How SingPost rebuilt its entire architecture before touching AI

Pathross says the data and architecture foundations must come before any AI programme.

Singapore Post Ltd (SingPost) has torn out and rebuilt its entire technology architecture from scratch, with its Chief Information Technology Officer arguing that no AI programme can deliver results without fixing the foundation first.

Gavin Pathross, speaking at the Singapore Business Review Technology Enterprise Summit, said the company’s cloud environment had been running on multiple platforms simultaneously, none of them the product of deliberate planning. The first task, he said, was not to build AI but to fix the plumbing.

“AI is really difficult to do if you do not have two things built well,” he said. “One is your enterprise architecture. Number two is your data architecture.”

SingPost’s cloud environment had spanned AWS, Azure, Ali Cloud, Google, and at least one other platform, the product of additions without a coherent strategy.

“This was not a multi-cloud by strategy,” Pathross said. “This was a multi-cloud by accident.”

The company has since consolidated everything onto Google’s tech stack, with Pathross citing the strength of its overall capabilities as the decisive factor. That unified foundation is what made it possible to start building AI solutions properly. The rebuild required a direct mandate from the top.

“To have your CEO give you the mandate to say that please rip out the entire architecture from before and rebuild it from the ground up is the first part that I think is critically important,” he said.

He reserved his sharpest observation for organisations that skip that step. “If your plumbing is wrong, forget about AI,” he said.

With the foundation in place, SingPost has 16 AI projects underway. Amongst the key projects is workload optimisation, which involves redistributing parcel and letter loads across its postmen force in real time, accounting for volume, parcel size, vehicle type, and route familiarity. The system also handles real-time delivery redirections mid-route, such as when a recipient changes their delivery address after a postman has already left the depot.

Pathross drew a sharp distinction between the types of AI organisations actually need. 

“If you have sufficient data, actually machine learning is more than sufficient. You do not need Gen AI,” he said. “What you really need is really understanding what data you have today.”

SingPost sits on a significant data asset. The company operates a sortation facility that 3D scans every parcel, and its postmen network touches every Singaporean on any given day. The constraint is not the data itself but the legal framework around using it, with PDPA compliance remaining an obstacle to fully exploiting what SingPost holds.

The talent gap is a parallel concern. Despite having 16 AI projects running, more than half of the IT team cannot build an agent.

“I can tell you more than half of my team cannot build an agent,” Pathross said. “That’s concerning. If my CFO and COO are building their own agents without the IT team, then my question to my IT team is: how are you going to help yourself?”

His prescription is self-reliance. Singapore has the infrastructure to support upskilling through SkillsFuture and NTUC programmes, but individuals need to take ownership of their own development rather than waiting for their organisations to do it for them.