Retailers urged to prioritise AI use cases over unified data platforms

Firms that see value track the true cost of AI.

Retailers chasing a unified data platform before deploying artificial intelligence (AI) are aiming at the wrong thing and instead must pick one use case first.

James Wilson, a partner, technology consulting, advisory and head of consumer & retail at KPMG Singapore, said companies should focus on the data needed for a specific decision rather than trying to consolidate all their data first. "A local store does not need to know everything about all the data in the organisation," he told the Retail Asia Summit 2026 in Singapore.

For a three-month push, he advised against starting where data sits in Excel spreadsheets, which he said needs a heavy lift.

Wilson sorted AI use cases into three types, including productivity, cost avoidance, and growth. “Productivity covers faster reporting and store administration whilst cost avoidance covers predictive models that flag stock-outs”.

Growth, which he called the ‘holy grail’, is where firms turn customer data into something new.

However, Wilson noted that the three types are not equally easy to fund.

“Productivity and cost-saving projects win budgets because they're tangible and you can measure the delta", he said.

Meanwhile, revenue projects have no baseline, making them a shot in the dark. Most firms are still at the proof-of-value stage on growth, he added.

Drawing on a quarterly KPMG survey of global AI leaders, Wilson said firms that see value track the true cost of AI.

He warned that one client had a huge list of use cases but cannot prioritise them. He advised selecting those with a solid business case and a return on investment, then building modular architecture, data flows and AI governance to scale.

For regional chains, Wilson said the centre should own governance, standards, architecture and success measures. “Local markets should handle customisation,” he said.

He also warned that spreadsheet data tends to lack governance, with many files managed by one person on their laptop. Feeding them to AI models, Wilson said, may not produce the right answer.