Same AI price, very different costs: what to automate in South-east Asia - The Business Times
A CUSTOMER opens a chat to dispute a charge. Resolving it takes a staff member about 20 minutes.
In Singapore, that costs the company about US$7.90. In Indonesia, it costs about US$0.60.
Yet, a leading customer service software firm charges the same price in both markets for an artificial intelligence agent to handle that chat.
Board members across South-east Asia are now asking their executive teams which workflows to hand to AI agents.
Our answer is that the business case depends on where the work is done. We compare the cost of the staff who do the work today with the cost of the AI that would do it instead.
We cover six markets: Singapore, Malaysia, Thailand, Vietnam, the Philippines and Indonesia. Our staff costs are based on official wage data for clerical or similar roles, and a 50 per cent addition for employer contributions, supervision and tools.
One customer service software firm charges from US$0.99 each time its AI agent resolves a chat. But AI agents may not be fully autonomous in customer service, so our calculations factor in human staff costs.
The same software firm reports that its AI agent resolves three chats in four on average. Results vary widely by business, so we assume a more cautious scenario, where the AI resolves six chats in 10. Staff still handle the other four.
The human part of the job then moves from routine handling to reviewing and deciding. The staff checks one in 10 chats, taking three minutes each time.
For a business handling 20,000 chats a month, we assume half a full-time employee is required just to manage and maintain the agent.
By our calculations, at US$0.99 per outcome and even after a business pays for the staff it still needs, an AI agent cuts the cost of a chat in Singapore by about half.
In Malaysia, it cuts the cost by about a quarter, and in Thailand by about a tenth. In Vietnam, it saves nothing. In the Philippines and Indonesia, it adds to the cost.
These figures are illustrative estimates based on a single workflow and average wages, which mask wide variances between firms. In reality too, large clients often negotiate published prices down.
Our calculation also excludes the cost of building an agent, connecting it to existing systems and keeping it up to date as models change.
Businesses must therefore run the math using their own data.
The success rate matters too. If the agent resolves only half of all chats, the savings in Singapore fall to about 40 per cent.
If the agent resolves three in four, the savings rise to about 60 per cent. Singapore still saves the most and Indonesia still loses the most.
The reference firm charges for each chat its agent resolves. In higher-wage markets where staff cost more than about US$1 a chat, an agent with a higher resolution rate saves more money.
But where staff costs less, as in the Philippines and Indonesia, every extra chat the agent resolves actually adds to the cost.
Some firms build their own agents rather than buy one from a vendor. For them, the choice of AI model matters a great deal.
For this scenario, we estimated costs for six leading models, based on their published prices. (Models we studied were Claude Sonnet 5.5, GPT-6 Astra, Gemini 3.8 Flash, GLM-5.2, DeepSeek-V4-Flash and Qwen 3.5.)
With the cheapest model, the self-built AI agent costs less than staff in every market. With the most expensive, it costs more everywhere except Singapore.
Engineering also makes a difference. When building an agent, design can directly affect the bottom line.
Take a mid-priced model: an agent that wastefully rereads twice as much company material at every step would lose its savings in Thailand, Vietnam, the Philippines and Indonesia.
An agent engineered to cache material, instead of rereading it, would pay off in all six markets.
Today, most AI models run in a few giant data centres, mostly outside South-east Asia, and are sold at one global price.
As more countries insist on data sovereignty, keeping data and AI processing at home, more AI is likely to run locally. It would then carry local costs, including power and the cost of meeting each country’s rules and regulations.
Power is the easiest of these costs to measure. Electricity for a data centre costs about three times as much in Singapore as in Indonesia. Still, that makes the same AI only about 10 per cent more expensive.
By contrast, staff costs nearly 13 times as much. So even if AI runs locally, wages are likely to decide where AI agents make commercial sense. The cost of compliance is the open question.
Leading AI companies currently spend far more than they earn.
OpenAI expects to burn through about US$280 billion between 2026 and 2030, according to company projections reported by the Financial Times.
Prices may therefore rise. On our estimates, if a mid-priced model trebled its token prices, the savings from AI would disappear everywhere in South-east Asia except Singapore.
Exchange rates add a second risk. AI prices are set in US dollars. A weaker ringgit, baht, dong or rupiah makes the AI agent more expensive in local terms, while local salaries stay roughly the same.
Still, costs may not be the only consideration. AI agents work around the clock; they can absorb sudden peaks in demand, help launch new services faster and create new ways to serve customers.
Where an AI agent saves little in operating costs, these benefits could still justify deploying one.
Start with the process. Map each workflow and remove unnecessary steps. Every step an AI agent performs adds cost.
Then ask three simple questions of each workflow: One, what does it cost us today? Two, what will it cost with an AI agent, including the staff it still needs? Three, how far could AI prices rise, or the local currency fall, before the business case fails?
Answer these questions separately for each market. Where the savings are small or negative, name the other benefits you expect in advance and measure them.
Crucially, businesses should keep the decision reversible for as long as possible.
Once teams are reshaped around AI agents, the know-how to do the work without them may be lost, and rebuilding that is slow and costly.
A phased roll-out is sensible. Each phase shows what the agent can handle before more of the workflow depends on it.
Avoid tying the workflow to a single provider or model. If that provider changes its terms, retires a model or fails, the business should be able to switch without starting over.
AI may cost the same in every market, but the case for it is won or lost, workflow by workflow and market by market.
Daniel Liebau is an adjunct associate professor with the Asian Institute of Digital Finance at the National University of Singapore, and the founding managing director of Lightbulb Capital, an innovation advisory firm. Damien Kopp is the founder and managing director of RebootUp, an AI advisory firm, and associate faculty, executive development, at Singapore Management University.
Workings are available from the authors on request.
Court clears path for Frasers’ S$391.9 million Centrepoint en bloc deal
Could the proposed COE changes mean higher premiums for all car buyers?