Malaysia's current Prime Minister, Anwar Ibrahim, just built a digital version of himself to explain policy and handle routine requests.  

The prime minister who cloned himself  

​Meet PMX AI. Malaysian Prime Minister Anwar Ibrahim launched an agentic avatar of himself, live on WhatsApp (though briefly offline for updates shortly after launch), built by local firm Zetrix AI.

It's really two systems in one face. As an agentic feature, it renews your driving licence, confirms the payment, and points you toward government training programmes or job openings, the same counters any citizen would otherwise queue at. That part needs no real expertise, just execution.

The other part does. Trained on his real speeches and policy record, it explains government policy and programmes the way he would, in English, Malay, and regional dialects. That part is genuinely him, knowledge and voice, reproduced on demand.

One layer is task automation borrowing a familiar face. The other is captured expertise, and that's the harder one to build.

Now picture doing both for your best person

​Your team already has both layers sitting unused.

The routine layer looks like expense approvals, meeting scheduling, invoice processing, hours lost to work that needs no real judgment at all. Most of it still runs through a human because no one's built the automated version, not because it requires expertise.

The expertise layer is rarer and harder to copy. Your best account executive's way of explaining the product, the language that closes a stalled deal. Your sharpest campaign lead's read on what messaging lands with which segment. None of it exists outside their heads. Average sales rep tenure sits at 18 months, turnover runs 35% a year. Agencies lose 40% of institutional knowledge every time a senior executive walks out.

Both layers exist. Most companies have built neither.

Start with the layer that doesn’t need permission

​The two layers deserve two different timelines.

The routine layer needs no debate. If a workflow runs on rules, not judgment, automating it carries almost no risk. List every task your team repeats the same way, then pick the one eating the most hours. That's your first build, not a strategy debate.

The expertise layer takes longer, but the first step is just as concrete. You don't need to decide today whether an AI should ever speak in your best rep's voice to a customer. Start smaller: record their next three calls or campaign reviews, and write down how they actually explain things, the phrasing, the framing, the answers to common objections. That's the raw material. Turning it into something reusable comes later.

Most companies treat both as one big, uncertain AI decision, and end up doing nothing. One is safe to start this week. The other just needs a first recording saved.