where we can, we fix the problem at the point of entry instead: validation in the forms, dropdowns instead of free text, a single customer ID across systems. It's ordinary software engineering, and it does more for AI accuracy than any amount of model tuning.
Add events, then automation
Once data and actions are reachable, we add events for the moments that matter: a booking is made, a payment fails, stock drops below a threshold. These feed a lightweight message queue that AI services and notifications can subscribe to.
This is where AI starts to feel useful in daily work. A forecast updates when sales come in, not overnight. An agent drafts a reply when a complaint arrives, not when someone remembers to ask it.
Keep security and PDPA in the design
Opening up a core system means more ways in. Every new API needs the same access rules as the screens it replaces, and every AI service should see only the fields it needs. We log who and what accessed personal data, and we make sure data doesn't leave Malaysia without the right safeguards. Our PDPA and AI checklist covers the questions to ask.
A practical sequence
- Pick one AI use case with a clear owner and a measurable result.
- Map the systems it touches: what it needs to read, what it needs to do, and when.
- Build only those APIs and events, properly secured and documented.
- Ship the use case and measure it against the baseline.
- Reuse the layer for the next use case. Each one gets cheaper.
After two or three use cases, most clients find they have a modern integration layer they can also use for new mobile apps and partner portals. The AI project paid for the modernisation.
Running an older core system?
We can review it with your team and show you the shortest path to your first AI use case, with a scope, timeline and cost range. Talk to our engineering team →
