Sequencing AI Initiatives for ASEAN Conglomerates

Why most multi-business-unit AI programmes stall — and the sequencing pattern that gets them past month four into a compounding portfolio of Digital Workers.

The conglomerate problem

Multi-business-unit conglomerates across ASEAN — Malaysian and regional alike — have a sequencing problem unique to their structure. Each BU has its own AI priorities. The group function wants horizontal leverage. The board wants a coherent narrative. And the first major AI investment, almost without fail, becomes a referendum on internal politics rather than on outcomes.

The result: 18 months of pilots scattered across BUs, none of them compounding, with a frustrated board and a CIO who has lost credibility. We see this pattern in five out of every six conglomerate AI programmes we audit.

The sequencing failure mode

The instinct is to "let each BU pick its own use case." This feels fair, empowering, and decentralised. It is also wrong, and predictably so. Three things happen:

1. BUs pick visible-but-fragile use cases. Customer-facing chatbots win the BU-level pitch but are technically hard, regulatorily exposed, and slow to ROI.
2. No infrastructure is shared. Each BU ends up with its own vendor stack, its own data pipelines, its own governance, its own bills. The horizontal leverage the group function was promised never materialises.
3. Failures look like AI failures. When one BU pilot fails (and statistically one will), it tarnishes the whole programme — even though the failure was actually about that specific BU running solo without infrastructure support.

The sequencing pattern that works

For ASEAN conglomerates specifically, the sequence that survives contact with quarterly results looks like this:

Phase 1 (months 0-3): Group function pilots, not BUs

Start the AI programme inside group function workflows — finance, HR, legal, procurement — that affect every BU but belong to none. These workflows have three properties that make them ideal first pilots: they are high-volume (real ROI), they are politically neutral (no BU loses budget), and the data is already centralised at group level (no integration ordeal).

Examples: group AP automation, contract-review triage, group recruitment screening, internal audit working-paper drafting. Boring? Yes. Foundational? Also yes.

Phase 2 (months 3-6): Infrastructure pattern emerges

By the time the group function pilots are 90 days in, the infrastructure pattern — what platform, what governance, what audit, what cost model — has been forged through real use. Document it. That documentation becomes the operating spec every BU pilot in phase 3 inherits.

Phase 3 (months 6-9): BU pilots, but only against the pattern

Now you can run BU pilots — but each one must use the platform, governance, and audit pattern proven in Phase 1. The BUs get the autonomy to pick the use case, but not the autonomy to reinvent the stack. This is the leverage moment. Each BU pilot now takes 6 weeks instead of 11, because the foundation already exists.

Phase 4 (months 9-12): Portfolio review

By month 12, you should have 5-8 Workers in production across the group. Some are clearly compounding. Some will be retired. The group AI committee — which by now has real authority — makes the portfolio investment calls for year two. The programme is now operating, not piloting.

We made every mistake on this list before someone showed us this sequence. We restarted in 2025 from group function workflows, and 14 months later we have 9 Digital Workers across the group with a governance model the board actually trusts.
— Group Chief Strategy Officer, Malaysian diversified conglomerate

What this requires from leadership

• Group CEO or COO active sponsorship. Group-function-first sequencing is unpopular at BU level. It needs top cover.
• A real group AI function. One person, named, with authority. Not a committee. Not a project.
• Patience for the first 90 days. The BU heads will press. Hold the sequence anyway.
• Honest sunset criteria. Some Workers will fail. The programme depends on retiring them gracefully without it becoming a verdict on AI.

How long does this actually take?

For a 3-BU group, plan 12 months. For a 5-7 BU group, plan 14-16. Trying to compress beyond that produces the same pattern that failed the first time. The sequence works because of its sequence, not its speed.

Map your own conglomerate sequencing →